betaFeedback

WorksMathematics

How much can a human being lift? A Bayesian latent capability model for Olympic weightlifting

Abstract

Estimates of the limits of human performance are usually built on record books, on the premise that a record is the best an athlete could do. In Olympic weightlifting the premise fails, because the competitor declares the weight before each of three attempts per lift and only the heaviest successful attempt is recorded, so a record is a lower bound, chosen by the athlete, on a maximal capability never observed. We treat the maximal lift of each competitor on each competition day as a latent variable, of which every attempt is a trial and every declared weight a measurement. Its distribution at a given bodyweight follows an extreme value mixture model with an estimated threshold, and an athlete-specific random effect links the competitions of each athlete. The model is fitted to the 117,892 attempts of 5,040 athletes in 141 international competitions between 2017 and 2026. The upper tail of capability is bounded, with posterior probability above 0.999 for men, 0.96 for the women's snatch and 0.94 for the women's clean and jerk. At 110 kilograms of bodyweight the men's ceiling lies one to three per cent above the heaviest lifts, against twelve and sixteen per cent for women at 86 kilograms. The record book alone places the men's ceiling six and twelve kilograms higher, because the best lift of a day ending on a success overstates capability.

Introduction

Consider a lifter who opens the snatch at 180 kilograms and makes it, raises to 185 and makes it, and then misses at 190. The official result is 185 kilograms. It shows that the lifter’s capability on that day was at least 185 kilograms, but not by how much it exceeded that value. The miss at 190 is weak evidence of an upper limit, since a lift can fail through timing, bar position or nerve as well as through lack of strength, and 188 kilograms might have succeeded. The opening weight of 180 kilograms is stronger evidence about where the lifter placed that limit, because lifters choose an opener they expect to lift. The recorded result is thus one of three choices made by an athlete who knows more about his or her own capability than the result reveals.

Statistical estimates of the limits of human performance start from record books. Einmahl and Magnus [13] treated the best performances in athletics as draws from a distribution and estimated its right endpoint, the limit of the discipline, from the observed extremes. The same approach has since been applied to sprinting [14, 12], to longevity [1, 17], to swimming, with progressively richer models [38, 39]. In every case the recorded result is taken to be the performance. In Olympic weightlifting that assumption fails, and it fails in a structured way. A competitor has three attempts at the snatch and three at the clean and jerk. The weight is declared before each attempt, may never decrease, and the heaviest success is credited. About half of all attempt sequences in international competition contain both a made lift and a heavier miss, so the capability of those competitors is bracketed from both sides. Most of the rest end on a successful final attempt, so the capability of those competitors is bounded only from below. The second group contains the champions, whose lifts matter most for the upper tail and about whom the record says least.

A natural starting point is a latent variable model in which each competitor has an unobserved capability on each day and each attempt is a pass-or-fail trial whose success probability falls as the bar rises above that capability. The difficulty lies in how the weights are chosen. They are not fixed in advance but set in sequence by athletes who know their own strength, in response to the outcomes of earlier attempts. After a successful opener the second attempt succeeds about six times in ten whether the bar is raised by two kilograms or by ten, so the declared weights carry much of the information about capability and the outcomes little. A model that treats the weights as a fixed design ignores that information.

This paper models the unobserved capability together with the choices made about it. Each competitor on each competition day is given a true maximal lift for the snatch and one for the clean and jerk. Across competitors of a given bodyweight these capabilities follow an extreme value mixture model [4, 36], with a mixture of normal distributions for the body and a generalized Pareto tail above a threshold that is estimated with everything else. The shape of that tail decides whether strength is bounded. Every attempt is a trial against the capability, with a logistic success probability whose scale is the amount of luck in a lift. Every declared weight is a measurement of it, set below capability by a margin whose typical size and spread depend on the situation in which the weight was chosen, an opener, a raise after a success or a raise after a miss, subject to the rule that the bar must go up after a success.

Not everything can be learned from competition data. If every declared weight were placed further below capability and the attempt-order effects adjusted accordingly, the likelihood would change only through the outcomes of first attempts. The data therefore say little about how far below their limit athletes declare. This margin decides the distance between a record and the capability behind it. Left free, the fit moves towards athletes who declare far below their limit and attempts whose outcome is close to a coin flip, a description that the measured reliability of maximal strength rules out. We therefore fix the scale of attempt-to-attempt variation from that external evidence and report how the results change with it.

The upper tail of capability is bounded with posterior probability above 0.999 for men. For the women’s two lifts it is 0.96 and 0.94. At 110 kilograms of bodyweight the men’s ceiling lies one to three per cent above the heaviest lifts since 2019. The women’s ceilings at 86 kilograms lie twelve and sixteen per cent above theirs. Fitted to the best lifts alone, the same distribution places the men’s ceiling six and twelve kilograms higher, because the best lift of a day ending on a successful attempt overstates capability.

The capability distribution also has a practical use. To compare lifters of different bodyweights, the sport multiplies each total by a coefficient that depends on bodyweight [37], refitted to the world records every Olympic cycle. A lift could instead be scored by its position in the capability distribution at the lifter’s bodyweight, such as the quantile it reaches, which does not depend on a handful of records. More generally, whenever performances come from challenges that athletes choose for themselves, those choices carry information about ability and belong in the model, although how far below their limit athletes choose to go cannot be learned from competition data alone.

Data

The data are the results of the senior international competitions sanctioned by the sport’s world governing body between March 2017 and May 2026, namely world and continental championships, Olympic Games, World Cup and Grand Prix meetings, and the qualifying events of each Olympic cycle. They come from the open OpenWeightlifting project (https://github.com/euanwm/OpenWeightlifting), which compiles the governing body’s result sheets. The processed attempt-level data and all code are available at https://github.com/manueleleonelli/weightlifting-capability. For every competitor at every competition the archive records name, bodyweight at the weigh-in to two decimals, the weight on the bar at each of six attempts and whether each attempt succeeded.

The unit of analysis is the athlete-day-lift, one competitor, one competition and one of the two disciplines, with up to three attempts. The data contain 117,892 attempts by 5,040 athletes, 2,646 men and 2,394 women, in 141 competitions. They make up 39,883 athlete-day-lifts on 20,198 athlete-days, of which 19,685 have both lifts. The same athlete appears in four competitions on average and in up to thirty-seven (§ 4.4).

The archive goes back to 1998 (Table 1), but the earlier years differ from the present in two ways that a smooth time trend cannot absorb. The weight categories of that era were abolished in 2018 together with every standing world record. The re-analysis of stored samples from the 2008 and 2012 Olympic Games produced sixty-one retrospective doping violations in weightlifting, thirty-four of them by medallists ([24]). At 86 kilograms of bodyweight, for instance, the heaviest women’s lifts of all years, from 2009 and 2012, have not been approached since. A model fitted to all years would carry such lifts to the present through its time trend. The analysis therefore starts in March 2017, and § 17 of the supplement repeats it on the whole archive.

PeriodCompetitionsAthlete-day-liftsAthletes
1998–2004145,8651,602
2005–20115014,3792,958
2012–20166918,2593,475
2017–2018339,0122,167
2019–202610830,8714,055

Table 1. Coverage of the archive by period.

Bodyweight enters the model as a continuous covariate, not through the weight categories, whose boundaries were redrawn in 1993, 1998, 2018 and 2025, twice within the analysis period. A record series by category breaks at each such revision, unlike bodyweight.

Background

Extreme value models for sporting records

Three lines of work bear on the problem. The first estimates the right endpoint of a performance distribution directly from the top performances. Einmahl and Magnus [13] did this for athletics, obtaining both an ultimate record for each event and a measure of how close the current record is to it. Einmahl and Smeets [14] applied the method to the 100 metres, and Einmahl and He [12] refined it, finding the men’s record within hundredths of a second of its estimated limit, with a much larger margin for women. The same tools have asked whether human longevity is bounded [1, 17] and how far the dominance of three players in men’s tennis lay beyond the rest of the field [26].

The second line models the whole population of performances rather than only the best. Stephenson and Tawn [40] built a model for athletics in which all performances are described at once, an idea anticipated by the record progression analysis of Berthelot et al. [6]. Robinson and Tawn [34] asked whether an individual record, the women’s 3000 metres world record set in 1993, is anomalous relative to the distribution that produced it. Nguyen et al. [32] describe the ten events of the decathlon with a Bayesian compositional model with athlete effects and find a score above 9,200 points possible but unlikely.

The third line, closest to the present work, is Spearing et al. [38], who model swim times through their rate of occurrence with covariates in location and scale. They were motivated by a defect of the official points system, which scores a swim against the current world record, itself a random extreme, rather than against the whole distribution. Weightlifting’s Sinclair coefficient [37], recalibrated on records each Olympic cycle, has the same defect.

Latent trials, adaptive designs and informative censoring

A made lift bounds capability from below and a missed lift bounds it, weakly, from above, so the natural description of the problem is as censoring. The literature on extremes under censoring [5, 15, 45] assumes censoring independent of the quantity of interest. Here it is not. The athlete chooses the weights knowing his or her capability, so the censoring is informative [28].

A pass-or-fail trial whose success probability depends on the difference between a latent ability and a known difficulty is the Rasch model of item response theory [33]. Two of its properties matter here. The number of successes is sufficient for the ability, so the conditional likelihood given that number involves the difficulties alone [3], a device that has been proposed for identifying the scale of the trial without a model for the ability. In adaptive testing the items are chosen in sequence in the light of earlier responses, and the likelihood that conditions on the chosen items remains valid when the selection depends on the observed responses alone, but not when it depends on the ability itself [31]. The same distinction is Rubin’s condition for an ignorable design [35]. The declared weights of a weightlifter are chosen in the light of the ability, so neither the conditional likelihood nor the likelihood that conditions on the weights is valid, and § 4 replaces conditioning with a model for the choice.

Extreme value mixture models and dependence

Any model with a generalized Pareto tail has to say where the tail begins. The classical approach fixes a threshold from diagnostic plots and fits the exceedances alone [10, 7]. Extreme value mixture models [36] instead combine a model for the bulk below the threshold with a generalized Pareto distribution above it and estimate the threshold with everything else. Behrens et al. [4] introduced the construction with a parametric bulk, and do Nascimento et al. (2012) made the bulk a mixture of gamma distributions. This paper uses a mixture of normal distributions instead, since the standardized error of capability takes values on the whole real line, and estimates the threshold inside a latent variable model.

For the dependence between two extremes the basic distinction is between asymptotic dependence, where two variables that are both extreme remain related however far out one looks, and asymptotic independence, where the relation fades [25, 8]. The conditional model of Heffernan and Tawn [22] and the family of Wadsworth et al. [44] fit data without deciding the class in advance. Bivariate analysis of Olympic results goes back at least to Adam and Tawn [2].

Statistical work on weightlifting

The literature specific to the sport is mostly about scaling performance by bodyweight, from the dimensional argument of Lietzke [27] and the record analysis of Croucher [9] to the coefficient of Sinclair [37] that is still in official use and later revisions [16]. Bayesian hierarchical models of individual performance are common in sports statistics [23].

The closest modelling work is Griffin et al. [19], who build Bayesian models of how individual sprinters and weightlifters progress with age, extended to within-season trajectories by Griffin et al. [20]. That work models the body of the performance distribution and takes the recorded best as the performance. Attempt selection has been studied descriptively, mostly in powerlifting [42], and responds to competitive position, which is why the censoring cannot be ignored. Neither the tail of strength nor the attempt structure has been modelled before.

Methodology

The model has four layers. The first is the distribution of capability across athlete-days at a given bodyweight and year, whose upper endpoint is the ultimate limit. The second relates capability to the success of each attempt, and the third treats each declared weight as a measurement of capability. The fourth links the competitions of the same athlete, and § 4.6 combines the layers into a likelihood.

The distribution of capability

Index by ii an athlete-day, one competitor at one competition, and by l∈{s,c}l \in\{s,c\} the two disciplines, snatch and clean and jerk. Let ZilZ_i^l be the heaviest weight that competitor ii could have raised in discipline ll on that day. We call it the capability, and § 4.2 makes the definition exact. Capability depends on the size of the athlete and on the era, and is modelled as

Zil=μl(xi)+ωl(xi)εil,(1)Z_i^l = \mu_l(x_i) + \omega_l(x_i)\varepsilon_i^l, \tag*{(1)}

a location and a scale that shift with covariates, and a standardized error that carries the shape of the distribution. The covariates are three. The first is bodyweight at the weigh-in, entered as bi=(log⁡bodyweight⁡i−mˉ)/sb_i = (\log\operatorname{bodyweight}_i - \bar{m})/s, centred and scaled by the sample mean and standard deviation of log bodyweight. The second is its square, made orthogonal to the intercept and to bib_i, scaled to unit standard deviation and written bi(2)b_i^{(2)}. The third is the competition year centred on its sample mean, tit_i. The location is linear in all three,

μℓ(xi)=βℓ0+βℓ1bi+βℓ2bi(2)+βℓ3ti,(2)\mu_{\ell}(x_i) = \beta_{\ell0} + \beta_{\ell1} b_i + \beta_{\ell2} b_i^{(2)} + \beta_{\ell3} t_i, \tag*{(2)}

so that capability rises with bodyweight at a rate that is allowed to flatten for the heaviest athletes, and drifts with time. The scale is log-quadratic in bodyweight,

log⁡ωℓ(xi)=γℓ0+γℓ1bi+γℓ2bi(2),(3)\log\omega_{\ell}(x_i) = \gamma_{\ell0} + \gamma_{\ell1} b_i + \gamma_{\ell2} b_i^{(2)}, \tag*{(3)}

so that the spread of capability among athletes of the same size may widen or narrow with size, and differently at the two ends of the range. Equation (1) is a regression, in which μℓ(xi)\mu_{\ell}(x_i) locates the capability of athletes of that size and year while ωℓ(xi)\omega_{\ell}(x_i) measures their spread.

The error εiℓ\varepsilon_i^{\ell} carries the shape of the distribution. The population of athlete-days is drawn from fields of very different depth, from continental championships and regional games, whose competitors trail far below the typical lift, to world championships and Olympic Games, whose competitors are compressed into a narrow band above it. The body of the error is therefore a mixture of three normal distributions,

hℓ(e)=∑k=13wℓk1sℓkϕ(e−mℓksℓk),Hℓ(e)=∑k=13wℓkΦ(e−mℓksℓk),(4)h_{\ell}(e) = \sum_{k=1}^{3} w_{\ell k} \frac{1}{s_{\ell k}} \phi\left(\frac{e-m_{\ell k}}{s_{\ell k}}\right), \qquad H_{\ell}(e) = \sum_{k=1}^{3} w_{\ell k} \Phi\left(\frac{e-m_{\ell k}}{s_{\ell k}}\right), \tag*{(4)}

where ϕ\phi and Φ\Phi are the standard normal density and distribution function. The weights wℓkw_{\ell k} are positive with unit sum. The components have means mℓkm_{\ell k} and standard deviations sℓks_{\ell k}. The first component has mℓ1=0m_{\ell1} = 0 and sℓ1=1s_{\ell1} = 1, which fixes the location and scale of the error. Then μℓ(x)\mu_{\ell}(x) is the centre of the reference component and ωℓ(x)\omega_{\ell}(x) its spread in kilograms. The other two components are free, and with them the body can be skewed, heavy on either side or bimodal as the fields require. The upper tail is where extreme value theory enters. Above a threshold uℓu_{\ell} the error follows a generalized Pareto distribution,

gℓ(e)={hℓ(e),e≤uℓ,{1−Hℓ(uℓ)}1σℓ(1+ξℓe−uℓσℓ)−1/ξℓ−1,e>uℓ,(5)g_{\ell}(e) = \begin{cases} h_{\ell}(e), & e \le u_{\ell},\\ \{1-H_{\ell}(u_{\ell})\}\dfrac{1}{\sigma_{\ell}}\left(1+\xi_{\ell}\dfrac{e-u_{\ell}}{\sigma_{\ell}}\right)^{-1/\xi_{\ell}-1}, & e > u_{\ell}, \end{cases} \tag*{(5)}

where ξℓ\xi_{\ell} is the shape of the tail and σℓ>0\sigma_{\ell} > 0 its scale. The tail carries the mass 1−Hℓ(uℓ)1-H_{\ell}(u_{\ell}) that the body places above the threshold, so that the two pieces integrate to one. This is the extreme value mixture model of Behrens et al. [4] with the mixture body of do Nascimento et al. (2012), with normal components in place of gamma components because the error takes values on the whole real line. The density may jump at the threshold, as in those papers, since nothing in the problem requires the body and the tail to meet smoothly. The generalized Pareto is the form that exceedances of a high threshold take for any distribution with a well-behaved tail (Davison and Smith, [10]; Coles, [7]), and its shape answers the question of the paper. For ξℓ<0\xi_{\ell} < 0 the distribution has a finite upper endpoint. For ξℓ=0\xi_{\ell} = 0 the tail is exponential, for ξℓ>0\xi_{\ell} > 0 heavier than exponential, and in both cases unbounded. The regression in (1) removes location and scale before the tail is modelled, as in McNeil and Frey [30], so that the standardized error is the athlete’s position relative to competitors of the same size and era.

Write GlG_l for the distribution function of (5), equal to Hl(e)H_l(e) for e≤ule \le u_l and to 1−{1−Hl(ul)}{1+ξl(e−ul)/σl}−1/ξl1-\{1-H_l(u_l)\}\{1+\xi_l(e-u_l)/\sigma_l\}^{-1/\xi_l} above the threshold. Two quantities follow. For a probability p>Hl(ul)p>H_l(u_l), the capability exceeded by a fraction 1−p1-p of athlete-days at covariate value xx is

Ql(p;x)=μl(x)+ωl(x)[ul+σlξl{(1−p1−Hl(ul))−ξl−1}](6)Q_l(p;x)=\mu_l(x)+\omega_l(x)\left[u_l+\frac{\sigma_l}{\xi_l}\left\{\left(\frac{1-p}{1-H_l(u_l)}\right)^{-\xi_l}-1\right\}\right] \tag*{(6)}

and when ζl<0\zeta_l<0 the distribution has the finite upper endpoint

ζl(x)=μl(x)+ωl(x)(ul+σl−ξl).(7)\zeta_l(x)=\mu_l(x)+\omega_l(x)\left(u_l+\frac{\sigma_l}{-\xi_l}\right). \tag*{(7)}

Equation (7) is what an ultimate limit means in this paper, the heaviest weight that any competitor of bodyweight and era xx could raise. Uncertainty about it has two parts, the probability P(ζl≥0)P(\zeta_l\ge0) that no such limit exists and the distribution of ζl(x)\zeta_l(x) given that one does. Because ζl→∞\zeta_l\to\infty as ζl→0−\zeta_l\to0^-, that distribution has a heavy upper tail whenever shapes close to zero are plausible. We therefore report the probability that a ceiling exists, the probability that it lies within ten, twenty-five or fifty per cent of what has been lifted, and its median.

From capability to attempts

Capability is never observed. What is observed, for each athlete-day and discipline, is a sequence of up to three declared weights wi1l≤wi2l≤wi3lw_{i1}^l\le w_{i2}^l\le w_{i3}^l, nondecreasing by rule, with a success indicator SijlS_{ij}^l for each. The attempt model says how the outcomes arise from ZilZ_i^l. Whether an attempt succeeds depends on more than strength, namely on the timing of the pull, the position of the catch and the athlete’s nerve at that moment. We represent this by giving each attempt its own momentary maximum, Zilexp⁡{τl(δjl+Lijl)}Z_i^l\exp\{\tau_l(\delta_j^l+L_{ij}^l)\}, where LijlL_{ij}^l is a standard logistic variable independent across attempts, and the attempt succeeds when the declared weight lies below that moment’s maximum. The spread of the momentary maximum is proportional to capability, because a lifter whose maximum is 200 kilograms varies from attempt to attempt by more kilograms than one whose maximum is 100, while the relative scale τl\tau_l is the same for both. Later attempts are allowed to be systematically easier or harder than the first at the same margin through δjl\delta_j^l, with δ1l=0\delta_1^l=0. Since LL has distribution function F(v)=(1+e−v)−1F(v)=(1+e^{-v})^{-1} and is symmetric about zero,

P(Sijl=1∣Zil)=F(log⁡(Zil/wijl)τl+δjl).(8)P(S_{ij}^l=1\mid Z_i^l)=F\left(\frac{\log(Z_i^l/w_{ij}^l)}{\tau_l}+\delta_j^l\right). \tag*{(8)}

Equation (8) is what defines capability. A first attempt at exactly ZilZ_i^l succeeds with probability one half, an attempt a fraction τl\tau_l below it succeeds with probability 0.73, and one 2τl2\tau_l below with probability 0.88. Conditionally on ZilZ_i^l and the shifts, the three attempts are independent. With log⁡Zil/τl\log Z_i^l/\tau_l as the ability and log⁡wijl/τl−δjl\log w_{ij}^l/\tau_l-\delta_j^l as the difficulty, (8) is a Rasch model (Rasch, [33]) with continuous, known difficulties.

Declared weights as measurements

A declared weight is the athlete’s own statement about his or her capability, which we treat as a measurement with an error that depends on the situation in which the weight was chosen. Combined with the attempts of § 4.2, this gives a selection model in the sense of Little and Rubin [28]. There are three situations, k=1,2,3k = 1, 2, 3. The opener is chosen before any outcome is known. A raise after a successful attempt must exceed the previous bar by rule. A raise after a missed attempt is the minority choice, since about three athletes in four repeat the weight after a miss. A repeat declares nothing new and enters the model only through its outcome. In situation kk the declared weight at attempt jj is capability reduced by a margin,

log⁡wijl=log⁡Zil−Rijl,(9)\log w_{ij}^{l} = \log Z_i^{l} - R_{ij}^{l}, \tag*{(9)}

whose distribution, as for a measurement error, is specified given the quantity measured. Given ZilZ_i^{l}, the margin has a two-piece normal distribution with mode mklm_k^{l}, the margin the athlete typically leaves. Its scale skl,+s_k^{l,+} above the mode describes how far below capability some athletes declare, and its scale skl,−s_k^{l,-} below the mode how far above capability they reach. Both raises, k=2,3k = 2, 3, add at least one kilogram to the bar, wijl≥wi,j−1l+1w_{ij}^{l} \ge w_{i,j-1}^{l} + 1, which truncates the margin from above at log⁡{Zil/(wi,j−1l+1)}\log\{Z_i^{l}/(w_{i,j-1}^{l} + 1)\}. The density is divided by the two-piece normal distribution function there. Otherwise the margin is independent of capability, so the declarations follow the distribution of capability, shifted down and spread by the margins.

Repeated competitions of the same athlete

The population of athlete-days is not exchangeable. The analysis set contains 2,646 men and 2,394 women, who appear in up to thirty-seven competitions each. The top of the capability distribution is made up of the repeated competitions of a few champions. We let the competitions of the same athlete share a factor on the probability scale. Writing Uil=Gl(ϵil)U_i^{l} = G_l(\epsilon_i^{l}) for the error on the uniform scale and Cil=Φ−1(Uil)C_i^{l} = \Phi^{-1}(U_i^{l}) for the same error on the normal scale, the competitions of athlete aa have

Cil=ρa,lVal+1−ρa,l ϵil,(10)C_i^{l} = \sqrt{\rho_{a,l}}V_a^{l} + \sqrt{1-\rho_{a,l}}\,\epsilon_i^{l}, \tag*{(10)}

with ValV_a^{l} and ϵil\epsilon_i^{l} independent standard normal. Every athlete-day keeps the marginal distribution (5), so its quantiles and endpoint are unchanged, while the correlation ρa,l\rho_{a,l} between two competitions of the same athlete measures how much of the spread of capability is persistent.

The two disciplines together

A competitor’s snatch and clean and jerk on the same day are related, since they draw on the same body, the same preparation and the same form. The relation is placed on the errors on the uniform scale of § 4.4. The pair (Uis,Uic)(U_i^{s}, U_i^{c}) is given a bivariate tt copula with correlation ρ\rho and degrees of freedom 1/ψ1/\psi, ψ∈[0,1/2]\psi\in[0, 1/2]. The correlation says how strongly the two disciplines move together in the body of the distribution. The degrees of freedom say whether they move together in the extremes. For ψ=0\psi= 0 the copula is Gaussian and the two disciplines are asymptotically independent (Ledford and Tawn [25]). The probability that the clean and jerk is extreme given that the snatch is extreme then goes to zero as the level rises. For any ψ>0\psi> 0 the copula is asymptotically dependent and that probability has a positive limit. Which of the two holds decides whether a competitor can be at the top of both lifts on the same day, and so how large a total can be.

The likelihood

For athlete-day-lift (i,l)(i,l) write ηijl(z)=log⁡(z/wijl)/τl+δjl\eta_{ij}^{l}(z)=\log(z/w_{ij}^{l})/\tau_l+\delta_j^l for the argument of the logistic distribution function FF in (8) at capability zz, Jil\mathcal{J}_i^l for the set of attempts whose weight is a declaration in one of the three situations of § 4.3, kijk_{ij} for the situation of attempt j∈Jilj\in\mathcal{J}_i^l, and dkld_k^l and DklD_k^l for the two-piece normal density and distribution function of the margin in situation kk. A weight repeated after a miss is left out of Jil\mathcal{J}_i^l. Given capability zz, the attempts and declarations of the athlete-day-lift contribute

ℓil(z)=∏jF{ηijl(z)}Sijl[1−F{ηijl(z)}]1−Sijl∏j∈Jildkijl{log⁡(z/wijl)}wijlDkijl[log⁡{z/(wi,j−1l+1)}].(11)\ell_i^l(z)=\prod_j F\{\eta_{ij}^l(z)\}^{S_{ij}^l}\left[1-F\{\eta_{ij}^l(z)\}\right]^{1-S_{ij}^l}\prod_{j\in\mathcal{J}_i^l}\frac{d_{k_{ij}}^l\{\log(z/w_{ij}^l)\}}{w_{ij}^lD_{k_{ij}}^l\left[\log\{z/(w_{i,j-1}^l+1)\}\right]}. \tag*{(11)}

where the first product runs over all attempts taken and D1l≡1D_1^l\equiv1, because the opener is not truncated. By (1) and (5) the capability of athlete-day ii has density and distribution function

gil(z)=1ωl(xi)gl{z−μl(xi)ωl(xi)},Gil(z)=Gl{z−μl(xi)ωl(xi)}.g_i^l(z)=\frac{1}{\omega_l(x_i)}g_l\left\{\frac{z-\mu_l(x_i)}{\omega_l(x_i)}\right\},\qquad G_i^l(z)=G_l\left\{\frac{z-\mu_l(x_i)}{\omega_l(x_i)}\right\}.

Given the athlete factor Val=vV_a^l=v, the athlete-days of athlete aa are independent, and by (10) the conditional density of ZilZ_i^l is gil(z)κρa,l[Φ−1{Gil(z)},v]g_i^l(z)\kappa_{\rho_{a,l}}[\Phi^{-1}\{G_i^l(z)\},v], where

κρ(c,v)=11−ρ ϕ(c)ϕ(c−ρv1−ρ).\kappa_\rho(c,v)=\frac{1}{\sqrt{1-\rho}\,\phi(c)}\phi\left(\frac{c-\sqrt{\rho}v}{\sqrt{1-\rho}}\right).

is the ratio of the conditional to the marginal density of CilC_i^l. The likelihood of discipline ll, with parameters θl\theta_l, is therefore

Ll(θl)=∏a∫−∞∞∏i∈Ia(∫0∞ℓil(z)gil(z)κρa,l[Φ−1{Gil(z)},v] dz)ϕ(v) dv.(12)L_l(\theta_l)=\prod_a\int_{-\infty}^{\infty}\prod_{i\in\mathcal{I}_a}\left(\int_0^\infty\ell_i^l(z)g_i^l(z)\kappa_{\rho_{a,l}}\left[\Phi^{-1}\{G_i^l(z)\},v\right]\,dz\right)\phi(v)\,dv. \tag*{(12)}

For each athlete, the capability in every competition is integrated out given the athlete factor, which is then integrated out. With ρa,l=0\rho_{a,l}=0, κ=1\kappa=1 and the athlete-days are independent.

The two disciplines are joined in a second stage (§ 4.9), in which the athlete factor is left out and the athlete-days are treated as independent. A competitor who completes both lifts on day ii contributes an integral over the two capabilities,

Lsc(ρ,ψ)=∏i∬ℓis(zs)ℓic(zc)gis(zs)gic(zc)cρ,ψ{Gis(zs),Gic(zc)} dzs dzc,(13)L_{\mathrm{sc}}(\rho,\psi)=\prod_i\iint\ell_i^s(z^s)\ell_i^c(z^c)g_i^s(z^s)g_i^c(z^c)c_{\rho,\psi}\{G_i^s(z^s),G_i^c(z^c)\}\,dz^s\,dz^c, \tag*{(13)}

where cρ,ψc_{\rho,\psi} is the density of the t copula of § 4.5.

Identification

Comparing declarations with outcomes in (12) might seem to identify τl\tau_l, since the declarations locate capability relative to the bar and the outcomes show how often a bar at that distance is lifted. Proposition 1 shows that this holds only in part.

Proposition 1. Fix τl\tau_l and let θl\theta_l collect the other parameters of discipline ll. For Δ∈R\Delta\in\mathbb{R}, let TΔθlT_\Delta\theta_l be obtained from θl\theta_l by multiplying the location coefficients βl0,…,βl3\beta_{l0},\ldots,\beta_{l3} by eΔe^\Delta, adding Δ\Delta to the scale intercept γl0\gamma_{l0} and to the margin modes ml1,ml2,ml3m^1_l,m^2_l,m^3_l, and subtracting Δ/τl\Delta/\tau_l from the attempt-order shifts δl2\delta^2_l and δl3\delta^3_l, with all other parameters unchanged. Then, for every θl\theta_l and Δ\Delta,

Ll(TΔθl)=Ll(Δ)(θl),L_l(T_\Delta\theta_l)=L_l^{(\Delta)}(\theta_l),

where Ll(Δ)L_l^{(\Delta)} is the likelihood (12) with ηi11(z)\eta^1_{i1}(z) replaced by ηi11(z)+Δ/τl\eta^1_{i1}(z)+\Delta/\tau_l in (11) for every athlete-day-lift. In particular, the likelihood without the outcomes of the first attempts is constant along {TΔθl:Δ∈R}\{T_\Delta\theta_l:\Delta\in\mathbb{R}\}.

The proof is in Section S9 of the supplement. The level of capability relative to the declared weights, which decides how far a record lies below the ceiling, is therefore learned only from the outcomes of the openers, and only for a given τl\tau_l. A larger τl\tau_l with all margins larger can leave the opener success rate unchanged while weakening the dependence of each outcome on its margin. Since τl\tau_l and Δ\Delta move together, the data determine them weakly. Left free, the fit drifts along the direction of Proposition 1 to a scale of 14 per cent (Section 5.2).

The comparison with a Rasch model shows why. A Rasch model sets the scale of the logistic to one by convention, which defines the unit of ability, and with known difficulties it identifies the level of ability as well. Here the unit is fixed by the kilogram. The scale τl\tau_l is the variability of a lifter from one attempt to the next, so its value changes the distance between a record and the ceiling. The difficulties are known but chosen by the athletes a few per cent below their capability, through margins whose mode is unknown, and the attempt-order shifts are free, which leaves only the openers to fix the level. Attempts far from capability, which would reveal the slope of the success curve, are almost never made.

Physiology bounds τl\tau_l. The test-retest coefficient of variation of a one-repetition maximum has a median of 3.3 per cent among participants with resistance-training experience in the studies reviewed by Grgic et al. [18], and the variation between attempts minutes apart on the same day cannot exceed the variation between sessions. As Manski [29] and Gustafson [21] recommend for parameters that the data do not pin down, we bring in this outside knowledge, reporting the conclusions as a function of it. We fix τl\tau_l at 0.02, for which the momentary maximum has a standard deviation of 3.6 per cent of capability and a lift two per cent under capability succeeds three times in four. Every quantity of interest is also reported at 0.01 and 0.03. The bound implies τl≤0.018\tau_l \le0.018, so the reference value lies slightly above it. It is kept as a round value between an anchor well inside the bound and one outside it. The conclusions change little across that range (Section 5.7).

Priors

Inference is Bayesian. Every parameter except the attempt scale τl\tau_l, fixed in Section 4.7, receives a prior distribution. All statements about uncertainty in the paper are posterior probabilities and intervals. The posterior is explored by Markov chain Monte Carlo (Section 4.9). The priors are weakly informative. The location coefficients have βlk∼N(0,3002)\beta_{lk}\sim N(0,300^2). The scale coefficients have γl0∼N(log⁡15,0.72)\gamma_{l0}\sim N(\log15,0.7^2), so that the spread of capability among athletes of average size is centred on 15 kilograms, and γl1,γl2∼N(0,1)\gamma_{l1},\gamma_{l2}\sim N(0,1).

For the body, the means of the two free components have ml2,ml3∼N(0,32)m_{l2},m_{l3}\sim N(0,3^2), the log standard deviations log⁡slk∼N(0,1)\log s_{lk}\sim N(0,1) on (0.05,20)(0.05,20), and the weights are wlk∝ealkw_{lk}\propto e^{a_{lk}} with al1=0a_{l1}=0 and al2,al3∼N(0,1.52)a_{l2},a_{l3}\sim N(0,1.5^2). The two free components have the same prior and are not ordered, because an ordering creates a boundary at which the sampler sticks when they overlap. Every quantity reported below is invariant to their labels. The threshold is a parameter, with ul∼N(1.5,1)u_l \sim N(1.5,1) on (0.2,4)(0.2,4), centred between the 0.9 and 0.95 quantiles of the reference component, so that the uncertainty about where the tail begins is carried into every quantity derived from the tail. The tail has log⁡σl∼N(log⁡0.5,1)\log\sigma_l \sim N(\log0.5,1) on (0.01,10)(0.01,10) and ξl∼N(0,0.32)\xi_l \sim N(0,0.3^2) truncated to (−1.5,0.4)(-1.5,0.4), which gives probability 0.55 to a bounded tail. Reweighting the posterior draws to a wider prior for ξl\xi_l or ulu_l, or to a threshold below 3, moves the median ceilings by at most 1.1 kilograms (§ 15 of the supplement). The wider prior for the threshold lowers the probability of a finite ceiling for the women’s clean and jerk from 0.94 to 0.89 and moves the other three by at most 0.01.

The attempt-order shifts have δjl∼N(0,1.52)\delta_j^l \sim N(0,1.5^2), the margin modes mkl∼N(0.03,0.12)m_k^l \sim N(0.03,0.1^2) on (−0.2,0.3)(-0.2,0.3) and the log scales of the margins N(log⁡0.03,1)N(\log0.03,1) on (0.003,1)(0.003,1). The athlete correlation ρa,l\rho_{a,l} is uniform on (0,0.95)(0,0.95). The copula parameter ρ\rho is uniform on (0.88,0.98)(0.88,0.98) and ψ\psi on [0,0.5][0,0.5], ranges whose upper bounds the posterior does not reach.

Estimation

The capabilities are integrated out of (12) by quadrature rather than sampled. The inner integrand is smooth except at the threshold and at the modes of the margin densities, some of whose scales are as small as 0.3 per cent of the bar, so the range is cut at these points, each piece integrated by Gauss–Legendre quadrature. The athlete factor is integrated by twelve-point Gauss–Hermite quadrature, centred and scaled for each athlete. Both rules agree with brute-force integration (Section S1 of the supplement). The likelihood of one discipline has twenty-eight parameters and takes about a fifth of a second to evaluate. Integrating the capabilities out also helps with the threshold. When a threshold is estimated from observed data, the likelihood jumps as observations cross it. Here ulu_l enters only through the integrand, and the likelihood is smooth in it.

The posterior is harder to explore. Raising the threshold moves mass from the body to the tail, which the scale of the generalized Pareto, the free mixture components and the regression coefficients compensate, so these parameters lie along curved ridges. Two devices address the problem. The first replaces σl\sigma_l in the sampler by the threshold-stable scale σl∗=σl−ξˉlul\sigma_l^* = \sigma_l - \bar{\xi}_l u_l ([7 Section 4.3.4]), which does not change when the threshold moves and in terms of which the endpoint of the standardized error is σl∗/(−ξˉl)\sigma_l^*/(-\bar{\xi}_l). The second is the sampler, differential-evolution Markov chains with sampling from the past and snooker updates ((ter Braak and Vrugt, 2008)), whose proposals move along differences between past states of the chains, following the shape of the posterior, curved ridges included.

Four chains, started around the posterior mode, run for 120,000 generations each, or 360,000 for the two women’s lifts, whose posteriors put part of their mass on shapes at or above zero. The first 30,000 generations are discarded and every twentieth state of the rest is kept. The posterior of the women’s snatch has two regions, one with a higher threshold and a shape closer to zero. They agree on the one-in-a-thousand level but differ on the endpoint, and the reported posterior mixes them with weights estimated by a further set of chains that moves between them. The one-in-a-thousand level and the endpoint have R^≤1.05\hat{R} \le1.05 in every fit, within each region for the women’s snatch (§ 16 of the supplement).

The two disciplines are fitted separately in this way, and the copula in a second stage. The likelihood (13) is evaluated on a grid of (ρ,ψ)(\rho,\psi) values at five posterior draws of the marginal parameters, and the posterior over the grid is averaged across those draws (§ 18 of the supplement). This stage also gives draws of the pair of capabilities for every athlete-day, which drive the dependence diagnostics of § 5.6.

Results

What the raw attempts show

Success becomes rarer with each attempt. In the men’s snatch the proportion of successful attempts is 0.771 at the first attempt, 0.633 at the second and 0.437 at the third, in the clean and jerk 0.795, 0.601 and 0.348. The women’s rates are within four points of the men’s. The median raise is three to four per cent of the bar.

Most athlete-day-lifts do not end cleanly. In 52 per cent of men’s snatches and 58 per cent of clean and jerks the sequence contains both a made lift and a heavier miss, so capability is bracketed on both sides. In 44 and 37 per cent the final attempt succeeded, so capability is bounded only from below. In 4 and 5 per cent every attempt failed. The women’s shares are similar. An analysis that equates capability with the best made lift ignores the heavier miss when capability is bracketed. When the final attempt succeeded, it treats a lower bound as the capability itself, and this is how most of the heaviest lifts end.

The declared weights are chosen to equalize the chance of success. Figure 1 shows, for second attempts that follow a successful opener, the proportion that succeeded against the size of the raise. For raises between two and ten kilograms the proportion lies between 0.58 and 0.69 in all four combinations of sex and lift, with no trend. Third attempts after a successful second attempt decline from 0.42 to 0.49 at two kilograms to 0.26 to 0.36 at eight to ten. If raises were chosen without knowledge of capability, a larger raise would leave a smaller margin, and success would fall steeply with its size. The flat curve of the second attempt indicates that athletes raise by more when they have more in reserve.

Figure 1. Proportion of second attempts that succeeded after a successful opener, by the size of the raise in kilograms, and the proportion predicted by the fitted model at its posterior mode for each attempt from the weights declared before its outcome was known. For the women’s snatch the mode is that of the region with the larger posterior mass. Raises of one kilogram are rare, eight to twenty-one attempts in each combination.

The outcome of an opener also depends little on how it compares with the athlete’s previous best. Table 2 classifies openers by their ratio to the athlete’s best made lift in the previous four hundred days. The median opener is 96 to 97 per cent of that best, and its success rate falls only from 0.88 to 0.93 for openers below 88 per cent of it to 0.71 to 0.79 for openers at or above it, in all four combinations. Openers above the previous best, declared by athletes who have improved, succeed as often as those just below it. If the previous best were the capability, a trial whose scale is a few per cent of the bar would move from near certainty to near failure across this range. Either the previous best is a poor measure of the capability of the day, or the outcome of a lift depends weakly on its margin, and the archive by itself cannot say which. § 5.2 shows that this is the central fact about what the data identify.

Opener / previous bestMen, snatch nnMen, snatch madeMen, c&j nnMen, c&j madeWomen, snatch nnWomen, snatch madeWomen, c&j nnWomen, c&j made
below 0.881910.881660.892790.892300.93
0.88 to 0.912780.892340.864160.843380.88
0.91 to 0.947920.836980.869870.808520.87
0.94 to 0.972,0680.791,7330.811,8660.781,7220.84
0.97 to 1.002,0260.722,0780.761,7170.721,8020.79
1.00 to 1.034500.725590.713800.724630.78
above 1.033710.733810.772990.733660.79

Table 2. Openers relative to the athlete’s heaviest lift made in the previous four hundred days, for athlete-days with such a lift: number of openers in each band of the ratio and proportion that succeeded.

Strength grows with bodyweight more slowly than the classical scaling argument predicts. A regression of the log best made lift on log bodyweight and year gives exponents from 0.465 to 0.509, well below the two-thirds surface law of Lietzke [27].

The attempt scale

Figure 2 follows the direction of Proposition 1 for the men’s snatch, fixing the attempt scale τ\tau at one to eight per cent of the bar and re-estimating everything else (Section S2 of the supplement). It uses a simplified model, without the athlete factor and with one lower margin scale shared across the three situations, cheap enough to trace the whole direction. The profile log likelihood rises without a maximum in the range that physiology allows, by 1,086 units from τ=0.01\tau= 0.01 to τ=0.03\tau= 0.03. The margins grow roughly in step with τ\tau, as the proposition says they must, while the tail shape stays between −0.32-0.32 and −0.39-0.39. The endpoint at 110 kilograms of bodyweight rises from 202.6 kilograms at τ=0.01\tau= 0.01 to 224.3 in the free fit, which ends at τ=0.120\tau= 0.120 with openers 17 per cent below capability. The final model behaves alike. Left free, it ends at τ=0.144\tau= 0.144, 1,349 units of log likelihood above the fit at τ=0.02\tau= 0.02, with openers 17 per cent below capability, the tail shape at the upper limit of its support and the men’s snatch ceiling infinite. The archive prefers a description in which a world record is a lucky lift by an athlete whose capability lies far above what is declared. That description contradicts the measured variability of maximal strength. The rest of this section fixes τ\tau at 0.02, with the other two anchors alongside.

Figure 2. The simplified model of § 5.2 for the men’s snatch, along the direction of Proposition 1: profile log likelihood relative to its maximum, modes of the opener margin and of the raise after a success, and the endpoint and one-in-a-thousand capability at 110 kilograms of bodyweight in 2026, against the fixed attempt scale τ\tau on a log axis. The triangle is the fit with τ\tau free. The dotted line marks the reference anchor τ=0.02\tau= 0.02 and the dashed line in the right panel the local record of the current era.

Attempts and declarations

Table 3 gives the posterior for the attempt and declaration parameters at the reference anchor. With τ=0.02\tau= 0.02 an opener two per cent below capability, three kilograms on a bar of 150 kilograms, succeeds about three times in four, and one four per cent below about nine times in ten. At the same margin between capability and bar, the second attempt succeeds with 1.4 to 2.3 times the odds of the opener and the third with 1.2 to 2.9 times. This reflects partly the nerves of the first lift and partly that an athlete who raises is one whose previous lift went up easily. The low raw success rate of third attempts therefore comes from where the bar is placed.

Table 3. Posterior medians and 95 per cent intervals for the attempt and declaration parameters at the reference anchor τ=0.02\tau= 0.02: the attempt-order shifts on the log odds scale, and the modes and scales of the margin between capability and the declared weight, in per cent of the bar, for openers and raises after a success.

Men, snatchMen, c&jWomen, snatchWomen, c&j
δ2\delta_{2}0.66 (0.59, 0.72)0.37 (0.30, 0.44)0.82 (0.75, 0.90)0.43 (0.35, 0.50)
δ3\delta_{3}0.75 (0.68, 0.81)0.15 (0.08, 0.22)1.06 (0.99, 1.14)0.24 (0.16, 0.31)
Opener — mode3.3 (3.0, 3.5)3.6 (3.4, 3.9)1.8 (1.5, 2.1)0.8 (0.6, 1.0)
upper scale1.7 (1.6, 1.9)1.8 (1.6, 1.9)4.2 (4.0, 4.4)4.6 (4.5, 4.7)
lower scale2.2 (2.1, 2.3)2.3 (2.1, 2.4)2.0 (1.8, 2.2)0.3 (0.3, 0.5)
Raise after a success — mode1.5 (1.4, 1.7)2.0 (1.9, 2.2)1.6 (1.2, 2.0)0.1 (−0.2-0.2, 0.4)
upper scale0.3 (0.3, 0.3)0.3 (0.3, 0.3)3.0 (2.4, 3.9)11.4 (8.2, 20.9)
lower scale2.9 (2.9, 3.0)3.1 (3.0, 3.1)3.2 (3.1, 3.4)2.3 (2.2, 2.5)

Table 3.

The margin densities differ in shape between the sexes but not in their averages, and the average opener margin is 2.9 to 4.2 per cent of the bar in all four combinations. The men’s opener margin is close to symmetric about a mode of 3.3 and 3.6 per cent, while the women’s has modes of 1.8 and 0.8 per cent with a long upper tail of conservative openers. After a success men raise to about their capability, while women again show a low mode and a long upper tail. The margin after a miss, the minority choice, is weakly determined (§ 15 of the supplement). Three scales sit at their lower bound of 0.3 per cent of the bar, among them the upper scale of the men’s raises after a success. At bars of 100 to 250 kilograms the bound is 0.3 to 0.75 kilograms, below the resolution of the declared weights, which are whole kilograms. At this anchor the athlete opens three to four per cent below what he or she can lift, raises to about capability after a success, and treats a third attempt as a toss-up.

The attempt model assumes that a declared weight says nothing about the form of the attempt beyond capability and the earlier outcomes. Predicting each attempt from everything observed before its outcome, the model matches the success rates by attempt to within 0.035 in every combination (§ 15 of the supplement), but it predicts a decline with the size of the raise where the observed curve is flat (Figure 1). In the men’s snatch the predicted rate falls from 0.66 at two kilograms to 0.40 at eight to ten, against observed rates between 0.60 and 0.69. The margin density of a raise after a success has a lower scale of about three per cent, so the model reads a large raise partly as a bar above capability and predicts failure in proportion. A flat predicted curve would require the outcome of a lift to depend less on its margin, which is the direction of Figure 2. The misfit is the visible price of the anchor.

The body and the tail

Table 4 collects the parameters of the capability distribution. In every combination the body has a long lower tail, which carries the athlete-days of the weak fields. Its components trade off against one another and against the location coefficients, so they are less well determined than the distribution they make up. The threshold lies 1.5 to 2.0 scale units above the location, at the 64th to 93rd percentile of the error, and is the least well determined parameter of the tail. In the men’s clean and jerk it lies at the 64th percentile, so the tail covers the upper third. The men’s tail is bounded with posterior probability above 0.999, with shapes ξs=−0.31\xi_s = -0.31 (−0.36-0.36, −0.25-0.25) and ξc=−0.40\xi_c = -0.40 (−0.44-0.44, −0.36-0.36). For women the shapes are −0.23-0.23 (−0.29-0.29, 0.05) in the snatch and −0.18-0.18 (−0.28-0.28, 0.14) in the clean and jerk, negative with posterior probabilities 0.96 and 0.94. The upper ends of the women’s intervals come from the part of each posterior that sits at shapes near zero, which for the snatch is the higher-threshold region of § 4.9. The athlete correlation ρa\rho_a is 0.90 to 0.93, so about ninety per cent of the variation in capability is persistent within athletes.

Table 4. Posterior medians and 95 per cent intervals for the capability distribution at the reference anchor: the threshold ulu_l on the scale of the standardized error, the probability Hl(ul)H_l(u_l) that the body places below it, the scale and shape of the generalized Pareto tail, the correlation between two competitions of the same athlete, the probability that the tail has a finite endpoint, and the time trend of the location.

Men, snatchMen, c&jWomen, snatchWomen, c&j
ulu_l1.75 (1.12, 2.56)2.00 (0.93, 3.26)1.47 (0.94, 2.48)1.63 (1.01, 3.17)
Hl(ul)H_l(u_l)0.846 (0.805, 0.877)0.643 (0.596, 0.723)0.855 (0.799, 0.992)0.933 (0.857, 0.999)
σl\sigma_l0.42 (0.35, 0.53)0.71 (0.55, 0.95)0.96 (0.19, 1.64)0.57 (0.14, 0.85)
ξl\xi_l−0.31-0.31 (−0.36-0.36, −0.25-0.25)−0.40-0.40 (−0.44-0.44, −0.36-0.36)−0.23-0.23 (−0.29-0.29, 0.05)−0.18-0.18 (−0.28-0.28, 0.14)
ρa,l\rho_{a,l}0.906 (0.900, 0.912)0.895 (0.889, 0.902)0.921 (0.915, 0.926)0.929 (0.923, 0.934)
P(ξl<0)P(\xi_l < 0)1.001.000.960.94
Location trend (kg per year)0.32 (0.26, 0.37)0.36 (0.29, 0.44)0.60 (0.55, 0.64)0.79 (0.74, 0.84)

Table 4.

From the record book to capability

The difference between capability and the recorded best lift is a posterior quantity, and at the reference anchor it is within 0.7 kilograms on average in all four combinations. Its sign depends on how the day ended. When a heavier attempt was missed, capability lies 0.5 to 2.1 kilograms above the best made lift, since the miss bounds it from above. When the day ended on a success, it lies 0.6 to 2.0 kilograms below, since the final attempt was declared at a weight the athlete makes about as often as not. Of the one per cent of athlete-days with the largest best lift relative to bodyweight and era, 76 to 84 per cent ended on a successful attempt, against 39 to 49 per cent of all athlete-days with a made lift. At this anchor the record book therefore measures median capability with little bias on average and overstates it at the top. At a larger anchor it is a lower bound throughout, since along Figure 2 the gap grows in step with τ\tau.

Table 5 separates what each layer of the model contributes to the tail. Every row uses the same distribution of capability, the normal-mixture body with a generalized Pareto tail above an estimated threshold, and the same priors, so the rows differ only in what the model is fitted to. The first row is the record book. It takes the best made lift of each athlete-day as the capability and the athlete-days as exchangeable, and since lifts are recorded to the kilogram, each best lift contributes the probability of its kilogram interval. The second row adds the athlete factor of § 4.4 to the same best lifts. The third is the full model, with the attempts and the declarations. The last row uses no model for the body. It places a threshold at the 0.9 quantile of the best lift from a linear quantile regression on bodyweight and year, fits a generalized Pareto with a log scale linear in bodyweight to the exceedances, and takes intervals from a bootstrap over athletes (§ 13 of the supplement).

The tail of the capability distribution under successive layers of the model, with the same mixture body and threshold estimation throughout: posterior median and 95 per cent interval of the tail shape, posterior probability that the tail is bounded, and the posterior median of the endpoint at the reference bodyweight (110 kilograms for men, 86 for women) in 2026 with, in parentheses, the posterior probability that it lies within ten per cent of the local record of the current era. The last row uses no body: a quantile-regression threshold, a generalized Pareto for the exceedances and an athlete bootstrap, with the share of bootstrap samples with a negative shape and the point estimate of the endpoint.

Men, snatch (record 200)Men, c&j (record 242)Women, snatch (record 127)Women, c&j (record 161)
Record bookξ~l\tilde{\xi}_l−0.23-0.23 (−0.36-0.36, 0.100.10)−0.21-0.21 (−0.27-0.27, −0.14-0.14)−0.13-0.13 (−0.31-0.31, 0.150.15)−0.30-0.30 (−0.36-0.36, −0.21-0.21)
P(ξ~l<0)P(\tilde{\xi}_l < 0)0.930.931.001.000.840.841.001.00
ceiling212212 (0.840.84)257257 (0.940.94)150150 (0.020.02)180180 (0.070.07)
Record book, athlete factorξ~l\tilde{\xi}_l−0.24-0.24 (−0.57-0.57, 0.280.28)−0.11-0.11 (−0.49-0.49, 0.280.28)−0.18-0.18 (−0.22-0.22, −0.14-0.14)−0.15-0.15 (−0.56-0.56, 0.320.32)
P(ξ~l<0)P(\tilde{\xi}_l < 0)0.820.820.790.791.001.000.740.74
ceiling219219 (0.530.53)307307 (0.080.08)151151 (0.000.00)204204 (0.000.00)
Full modelξ~l\tilde{\xi}_l−0.31-0.31 (−0.36-0.36, −0.25-0.25)−0.40-0.40 (−0.44-0.44, −0.36-0.36)−0.23-0.23 (−0.29-0.29, 0.050.05)−0.18-0.18 (−0.28-0.28, 0.140.14)
P(ξ~l<0)P(\tilde{\xi}_l < 0)1.001.001.001.000.960.960.940.94
ceiling206206 (1.001.00)245245 (1.001.00)143143 (0.110.11)187187 (0.020.02)
No body (check)ξ~l\tilde{\xi}_l−0.24-0.24 (−0.34-0.34, −0.16-0.16)−0.22-0.22 (−0.31-0.31, −0.15-0.15)−0.32-0.32 (−0.43-0.43, −0.20-0.20)−0.21-0.21 (−0.33-0.33, −0.08-0.08)
P(ξ~l<0)P(\tilde{\xi}_l < 0)1.001.001.001.001.001.001.001.00
ceiling212212255255140140187187

Table 5. The tail of the capability distribution under successive layers of the model, with the same mixture body and threshold estimation throughout: posterior median and 95 per cent interval of the tail shape, posterior probability that the tail is bounded, and the posterior median of the endpoint at the reference bodyweight (110 kilograms for men, 86 for women) in 2026 with, in parentheses, the posterior probability that it lies within ten per cent of the local record of the current era. The last row uses no body: a quantile-regression threshold, a generalized Pareto for the exceedances and an athlete bootstrap, with the share of bootstrap samples with a negative shape and the point estimate of the endpoint.

Taken as an exchangeable sample, the record book puts the men’s ceiling at 110 kilograms of bodyweight at 212 kilograms in the snatch and 257 in the clean and jerk, six per cent above the local records of 200 and 242. The check without a body agrees with it to within two kilograms. Linking the competitions of the same athlete leaves the record book unable to settle the tail. With the athlete factor the probability of a finite ceiling falls to 0.82 and 0.79. The intervals for the shape reach 0.28. The exchangeable fit treats thirty competitions of one champion as thirty independent draws from the upper tail. Counted as the results of one athlete, the champions are a few draws from a population, and a few draws cannot fix where the tail ends. The entries with the athlete factor for the men’s clean and jerk and the women’s snatch are indicative, as their thresholds mix slowly (§ 16 of the supplement).

The attempts and the declarations fix it. In the full model the men’s tail is bounded with posterior probability above 0.999, with shapes of −0.31 and −0.40. The ceilings are 206 and 245 kilograms, six and twelve kilograms below the record book, three and one per cent above the local records. The comparison of capability with the best made lift explains the difference. A best made lift is capability measured with error, largest on the days that end with a successful final attempt, which make up most of the top of the archive. The latent distribution is free of that error, and its upper tail is shorter than the tail of the best lifts.

For women the layers disagree in both directions. The record book leaves the snatch unbounded with probability 0.16 when the athlete-days are exchangeable, and the clean and jerk with probability 0.26 when the athlete factor is added. In the full model the ceilings at 86 kilograms are 143 and 187 kilograms, twelve and sixteen per cent above the local records of 127 and 161. The check without a body gives 140 and 187. In the snatch the record book places the ceiling seven kilograms above the full model, at 150, as it does for men. In the clean and jerk it places the ceiling below the full model, at 180 kilograms, because it estimates a shorter tail, with a shape of −0.30-0.30 against −0.18-0.18, and the full model’s tail for that lift is the least well determined in the paper.

Dependence between the lifts

The copula gives a correlation ρ\rho of 0.96 for men and 0.97 for women. Its tail parameter ψ\psi is 0.15 for men, a tt copula with seven to ten degrees of freedom, and the Gaussian case ψ=0\psi= 0, under which the two lifts would be asymptotically independent, has posterior probability below 10−1110^{-11}. For women ψ\psi is 0.10 (0.05, 0.10), ten to twenty degrees of freedom, and the Gaussian case has posterior probability 10−510^{-5}. On the posterior draws of the latent capabilities, the probability χ(u)\chi(u) that the clean and jerk exceeds its uu-quantile given that the snatch exceeds its own is 0.59 (0.54, 0.66) at u=0.99u = 0.99 for men and 0.76 (0.71, 0.81) for women. It declines slowly and stays well above zero (Figure 3). The fitted copula gives 0.75 and 0.78 at that level. For women the two agree. For men the latent χ(u)\chi(u) is still falling at u=0.99u = 0.99, and a tt copula, whose χ(u)\chi(u) settles at a positive limit, may overstate the dependence at the top. The latent draws also carry independent posterior uncertainty about each capability, which dilutes the latent χ(u)\chi(u), so for men the dependence at the top lies between the two curves. The competition total therefore loses almost nothing to the need to be extreme in both lifts on the same day. At one in ten thousand athlete-days the men’s total at 110 kilograms is 445.8 (442.7, 449.0), 0.3 kilograms below the sum of the two marginal levels on the same draws, and the women’s deficit at 86 kilograms is 0.4.

Figure 3. Extremal dependence between the lifts at observable levels: χ(u)\chi(u) computed on the posterior draws of the latent capabilities, with intervals over the draws, against the value implied by the fitted copula.

Predictions

Table 6 gives the capability quantiles in 2026 at the reference anchor, with the local record of the current era, the heaviest lift made since 2019 within five kilograms of each bodyweight. Records of the current era are used because the location drifts with time. Lifts of 2017 and 2018 exceed them at seven of the forty combinations of bodyweight and lift, by one to four kilograms. The largest, a clean and jerk of 233 kilograms at 94.0 kilograms of bodyweight in 2017, lies 5.2 kilograms above the ceiling of 2017 at that bodyweight, about two per cent, within the variation of a single attempt, whose momentary maximum has a standard deviation of 3.6 per cent of capability. Up to 110 kilograms of bodyweight a local record is the maximum of 650 to 1,150 athlete-days, so under the model it should sit near or somewhat below the one-in-a-thousand level. At 110 kilograms that level is 199.3 kilograms in the snatch and 241.0 in the clean and jerk, against records of 200 and 242. The record lies within five kilograms of that level at twelve of the sixteen combinations of bodyweight and lift. For women the record does so at sixteen of the twenty combinations and lies furthest below the level at 86 kilograms, by five and seven kilograms. These levels describe the strongest plausible competitor of each size today. Competitions simulated from the posterior, attempts included, reproduce the local records at 36 of the 40 combinations (§ 15 of the supplement). The exceptions are the men’s snatch at 61 kilograms, one kilogram below the simulated range, and three of the four women’s records at the two heaviest bodyweights, which lie two to five kilograms above it.

BodyweightSnatch q0.99q_{0.99}Snatch q0.999q_{0.999}Snatch CeilingSnatch RecordClean and jerk q0.99q_{0.99}Clean and jerk q0.999q_{0.999}Clean and jerk CeilingClean and jerk Record
Men
61141.3 (140.2, 142.4)146.6 (145.2, 148.0)152146173.2 (172.1, 174.5)178.1 (176.8, 179.6)181183
67151.1 (150.1, 152.1)156.7 (155.4, 158.1)162160185.0 (184.0, 186.1)190.2 (188.9, 191.6)193197
73159.5 (158.6, 160.5)165.4 (164.1, 166.8)171169195.2 (194.2, 196.2)200.5 (199.3, 201.9)204209
81169.1 (168.1, 170.0)175.2 (173.9, 176.7)181182206.7 (205.6, 207.7)212.2 (211.0, 213.7)216211
89177.0 (176.0, 178.0)183.4 (182.1, 185.0)190188216.2 (215.1, 217.4)221.9 (220.6, 223.6)226229
96182.8 (181.9, 183.9)189.5 (188.1, 191.1)196188223.3 (222.2, 224.5)229.2 (227.8, 230.9)233233
102187.2 (186.2, 188.3)194.0 (192.6, 195.7)201190228.6 (227.5, 229.9)234.6 (233.1, 236.4)239233
110192.3 (191.2, 193.4)199.3 (197.8, 201.0)206200234.8 (233.5, 236.1)241.0 (239.3, 242.7)245242
130201.7 (200.4, 203.2)209.2 (207.5, 211.1)216204246.5 (244.9, 248.3)253.0 (251.0, 255.2)257250
150208.0 (206.1, 210.1)215.7 (213.6, 218.3)223217254.5 (252.2, 257.0)261.2 (258.6, 264.2)266253
Women
4995.4 (94.4, 96.5)100.8 (99.4, 102.4)10999120.7 (119.4, 122.2)128.4 (126.6, 130.7)144126
55102.7 (101.8, 103.7)108.4 (107.1, 109.9)117111129.5 (128.3, 130.8)137.6 (135.9, 139.8)154141
59106.8 (105.9, 107.8)112.7 (111.4, 114.2)122117134.6 (133.4, 135.9)143.0 (141.2, 145.2)160149
64111.3 (110.4, 112.3)117.5 (116.1, 119.0)127120140.2 (139.0, 141.5)148.9 (147.1, 151.2)166151
71116.6 (115.7, 117.7)123.1 (121.7, 124.7)133125146.9 (145.6, 148.3)156.0 (154.2, 158.4)174156
76119.8 (118.8, 121.0)126.5 (125.1, 128.1)137125151.1 (149.7, 152.5)160.4 (158.5, 162.9)179161
81122.7 (121.6, 123.8)129.5 (128.0, 131.2)140127154.8 (153.3, 156.3)164.3 (162.4, 166.8)183161
86125.2 (124.0, 126.4)132.1 (130.6, 133.8)143127158.1 (156.6, 159.7)167.8 (165.8, 170.3)187161
100130.7 (129.4, 132.2)138.0 (136.4, 139.7)149145165.6 (163.8, 167.6)175.7 (173.6, 178.5)196178
120135.8 (134.0, 138.0)143.5 (141.4, 145.9)155140173.3 (170.8, 176.0)183.9 (180.9, 187.3)205185
SexLiftBodyweightRecordP(finite)P(\mathrm{finite})+10%+25%+50%Ceiling0.010.020.03
MenSnatch731691.001.001.001.00171
MenSnatch891881.001.001.001.00190
MenSnatch1102001.001.001.001.00206207206211
MenSnatch1502171.001.001.001.00223
MenC&J732091.001.001.001.00204
MenC&J892291.001.001.001.00226
MenC&J1102421.001.001.001.00245247245248
MenC&J1502531.001.001.001.00266
WomenSnatch591170.960.850.920.93122
WomenSnatch711250.960.800.910.93133
WomenSnatch861270.960.110.890.93143143142146
WomenSnatch1201400.960.350.900.93155
WomenC&J591490.940.670.900.92160
WomenC&J711560.940.360.870.92174
WomenC&J861610.940.020.830.92187191187187
WomenC&J1201850.940.440.880.92205

Table 6. Posterior medians and 95 per cent intervals for capability quantiles in 2026 at the reference anchor, in kilograms, by sex and bodyweight, the posterior median of the ceiling, and the local record of the current era (the heaviest lift made since 2019 within five kilograms of the bodyweight). The 0.999 level is exceeded by one athlete-day in a thousand at that bodyweight.

Table 7 reports the ceiling. For men the probability of a finite ceiling is above 0.999 in both disciplines and at every bodyweight. The probability that it lies within ten per cent of the local record is at least 0.998. At 110 kilograms of bodyweight the median ceiling is 206 kilograms in the snatch and 245 in the clean and jerk, three and one per cent above the records of 200 and 242. Across the bodyweight range the men’s ceilings lie at most six per cent above the records. At five of the twenty men’s combinations the ceiling lies up to 2.4 per cent below the record, and in all five the record was lifted within 1.4 kilograms of the upper edge of the five-kilogram window. Against the ceiling at the lifter’s own bodyweight every local record lies below it, the closest being the men’s clean and jerk record of 209 kilograms, lifted at 76.7 kilograms of bodyweight and 0.6 kilograms below the ceiling there. For women the ceilings lie 3 to 16 per cent above the local records across the bodyweights of Table 6, furthest at 86 kilograms. Einmahl and He [12] found the same pattern for the women’s 100 metres and attributed it in part to the smaller number of women in their data.

Fitted to the whole archive since 1998, the model places the ceilings higher, at 210 and 250 kilograms for men and 157 and 196 for women (§ 17 of the supplement). The heaviest women’s lifts at 86 kilograms, 138 in the snatch and 166 in the clean and jerk, date from 2009 and 2012 (§ 2). The trend carries the tail they imply to 2026.

Within the anchors that physiology allows, the ceiling barely moves. Between τ=0.01\tau= 0.01 and 0.030.03 the posterior modes of the ceilings at the reference bodyweight move by at most five kilograms (Table 7) and the one-in-a-thousand level of the men’s snatch by 4.5. The attempt-order shifts absorb most of the change in τ\tau, which alters how a record is read. The mean difference between capability and best made lift in the men’s snatch is −1.4-1.4 kilograms at τ=0.01\tau= 0.01, −0.4-0.4 at 0.020.02 and 0.70.7 at 0.030.03. A simulation study (Section S2 of the supplement) shows the same on data generated at τ=0.02\tau= 0.02. Refitted at the true anchor, the model recovers the tail. Refitted at 0.01 or 0.03, it moves the attempt-order shifts and margins, while the one-in-a-thousand levels and ceilings move by at most 1.8 kilograms on average.

Discussion

The men’s results sit close to those obtained for other sports from record books. Einmahl and Magnus [13] found the men’s athletics records within a few per cent of their estimated limits, and Einmahl and He [12] found the men’s 100 metres record within hundredths of a second of its limit with a much larger margin for women. Here the men’s ceiling lies at most six per cent above the heaviest lifts of the current era and the women’s 3 to 16 per cent above theirs, across the bodyweights of Table 6. The same distribution fitted to the best lifts, as those analyses do,

places the men’s ceilings six and twelve kilograms higher than the full model does, a body-free generalized Pareto fit six and ten kilograms higher. Table 5 shows where the difference comes from. The record book reads the luck of a made lift as capability, which lengthens the tail. Once the competitions of the same athlete are linked it no longer settles where the tail ends. The extremal dependence between the two lifts resembles that found by Adam and Tawn [2] in bivariate Olympic data.

The treatment of the design places the paper beside the work on adaptive testing and informative censoring [31, 28]. When a latent quantity is probed by challenges the subject chooses, a model that conditions on the challenges loses the identification of the measurement scale. A model for the choice restores most of it. Proposition 1 names what remains open, the position of the latent scale relative to the challenges, which the paper settles with outside knowledge of how much a lift varies from one attempt to the next. The extreme value mixture model of Behrens et al. [4] and do Nascimento et al. (2012) carries over to a latent variable model with a gain. With the latent quantities integrated out, the marginal likelihood is smooth in the threshold, which removes the difficulty of threshold estimation with observed data. Estimating the threshold also matters. Held at a fixed quantile, the threshold moves the record-book ceilings by up to ten kilograms (§ 14 of the supplement).

Several limitations qualify the claims. The first is the anchor. At τ=0.02\tau= 0.02 the momentary maximum has a standard deviation of 3.6 per cent of capability, close to the between-session coefficient of variation of 3.3 per cent reported by Grgic et al. [18]. The within-day variation that τ\tau represents is plausibly smaller. Between τ=0.01\tau= 0.01 and τ=0.03\tau= 0.03 the posterior modes of the ceilings move by at most five kilograms (Table 7). The anchor acts mainly on the attempt-order shifts and on how a successful record is read. The attempt model is logistic in the margin with a single scale. It predicts a decline of success with the size of a raise that the data do not show. A floor on the miss probability would remove the misfit, but at a fixed τ\tau it drifts along the direction of Figure 2, where the data cannot tell a floor from a shift of the capability scale. A shared error in the athlete’s assessment of the day’s capability improves the fit. It moves the one-in-a-thousand levels by at most 1.3 kilograms and the ceilings by at most 2.4, except at the higher-threshold mode of the women’s snatch, whose endpoint responds to any change of a shape that sits near zero (Section S5 of the supplement). The model also understates the very top of the best lifts. In competitions simulated from the posterior the 0.99 quantile of the best made lifts, relative to bodyweight and era, is lower than observed in at least 95 per cent of the simulations for every sex and lift. Among the top one per cent of best lifts 54 to 64 per cent end on a made attempt, against 76 to 84 per cent in the data (Section S5 of the supplement). The model reads some of these made final attempts as luck, which pulls the inferred capability of the strongest days down. The ceilings are therefore more likely conservative than generous, and the difference from the record book in Table 5 is if anything an upper bound on its bias.

The model also leaves out some structure. The choice of weight responds to the competitive situation, whether a medal is at stake and what rivals have declared, which the model does not see. Nor does it describe the choice between repeating and raising a weight after a miss. The athlete factor is constant over a career, where a trajectory per athlete, as in Griffin et al. [19], would follow improvement and decline. Fitted to the whole archive rather than the years since 2017, the model raises the men’s ceilings by four and five kilograms, the women’s by fourteen and nine (Section S7 of the supplement). Bodyweight is measured at the weigh-in, so the model gives the capability of an athlete who makes a given weight, which suits questions about competition better than about physiology.

Weightlifting records are declared before they are lifted. Read through a measurement model, the declarations locate the limit that the records only bound. The capability distribution could also serve the sport, by scoring a lift through its position in the distribution at the athlete’s own bodyweight. Such a score would not depend on the records it is compared with. It would have the properties that Spearing et al. [38] established for swimming.

Funding

Supported by Spanish State Research Agency grant PID2023-153222OB-I00.

Code and data

The R and C++ code and the processed attempt-level data that reproduce every table and figure of the paper and of the supplement are available at https://github.com/manueleleonelli/weightlifting-capability.

Statement on AI use

The author used Claude (Anthropic), a large language model, as an assistant during the research. It was used to write and debug the R and C++ code of the analysis under the author’s direction, to run and check computations, to check the numbers quoted in the text against the outputs of the code, and to edit the text for clarity and length. The author checked all code and results and takes full responsibility for the paper.

Supplement

The quadrature rules

The likelihood (12) of the main text is computed by quadrature in two layers. Dropping the superscript ll, each athlete-day-lift requires

Ii(v)=∫0∞qi(z;v) dz,qi(z;v)=ℓi(z) gi(z) κρa[Φ−1{Gi(z)},v],I_i(v)=\int_{0}^{\infty}q_i(z;v)\,\mathrm{d}z,\qquad q_i(z;v)=\ell_i(z)\,g_i(z)\,\kappa_{\rho_a}\left[\Phi^{-1}\{G_i(z)\},v\right],

and each athlete aa requires ∫∏i∈IaIi(v) ϕ(v) dv\int\prod_{i\in\mathcal{I}_a} I_i(v)\,\phi(v)\,\mathrm{d}v. This section gives the rule for each integral and checks that the quadrature error is negligible.

The integral over capability

Support. Let wiw_i be the lightest made weight of athlete-day-lift ii and w‾i\overline{w}_i the heaviest missed weight, and for each declared attempt j∈Jij\in\mathcal{J}_i let mjm_j, sj−s_j^{-} and sj+s_j^{+} be the mode and the two scales of its margin. The integrand is negligible outside [ai,bi][a_i,b_i], where

ai=max⁡{wi(1−12τ),wi1em1−8s1−},a_i=\max\left\{w_i(1-12\tau),w_{i1}e^{m_1-8s_1^{-}}\right\},
bi=min⁡{w‾i(1+12τ),min⁡j∈Jiwijemj+8sj+}.b_i=\min\left\{\overline{w}_i(1+12\tau),\min_{j\in\mathcal{J}_i}w_{ij}e^{m_j+8s_j^{+}}\right\}.

The first terms come from the attempt model, since a lift at a weight twelve attempt scales away from capability has a success or failure probability of F(−12)≈6×10−6F(-12)\approx6\times10^{-6}. The other terms come from the margin densities, eight scales on either side of the mode. Only the opener bounds capability from below. The density of a raise is divided by its truncation probability, and for the smallest allowed raise, where the declared weight sits at the truncation point, this ratio does not decay below the mode. Without a made lift, the first term of aia_i is replaced by μ(xi)+ω(xi)min⁡k(mk−8sk)\mu(x_i)+\omega(x_i)\min_k(m_k-8s_k), eight standard deviations below the lowest component, and without a miss the first term of bib_i is dropped.

Breakpoints. On [ai,bi][a_i,b_i] the integrand is smooth except at the threshold zu=μ(xi)+ω(xi)uz_u=\mu(x_i)+\omega(x_i)u, where the capability density may jump, and at the modes zj=wijemjz_j=w_{ij}e^{m_j}, where a margin density switches from one scale to the other and its derivative jumps. Some of the scales are small, down to their lower bound of 0.3 per cent of the bar, so these kinks are sharp on the scale of the interval, and a margin with a small lower scale rises steeply just below its mode.

Rule. The interval is cut at zuz_u, at every zjz_j inside it, and at zj−=wijemj−4sj−z_j^{-}=w_{ij}e^{m_j-4s_j^{-}} for every margin whose lower side is narrow, sj−wij<2hs_j^{-}w_{ij}<2h. Each piece [c,d][c,d] is integrated by Gauss–Legendre quadrature with

n={24,d−c>10,min⁡{16,max⁡(4,⌈(d−c)/hcd⌉)},otherwise,n=\begin{cases} 24, & d-c>10,\\ \min\{16,\max(4,\lceil(d-c)/h_{cd}\rceil)\}, & \text{otherwise}, \end{cases}
h=max⁡{min⁡(0.7min⁡j∈Jisj+wij,1),0.25}.h=\max\left\{\min\left(0.7\min_{j\in\mathcal{J}_i}s_j^{+}w_{ij},1\right),0.25\right\}.

with lengths in kilograms, where hcd=hh_{cd}=h except on a piece [z‾j,z‾j][\underline{z}_j,\overline{z}_j], where it is the smaller of hh and max⁡(0.7sˉjwij,0.25)\max(0.7\bar{s}_j w_{ij},0.25). A narrow piece thus gets a node spacing below the narrowest margin scale it contains. A rule that splits [ai,bi][a_i,b_i] only at the threshold, with twenty-four nodes a side, misses the kinks. For the men’s snatch its total log likelihood is in error by about forty units, and an optimizer finds spurious maxima where the narrow densities fall between nodes.

The integral over the athlete factor

For athlete aa with nan_a competitions, let cˉi\bar{c}_i be the mean of Φ−1{Gi(Zi)}\Phi^{-1}\{G_i(Z_i)\} under the normalized integrand qi(z;0)q_i(z;0) of each competition, the capability on the normal scale as it would be estimated with ρa=0\rho_a=0. Treating these as observations of Ci=ρaVa+1−ρa ϵiC_i=\sqrt{\rho_a}V_a+\sqrt{1-\rho_a}\,\epsilon_i gives the approximate posterior mean and standard deviation of the factor,

v^a=ρa1−ρa∑i∈acˉi1+naρa/(1−ρa),s^a={1+naρa1−ρa}−1/2.\hat{v}_a=\frac{\sqrt{\rho_a}}{1-\rho_a}\frac{\sum_{i\in a}\bar{c}_i}{1+n_a\rho_a/(1-\rho_a)},\qquad \hat{s}_a=\left\{1+\frac{n_a\rho_a}{1-\rho_a}\right\}^{-1/2}.

The integral over VaV_a uses twelve-point Gauss–Hermite nodes xkx_k and weights ωk\omega_k centred and scaled at these values,

∫∏i∈aIi(v) ϕ(v) dv≈∑k=112ωkexk22s^aϕ(vk)∏i∈aIi(vk),vk=v^a+2s^axk.\int\prod_{i\in a} I_i(v)\,\phi(v)\,dv \approx\sum_{k=1}^{12}\omega_k e^{x_k^2}\sqrt{2}\hat{s}_a\phi(v_k)\prod_{i\in a}I_i(v_k),\qquad v_k=\hat{v}_a+\sqrt{2}\hat{s}_a x_k.

with each Ii(vk)I_i(v_k) computed as in Section S1.1. A single competition needs no outer integral.

The integral over the shared error

For the shared self-assessment error of § 15, the margins are taken relative to ZieηiZ_i e^{\eta_i} with ηi∼N(0,sη2)\eta_i\sim N(0,s_{\eta}^2), and the likelihood of an athlete-day-lift includes an integral over ηi\eta_i inside the integral over capability. The integrand in ηi\eta_i is a product of two-piece normal densities and truncation factors. It is cut at the points where a margin or a truncation point crosses its mode, with extra breakpoints near the truncation points. On each segment the Gaussian part is integrated in closed form and the remaining smooth factor by Gauss–Legendre quadrature with KηK_{\eta} nodes.

Accuracy

Table S1 compares each rule with a more accurate reference. In every case the discrepancy is far below the differences in log likelihood between the anchors (Table S9) and between the fits with and without the shared error (Table S12), which run to hundreds of units. In addition, the mixture density integrates to one and its survival function matches the integral of the density to six decimal places, and at ρ=0\rho=0 the copula likelihood of the second stage, computed on the product of the two grids, reproduces the sum of the two marginal log likelihoods to 10−1310^{-13}.

IntegralRuleReferenceLargest difference
CapabilityRule of S1.1Grid of 400,001 points10−310^{-3} per athlete-day-lift
Athlete factor12 Gauss–Hermite nodes20 nodes3×10−33 \times10^{-3} per athlete
Athlete factor12 Gauss–Hermite nodesGrid of 4,001 points3×10−33 \times10^{-3} per athlete
Shared errorKη=4K_{\eta}=4Grid of 200,001 points5×10−45 \times10^{-4} per athlete-day-lift
Shared errorKη=4K_{\eta}=4Kη=64K_{\eta}=642 in total

Table S1. Accuracy of the quadrature rules: absolute difference in log likelihood between each rule and its reference. The shared-error rows are at the posterior modes of Table S12.

Table S9. Posterior modes of the final model at the three anchors, for the women’s snatch at both modes of its posterior (§ 16): log likelihood, mode of the opener margin in per cent of the bar, attempt-order shifts, tail shape, and the one-in-a-thousand level and endpoint of the capability distribution at the reference bodyweight (110 kilograms for men, 86 for women) in 2026, with the mean difference between capability and the best made lift in kilograms.

SexLiftFitℓ\ellm^1\hat{m}_1δ^2\hat{\delta}_2δ^3\hat{\delta}_3ξ^\hat{\xi}q0.999q_{0.999}EndpointGap
MenSnatchτ=0.01\tau= 0.01−90,671.71.71.572.46−0.29198.8206.6−1.4
MenSnatchτ=0.02\tau= 0.02−90,025.43.30.660.75−0.31199.2206.0−0.4
MenSnatchτ=0.03\tau= 0.03−89,619.53.20.340.14−0.30203.3210.80.7
MenClean and jerkτ=0.01\tau= 0.01−92,718.51.91.211.70−0.36241.4247.1−0.6
MenClean and jerkτ=0.02\tau= 0.02−92,191.53.60.370.15−0.41240.8244.70.7
MenClean and jerkτ=0.03\tau= 0.03−91,913.94.80.07−0.42−0.41244.5248.52.2
WomenSnatchτ=0.01\tau= 0.01−73,599.20.92.023.13−0.05131.1184.8−1.1
WomenSnatchτ=0.02\tau= 0.02−72,594.91.80.831.07−0.11131.5152.2−0.4
WomenSnatchτ=0.03\tau= 0.03−72,337.30.60.420.35−0.10132.6156.50.3
WomenSnatch, second modeτ=0.01\tau= 0.01−73,592.90.92.023.13−0.22131.7143.4−1.1
WomenSnatch, second modeτ=0.02\tau= 0.02−72,588.41.80.831.07−0.24132.1141.9−0.4
WomenSnatch, second modeτ=0.03\tau= 0.03−72,326.50.60.420.35−0.20133.3145.90.3
WomenClean and jerkτ=0.01\tau= 0.01−77,705.5−0.51.431.95−0.16167.8191.1−0.4
WomenClean and jerkτ=0.02\tau= 0.02−77,333.00.70.430.24−0.18167.6186.90.6
WomenClean and jerkτ=0.03\tau= 0.03−77,268.42.10.01−0.43−0.18169.0187.11.9
MenSnatchτ\tau free (τ^=0.144\hat{\tau} = 0.144)−88,676.517.1−0.43−1.090.40236.4∞\infty20.8

Table S9.

SexLiftΔℓ\Delta\ells^η\hat{s}_{\eta}ξ^\hat{\xi} withoutξ^\hat{\xi} withq0.999q_{0.999} withoutq0.999q_{0.999} withEndpoint withoutEndpoint with
MenSnatch8750.027−0.31−0.32199.2199.6206.0205.7
MenClean and jerk8490.027−0.41−0.40240.8241.4244.7245.3
WomenSnatch950.017−0.11−0.04131.5131.1152.2191.6
WomenSnatch, second mode950.017−0.24−0.22132.1132.0141.9142.7
WomenClean and jerk6150.030−0.18−0.17167.6168.9186.9189.3

Table S12. The final model with and without the shared self-assessment error, posterior modes at the reference anchor: improvement in log likelihood, scale of the shared error, and the tail shape, the one-in-a-thousand level and the endpoint at the reference bodyweight (110 kilograms for men, 86 for women) in 2026, without and with the shared error.

The anchor

The simplified model along the anchor

Table S2 gives the fits behind Figure 2 of the main text: the simplified model of § 5.2 of the main text for the men’s snatch, without the athlete factor, with one lower margin scale shared

τ\tauΔℓ\Delta\ellm^1\hat{m}_1m^2\hat{m}_2δ^2\hat{\delta}_2δ^3\hat{\delta}_3ξ^\hat{\xi}u^\hat{u}q0.999q_{0.999}Endpoint
0.010−7362.91.21.552.42−0.321.23197.8202.6
0.015−2433.61.50.951.29−0.331.26198.1202.7
0.02004.51.80.640.71−0.341.32198.5202.8
0.0303506.02.60.260.06−0.351.45200.1203.9
0.0505658.34.8−0.10−0.52−0.371.29204.1207.5
0.08063911.88.3−0.31−0.86−0.381.27211.2214.2
free (0.120)65616.613.1−0.42−1.06−0.391.25221.3224.3

Table S2. The simplified model for the men’s snatch at fixed values of the relative attempt scale: log likelihood relative to the fit at τ=0.02\tau= 0.02, modes of the margin for openers and for raises after a success in per cent of the bar, attempt-order shifts, tail shape and threshold, and the one-in-a-thousand level and endpoint of the capability distribution at 110 kilograms of bodyweight in 2026, across the three situations and with the normal-mixture body and estimated threshold of the final model, with the relative attempt scale τ\tau fixed at the listed values and every other parameter re-estimated, and the fit with τ\tau free.

Each fit is a quasi-Newton maximization of the log posterior on a bounded reparametrization. Along a ridge a single start can stop at a local mode, so each fit at fixed τ\tau is started twice, from the posterior mode of the final model and from the fit at the neighbouring value of τ\tau, outwards from τ=0.02\tau= 0.02, and the better of the two is kept. The two starts end within a tenth of a unit of log posterior of each other at every value of τ\tau. The fit with τ\tau free starts from the kept fit at τ=0.08\tau= 0.08.

A simulation study

The simulation study checks, on data with a known truth, that the final model recovers its parameters at the correct anchor, and what a wrong anchor does to them. Attempts are simulated from the final model at its posterior mode at τ=0.02\tau= 0.02 for the men’s snatch and the women’s clean and jerk, on the design of the data. Each athlete-day-lift keeps its athlete, bodyweight, year and number of attempts. Capability is drawn with the athlete factor and the opener from its margin density. After a success the next weight is a raise, with the margin truncated as in § 4.3 of the main text. After a miss the weight is repeated with the observed share of repeats and otherwise raised. Declared weights are rounded to whole kilograms. Ten data sets are simulated for each combination and refitted by the posterior mode at τ=0.02\tau= 0.02, and the first five are also refitted at τ=0.01\tau= 0.01 and 0.03. Every fit starts from the true values of the regression, body, attempt and declaration parameters, with the tail neutral, shape 0 and threshold 1.5, the centre of its prior.

Table S3 gives the results. At the correct anchor the estimates centre on the truth in the men’s snatch. There the tail shape has a standard deviation of 0.04 across replicates, the one-in-a-thousand level a bias of −0.3 kilograms and a standard deviation of 0.6, and the endpoint a standard deviation of three kilograms. The women’s clean and jerk has a weaker tail. Its shape varies twice as much and is pulled towards zero, by about half a standard deviation on average, one replicate in ten is fitted with an unbounded tail, and the endpoint ranges from 178 to 221 kilograms in the other nine, while the one-in-a-thousand level has a standard deviation of 1.9 kilograms. This matches the wide posterior of that ceiling in Table 7 of the main text. A wrong anchor acts as Proposition 1 describes. At τ=0.01\tau= 0.01 the attempt-order shifts rise by about 1.1 and 1.7 and the margin modes fall by 1.0 to 1.6 points of the bar, and at τ=0.03\tau= 0.03 they move the other way. On average the tail shape and the athlete correlation barely move, and the one-in-a-thousand level and the endpoint move by at most 1.8 kilograms. The real data behave in the same way (Table S9).

TruthFit at τ=0.02\tau= 0.02Change at τ=0.01\tau= 0.01τ=0.03\tau= 0.03
Men’s snatch
Tail shape ξ\xi−0.308−0.293 (0.040)0.041−0.001
Threshold uu1.682.00 (0.28)0.000.01
Athlete correlation ρa\rho_a0.9060.905 (0.003)−0.0040.002
Shift δ2\delta_20.660.66 (0.03)1.11−0.45
Shift δ3\delta_30.750.75 (0.05)1.80−0.68
Opener mode m1m_1 (%)3.263.24 (0.11)−1.591.44
Raiser mode m2m_2 (%)1.501.53 (0.09)−0.991.14
q0.999q_{0.999} (kg)199.2198.9 (0.6)−0.51.8
Endpoint (kg)206.0206.5 (2.9)1.31.8
Women’s clean and jerk
Tail shape ξ\xi−0.177−0.139 (0.079)0.031−0.002
Threshold uu1.591.71 (0.12)0.11−0.02
Athlete correlation ρa\rho_a0.9280.928 (0.002)−0.0060.003
Shift δ2\delta_20.430.41 (0.03)1.06−0.45
Shift δ3\delta_30.240.23 (0.04)1.68−0.67
Opener mode m1m_1 (%)0.730.73 (0.08)−1.101.31
Raiser mode m2m_2 (%)0.130.18 (0.21)−1.201.61
q0.999q_{0.999} (kg)167.6167.8 (1.9)−0.61.6
Endpoint (kg)186.9190.8{sup}*`0.5{sup}†`1.5{sup}†`

Table S3. Simulation study. True values (the posterior mode at τ=0.02\tau= 0.02), mean and standard deviation of the estimates over ten simulated data sets refitted at τ=0.02\tau= 0.02, and mean change when the first five are refitted at τ=0.01\tau= 0.01 and 0.03, against the fit at τ=0.02\tau= 0.02 to the same data. Quantiles and endpoints at the reference bodyweight in 2026. *Median, one replicate has an unbounded tail. †Over the replicates with a bounded tail in both fits.

The tail without a body

Table S4 gives the check without a body used in the last row of Table 5 of the main text. The threshold u(x)u(x) is the qq-quantile of the best made lift of an athlete-day, estimated by linear quantile regression on the design of the main text (intercept, standardized log bodyweight, its orthogonal square and centred year). The exceedances y−u(x)y-u(x) follow a generalized Pareto distribution with a log scale linear in standardized log bodyweight and a constant shape, fitted by maximum likelihood treating the athlete-days as independent, and the intervals come from 1,000 bootstrap samples that resample athletes with all their competitions, which carries the dependence between the competitions of the same athlete into the intervals. The endpoint at covariate xx is u(x)+σ(x)/(−ξ)u(x)+\sigma(x)/(-\xi). The shape is negative at both thresholds in all four combinations. Between the two thresholds the endpoint moves by at most four kilograms for men and in the women’s snatch, and by seven in the women’s clean and jerk, whose shape falls from −0.21-0.21 to −0.31-0.31. At the higher threshold the shape is estimated from about five hundred exceedances, and its bootstrap interval widens to about 0.3. At the lightest women’s bodyweights the endpoint falls at or below the record (at 59 kilograms, 116 against a snatch record of 117 at the lower threshold and 149 against a clean and jerk record of 149 at the higher), a sign that a scale linear in log bodyweight is too rigid at the edge of the range.

SexLiftqqExceedancesξ^\hat{\xi}EndpointRecord
MenSnatch0.901,007−0.24-0.24 (−0.34-0.34, −0.16-0.16)212200
0.95502−0.21-0.21 (−0.39-0.39, −0.08-0.08)212
MenClean and jerk0.90965−0.22-0.22 (−0.31-0.31, −0.15-0.15)255242
0.95479−0.20-0.20 (−0.34-0.34, −0.08-0.08)259
WomenSnatch0.90933−0.32-0.32 (−0.43-0.43, −0.20-0.20)140127
0.95468−0.35-0.35 (−0.51-0.51, −0.23-0.23)140
WomenClean and jerk0.90917−0.21-0.21 (−0.33-0.33, −0.08-0.08)187161
0.95456−0.31-0.31 (−0.46-0.46, −0.19-0.19)180

Table S4. Generalized Pareto fits without a body: threshold quantile, number of exceedances, tail shape with athlete-bootstrap 95 per cent interval, endpoint at the reference bodyweight (110 kilograms for men, 86 for women) in 2026, and the local record of the current era.

Fixed and estimated threshold

Table S5 fits the record book, the best made lifts with the athlete-days exchangeable, with the threshold of the mixture body held at the 0.85, 0.90 and 0.95 quantiles of the standardized error and every other parameter re-estimated. The fits are by maximum likelihood with the density of each best lift, whereas the record-book rows of Table 5 of the main text are posterior summaries that use the probability of the one-kilogram interval around each lift and the priors of the full model. The ceiling moves with the threshold, by up to four kilograms in the men’s lifts and the women’s snatch and by ten in the women’s clean and jerk. A threshold fixed in advance carries an uncertainty that intervals at that threshold do not show. In the latent model the threshold is a parameter, and its posterior carries this uncertainty into every tail quantity.

Men, snatchMen, c&jWomen, snatchWomen, c&j
Threshold at 0.85ξ^\hat{\xi}−0.24-0.24−0.23-0.23−0.23-0.23−0.20-0.20
ceiling212255144188
Threshold at 0.90ξ^\hat{\xi}−0.24-0.24−0.23-0.23−0.28-0.28−0.22-0.22
ceiling212255142185
Threshold at 0.95ξ^\hat{\xi}−0.33-0.33−0.26-0.26−0.29-0.29−0.33-0.33
ceiling208253141178

Table S5. Tail shape and ceiling at the reference bodyweight in 2026 for the record book, the best made lifts with the athlete-days exchangeable, with the threshold of the mixture body held at three quantiles of the standardized error.

The threshold is harder to place in the record-book fits of Table 5 of the main text than in the full model. In the record-book fit of the women’s snatch with the athlete-days exchangeable, the posterior of the threshold is wide and the chains explore it slowly, with an effective sample size of 82 for the threshold against 128 for the shape (Table S13). At the higher thresholds the shape is estimated from a few hundred exceedances and its posterior spreads towards zero, which is why that row of Table 5 gives a finite ceiling with probability 0.84 only. In the full model the women’s snatch threshold is 1.47 (0.94, 2.48), an interval that spans the two regions of its posterior (§ 16).

SexLiftFitAcceptanceMax R^\hat{R}Min ESSMedian ESSξl\xi_lulu_lEndpointR^\hat{R} endpoint
MenSnatchFinal model0.051.031674904031804891.01
MenSnatchRecord book0.031.031483292551482771.02
MenSnatchRecord book, athlete factor0.031.021823875122455291.01
MenClean and jerkFinal model0.041.031123832521174171.03
MenClean and jerkRecord book0.031.031202988181878811.01
MenClean and jerkRecord book, athlete factor0.021.47882159231431.05
WomenSnatchFinal model, joint set0.011.24259855351571.06
WomenSnatchFinal model, higher threshold (0.22)0.021.18875473301843831.05
WomenSnatchFinal model, lower threshold (0.78)0.021.331033104172425531.05
WomenSnatchRecord book0.021.291189128821591.02
WomenSnatchRecord book, athlete factor0.021.2415111317373481.00
WomenClean and jerkFinal model0.021.05884982051742441.03
WomenClean and jerkRecord book0.021.41932183312561.01
WomenClean and jerkRecord book, athlete factor0.021.11361331541331711.02

Table S13. Convergence of the chains: acceptance rate, largest R^\hat{R} over the parameters, smallest and median bulk effective sample size over the parameters, and the bulk effective sample size of the tail shape, the threshold and the endpoint at the reference bodyweight in 2026, with the R^\hat{R} of the endpoint, for the final model and for the record-book fits without and with the athlete factor. Four chains of 4,500 draws for the final model (16,500 for the women’s lifts) and of 3,000 for the record-book fits. For the women’s snatch the row for the third set of eight chains, which moves between the two regions of the posterior, is followed by one row per region, with the weight of the region in parentheses. The rows for the regions use the draws of every chain in the region and compute R^\hat{R} over the chains with at least 1,000 draws there, cut to a common length.

Model checks and sensitivity analyses

Table S6 compares the observed success rate of each attempt with the rate predicted by the final model from everything observed before its outcome, the earlier weights and outcomes and the weight just declared, with the parameters at their posterior mode and the athlete factor switched off, so that the prediction uses the day alone. The rates by attempt agree to within 0.035 in every combination, and the rates by size of raise are in Figure 1 of the main text.

SexLiftAttemptObservedPredicted
MenSnatch10.7710.761
MenSnatch20.6330.619
MenSnatch30.4370.434
MenClean and jerk10.7950.784
MenClean and jerk20.6010.588
MenClean and jerk30.3480.347
WomenSnatch10.7720.756
WomenSnatch20.6510.623
WomenSnatch30.4660.455
WomenClean and jerk10.8260.818
WomenClean and jerk20.6320.598
WomenClean and jerk30.3720.362

Table S6. Observed and predicted success rates by sex, lift and attempt number, the prediction for each attempt computed at the posterior mode from the weights and outcomes observed before it. For the women’s snatch the mode is that of the region with the larger posterior mass.

Table S7 checks the record book itself. For 200 draws from the posterior, complete competitions are simulated for every athlete-day-lift of the data as in § 12.2, and the local record of the current era is recomputed at each bodyweight of Table 6 of the main text. The observed record lies within the central 90 per cent of the simulated records at 36 of the 40 combinations. The men’s snatch record at 61 kilograms lies one kilogram below that range, and the records of the heaviest women, in the snatch at 100 kilograms and in the clean and jerk at 100 and 120, lie two to five kilograms above it. On the scale of capability, (best−μ)/ω(\mathrm{best} - \mu)/\omega, the 0.999 quantile of the best made lifts is reproduced in three of the four combinations, but the 0.99 quantile is higher than in at least 95 per cent of the simulated data sets in all four. Among the top one per cent of best lifts, 0.76 to 0.84 come from a day that ended with a made attempt, against 0.54 to 0.64 in the simulated competitions, and at most one simulated data set in two hundred reaches the observed share. Under the model the outcome of an attempt depends on its margin and not on the level of capability, so the best days of the strongest lifters end with a miss as often as anyone’s. In the data they end with a make more often. The model therefore reads some of these made final attempts as luck, which pulls the inferred capability of the strongest days down.

BodyweightSnatchClean and jerk
RecordSimulatedRecordSimulated
Men
61146151 (147, 157)183184 (179, 190)
67160161 (156, 168)197195 (190, 202)
73169168 (163, 174)209203 (198, 211)
81182177 (172, 182)211214 (209, 221)
89188187 (182, 193)229225 (221, 233)
96188192 (187, 199)233231 (226, 238)
102190195 (190, 203)233236 (230, 243)
110200200 (194, 207)242241 (235, 248)
130204204 (195, 214)250248 (238, 257)
150217211 (201, 221)253255 (243, 265)
Women
4999103 (99, 109)126129 (123, 138)
55111111 (107, 117)141139 (134, 146)
59117117 (113, 122)149145 (141, 154)
64120119 (115, 126)151149 (143, 156)
71125124 (120, 132)156156 (150, 165)
76125128 (123, 135)161160 (152, 170)
81127130 (125, 136)161163 (155, 171)
86127130 (125, 137)161163 (156, 172)
100145131 (122, 140)178164 (155, 176)
120140134 (125, 146)185170 (159, 182)

Table S7. Predictive check of the record book. Local record of the current era (the heaviest lift made since 2019 within five kilograms of the bodyweight) and the median and 90 per cent interval of the same record in competitions simulated from 200 posterior draws, in kilograms.

Table S8 gives the margin of a raise after a missed attempt, which is left out of Table 3 of the main text. Such a raise is truncated above at log⁡{Z/(ωj−1+1)}\log\{Z/(\omega_{j-1}+1)\}, and after a miss this point is usually near or below zero, so the observed margins come from the part of the two-piece normal below its mode. The data determine only the shape of the density below the truncation point, which the mode and the two scales describe together, so the mode is not a typical margin and the three parameters are barely separable. In three of the four combinations the mode lies far above the truncation point, against the upper end of its support at 30 per cent (§ 4.8 of the main text), and the upper scale is barely determined, as its intervals show. In the women’s snatch the mode lies just below zero, with the lower scale at the lower end of its support. About three athletes in four repeat the weight after a miss, and the choice between repeating and raising is not modelled.

Men, snatchMen, c&jWomen, snatchWomen, c&j
Raise after a miss mode24.4 (15.5, 29.7)28.7 (24.4, 29.9)−0.5 (−0.7, −0.3)27.2 (20.5, 29.9)
upper scale3.1 (0.5, 20.8)2.9 (0.5, 21.5)1.2 (1.0, 1.6)2.9 (0.5, 19.9)
lower scale4.2 (3.5, 4.7)5.0 (4.6, 5.2)0.3 (0.3, 0.3)4.8 (4.2, 5.1)

Table S8. Posterior medians and 95 per cent intervals for the margin of a raise after a missed attempt at the reference anchor τ=0.02\tau= 0.02, in per cent of the bar.

Table S9 reports posterior modes of the final model for the four combinations of sex and lift at the three anchors, for the women’s snatch at both modes of § 16, and with τ\tau left free for the men’s snatch. Between τ=0.01\tau= 0.01 and 0.030.03 the tail shape moves by at most 0.05, the one-in-a-thousand level by at most 4.5 kilograms and the endpoint by at most five, except at the higher-threshold mode of the women’s snatch, where the shape moves from −0.11-0.11 to −0.05-0.05 at τ=0.01\tau= 0.01 and the endpoint from 152 to 185 kilograms. At the other mode, which holds most of the mass, the endpoint lies between 142 and 146 kilograms at the three anchors. The attempt-order shifts take up most of the change and fall in every combination as τ\tau grows. The mode of the opener margin rises between τ=0.01\tau= 0.01 and 0.02, roughly doubling in three of the four combinations. Between 0.02 and 0.03 it keeps rising in the clean and jerk, changes little in the men’s snatch and falls in the women’s snatch, where the lower scale of the opener margin shrinks with it along a direction that the data barely resolve. For that reason each anchor fit of the women’s snatch was also started from the margins of the fit at the other mode, and the table reports the higher of the two. With τ\tau free the fit moves far along the direction of Proposition 1, with the margins, the upper quantiles and the endpoint moving together. The optimizer ends at τ=0.144\tau= 0.144 with the tail shape at the upper limit of its support, the last row of the table.

Table S10 checks the priors of the tail. All posterior draws of each combination are reweighted by the ratio of an alternative prior to the original one, which gives the posterior under the alternative without new chains. The alternatives are a wider prior for the tail shape, ξl∼N(0,0.52)\xi_l \sim N(0, 0.5^2), a wider prior for the threshold, ul∼N(1.5,22)u_l \sim N(1.5, 2^2) on the same support, and a threshold bounded below 3. The effective sample size of the weights is at least 75 per cent of the number of draws. The median ceiling at the reference bodyweight moves by at most 1.1 kilograms and the one-in-a-thousand level by at most 0.1. The probability of a finite ceiling moves by at most 0.02, except in the women’s clean and jerk under the wider prior for the threshold, where it falls from 0.94 to 0.89, as that prior favours higher thresholds with fewer lifts above them. The upper limit of the interval for the women’s ceilings is set by the share of draws with an unbounded tail, and under the wider prior for the threshold it is infinite in both lifts.

SexLiftPriorESSξ\xiuuq0.999q_{0.999}P(finite)P(\mathrm{finite})Ceiling
MenSnatchReference18,000−0.31-0.311.75199.31.000206.1 (203.3, 209.9)
ξl∼N(0,0.52)\xi_l \sim N(0,0.5^2)17,928−0.31-0.311.75199.31.000206.0 (203.2, 209.8)
ul∼N(1.5,22)u_l \sim N(1.5,2^2)17,671−0.31-0.311.80199.31.000206.1 (203.3, 209.9)
ul<3u_l < 317,997−0.31-0.311.75199.31.000206.1 (203.3, 209.9)
MenClean and jerkReference18,000−0.40-0.402.00241.01.000245.0 (243.0, 247.3)
ξl∼N(0,0.52)\xi_l \sim N(0,0.5^2)17,938−0.41-0.412.00240.91.000244.9 (243.0, 247.2)
ul∼N(1.5,22)u_l \sim N(1.5,2^2)14,373−0.40-0.402.22241.01.000245.0 (243.1, 247.2)
ul<3u_l < 316,875−0.41-0.411.94241.01.000245.0 (243.0, 247.3)
WomenSnatchReference43,326−0.23-0.231.47132.10.955142.8 (139.0, 418.1)
ξl∼N(0,0.52)\xi_l \sim N(0,0.5^2)43,079−0.23-0.231.46132.10.960142.6 (138.8, 241.5)
ul∼N(1.5,22)u_l \sim N(1.5,2^2)42,682−0.22-0.221.48132.10.945143.0 (139.0, ∞\infty)
ul<3u_l < 343,314−0.23-0.231.47132.10.955142.8 (139.0, 418.1)
WomenClean and jerkReference66,000−0.18-0.181.63167.80.941187.1 (178.3, ∞\infty)
ξl∼N(0,0.52)\xi_l \sim N(0,0.5^2)65,203−0.18-0.181.63167.80.941186.6 (178.0, ∞\infty)
ul∼N(1.5,22)u_l \sim N(1.5,2^2)49,337−0.17-0.171.71167.90.888188.2 (178.3, ∞\infty)
ul<3u_l < 363,523−0.18-0.181.62167.70.957186.9 (178.3, 359.3)

Table S10. Prior sensitivity of the tail by reweighting of the posterior draws: effective sample size of the weights, posterior medians of the tail shape, the threshold and the one-in-a-thousand level, posterior probability of a finite ceiling, and posterior median and 95 per cent interval of the ceiling, infinite draws included, at the reference bodyweight (110 kilograms for men, 86 for women) in 2026.

Table S11 adds a linear term in the centred year to the log scale of capability in the record-book fit of the first row of Table 5 of the main text, at its posterior mode, whose values differ from the posterior medians of that table. The term is negligible. Its coefficient is under 0.3 per cent a year in absolute value, the log likelihood improves by at most 1.6 units, and the one-in-a-thousand levels move by under half a kilogram. The endpoint of the women’s snatch moves by 7.5 kilograms, because the shape of that fit is weakly determined (§ 14).

SexLiftΔℓ\Delta\ellγ^13\hat{\gamma}_{13}ξ^\hat{\xi} withoutξ^\hat{\xi} withq0.999q_{0.999} withoutq0.999q_{0.999} withEndpoint withoutEndpoint with
MenSnatch0.5−0.0015−0.31−0.31202.3202.2208.6208.4
MenClean and jerk1.20.0026−0.21−0.21243.9244.2256.3257.0
WomenSnatch0.60.0008−0.14−0.09133.8133.8148.4155.9
WomenClean and jerk1.6−0.0004−0.26−0.28171.7171.3183.5181.6

Table S11. The record-book fit with and without a linear time term in the log scale: improvement in log likelihood, coefficient of the centred year in the log scale, and the tail shape, one-in-a-thousand level and endpoint at the reference bodyweight in 2026 without and with the term.

Two variants of the attempt model were also fitted at the reference anchor. The first adds a shared self-assessment error to the final model. The athlete’s estimate of capability on the day is ZieηiZ_i e^{\eta_i}, with ηi∼N(0,sη2)\eta_i \sim N(0,s_{\eta}^{2}) common to the day’s declarations and log⁡sη∼N(log⁡0.02,1)\log s_{\eta} \sim N(\log0.02,1) on (0.002, 0.3). The margins are taken relative to that estimate. Table S12 compares the posterior modes with and without it for the four combinations of sex and lift, for the women’s snatch at both modes. The shared error is between 1.7 and 3.0 per cent of capability and improves the log likelihood by 95 units in the women’s snatch and by 600 to 900 units in the other three combinations. It takes up part of the attempt-order shifts, which fall by 0.03 to 0.24, and leaves the athlete correlation within 0.01 and the one-in-a-thousand level at the reference bodyweight within 1.3 kilograms. The tail shape moves by at most 0.02 and the endpoint by at most 2.4 kilograms, except at the higher-threshold mode of the women’s snatch, where the shape moves from −0.11 to −0.04 and the endpoint from 152 to 192 kilograms. That mode holds a fifth of the posterior mass (§ 16), and its endpoint responds to any change that moves the shape towards zero, as it does at the anchors (Table S9). At the mode that holds the rest of the mass the shape moves from −0.24 to −0.22 and the endpoint by 0.8 kilograms. The shared error does not change the success curve by size of raise, because the size of a raise is the ratio of two declarations made relative to the same estimate and so does not depend on the shared error. Each likelihood evaluation costs about twenty times as much as for the final model, which puts the chains out of reach, so the shared error is left out of the final model. The second variant, fitted to the men’s snatch with the normal-mixture body and without the athlete factor, puts a floor on the miss probability, P(S=1∣Z,w)=(1−π)F{log⁡(Z/w)/τ}P(S = 1 \mid Z,w) = (1-\pi)F\{\log(Z/w)/\tau\}. It improves the log likelihood by 819 units for one parameter, but does so by moving along the direction of Proposition 1, with π^=0.32\hat{\pi} = 0.32, openers declared 12 per cent below capability, the order shifts at −0.6 and −2.5 and the endpoint at 223 kilograms, against 203 without the floor. The floor and the shift of the capability scale are indistinguishable to the data, and the variant is not used.

A linear trend in time is kept in the location. Its coefficient is 0.32 and 0.36 kilograms a year for the men’s two lifts and 0.60 and 0.79 for the women’s (Table 4 of the main text).

Convergence of the chains

The final model was sampled, for each sex and lift, by four differential-evolution Markov chains with sampling from the past and snooker updates ((ter Braak and Vrugt, 2008)). The chains work in the coordinates of the main text with the threshold-stable tail scale σl∗=σl−ξlul\sigma_l^* = \sigma_l - \xi_l u_l, which does not move with the threshold ([7 Section 4.3.4]). The change of variables is exact, since the target density carries its Jacobian. The sampler was chosen because the threshold, the tail scale, the free mixture components and the regression coefficients lie along curved ridges of the posterior, on which a Gaussian random walk gave effective sample sizes of about twenty in 17,000 draws. Each chain ran for 120,000 generations, the first 30,000 were discarded and every twentieth state of the rest was kept, 4,500 draws per chain, or 360,000 generations and 16,500 draws for the two women’s lifts. The shared archive of past states was extended every ten generations. The initial archive and the starting states were drawn around the posterior mode with four times the variance of its Laplace approximation and with the threshold spread over half a scale unit on either side. The record-book fits of Table 5 of the main text were sampled in the same way with 200,000 generations, 50,000 discarded and every fiftieth state kept.

The posterior mode is the best of three optimizations. One starts from a fit to the best lifts and one from a high threshold with a positive shape. The third moves the first start along the ridge that links the scale of capability to the free mixture components. For the men’s lifts and the women’s clean and jerk the three optimizations end at the same mode or within 1.6 units of log posterior of it. For the women’s snatch they end in two regions. One has a threshold of 2.2 and a shape of −0.11. The other, 4.7 units of log posterior higher, has a narrower scale of capability, a threshold of 1.4 and a shape of −0.24. The replication package stores it as a start, because the optimizer reaches it only in some runs. Four chains were run around each mode, for 360,000 generations at the higher-threshold mode and 120,000 at the other. Every retained draw is assigned to the nearer of the two modes in the threshold, the log scale and the shape, each divided by its standard deviation within a chain. These chains cross between the regions rarely, so the share of each region in their pooled draws would depend on where they started. A third set of eight chains therefore starts from the last states of these eight, with an archive of 3,200 draws taken from both sets in equal numbers. Every second proposal is a jump of full length along the difference of two archive states, which can carry a chain from one region to the other. The proposals are symmetric, so the set has the posterior as its stationary distribution whatever the archive holds. After the first tenth of its 40,000 generations it crossed between the regions 41 times, about once every 7,000 generations per chain. The share of its draws in each region estimates the posterior mass of the region, 0.78 for the region of the narrow mode, with a standard error of 0.04 from ten batches of each chain. The draws of each region from all sixteen chains are pooled with these weights. The regions agree on the one-in-a-thousand level, 132.2 against 131.8 kilograms at the reference bodyweight, but differ on the endpoint, 142 kilograms against 158 with 21 per cent of the draws at shapes at or above zero. The checks at the mode in § 15 are made at both modes.

Table S13 gives the acceptance rates with the rank-normalized split R^\hat{R} and the bulk effective sample sizes of Vehtari et al. [43], for the women’s snatch within each region. In the men’s fits and the women’s clean and jerk, R^\hat{R} is at most 1.03 for the tail shape, the athlete correlation, the attempt and declaration parameters, the time trend and the derived quantities. It reaches 1.05 only among the means, scales and weights of the two free mixture components, which are exchangeable and not what the paper reports. Within the two regions of the women’s snatch the tail shape, the endpoint and the one-in-a-thousand level have R^\hat{R} at most 1.11 and effective sample sizes of at least 330, while the parameters of the body reach 1.18 and 1.33 along the ridge in the threshold and the log scale. The women’s clean and jerk puts six per cent of its mass on shapes at or above zero and crosses into that region rarely, which is why it needed the longer chains. The Monte Carlo standard error of the posterior median of the endpoint at the reference bodyweight is about a tenth of a kilogram for the men’s lifts and half a kilogram for the women’s clean and jerk. Within the narrow region of the women’s snatch it is 0.2 kilograms, and the pooled median moves by less than 0.2 kilograms when the weight of that region moves by one standard error. That of the one-in-a-thousand level is under a tenth of a kilogram in all four combinations. The replication package produces trace plots of both quantities. Of the record-book fits, the one with the athlete factor for the men’s clean and jerk has chains that disagree on the threshold and the location coefficient, with R^\hat{R} up to 1.47, so its entries in Table 5 of the main text are indicative only. In the women’s record-book fits the largest values of R^\hat{R}, up to 1.41, again belong to the mixture components, while the shape and the endpoint have R^≤1.03\hat{R} \le1.03. The threshold of the women’s snatch fit with the athlete factor has an effective sample size of 37.

The whole archive

Every analysis of the main text was repeated on the whole archive, the 78,386 athlete-day-lifts since 1998, with the same model, priors, anchors and sampler settings. Table S14 compares the final model on the two datasets and Table S15 repeats Table 5 of the main text on the whole archive.

SexLiftDatannξl\xi_lP(ξl<0)P(\xi_l < 0)q0.999q_{0.999}Endpointρa,l\rho_{a,l}Trendδ2\delta_2δ3\delta_3
MenSnatchsince 201710,492−0.31 (−0.36, −0.25)1.00199.32060.9060.320.660.75
since 199822,272−0.33 (−0.37, −0.24)1.00204.22100.9100.060.640.74
MenClean and jerksince 201710,179−0.40 (−0.44, −0.36)1.00241.02450.8950.360.370.15
since 199821,622−0.40 (−0.43, −0.35)1.00245.62500.8970.040.330.09
WomenSnatchsince 20179,695−0.23 (−0.29, 0.05)0.96132.11430.9210.600.821.06
since 199817,395−0.22 (−0.28, −0.05)1.00142.01570.9040.350.831.10
WomenClean and jerksince 20179,517−0.18 (−0.28, 0.14)0.94167.81870.9290.790.430.24
since 199817,097−0.21 (−0.42, 0.19)0.88178.41960.9100.450.410.27

Table S14. Table S14. The final model on the athlete-days since 2017 (the analysis of the main text) and on the whole archive since 1998: number of athlete-day-lifts, posterior median and 95 per cent interval of the tail shape, posterior probability of a finite endpoint, posterior medians of the one-in-a-thousand level and of the endpoint at the reference bodyweight in 2026, and the posterior medians of the athlete correlation, the time trend of the location in kilograms a year and the attempt-order shifts.

Men, snatch (record 200)Men, c&j (record 242)Women, snatch (record 127)Women, c&j (record 161)
Record book
ξl\xi_l−0.24 (−0.26, −0.20)−0.22 (−0.25, −0.18)−0.09 (−0.20, 0.03)−0.30 (−0.34, −0.25)
ceiling212 (1.00)256 (1.00)174 (0.00)182 (0.00)
Record book, athlete factor
ξl\xi_l−0.22 (−0.26, −0.15)−0.20 (−0.25, −0.08)−0.13 (−0.23, 0.03)−0.22 (−0.27, −0.17)
ceiling227 (0.01)276 (0.01)175 (0.00)202 (0.00)
Full model
ξl\xi_l−0.33 (−0.37, −0.24)−0.40 (−0.43, −0.35)−0.22 (−0.28, −0.05)−0.21 (−0.42, 0.19)
ceiling210 (0.99)250 (1.00)157 (0.00)196 (0.00)
No body (check)
ξl\xi_l−0.24 (−0.30, −0.19)−0.23 (−0.28, −0.18)−0.30 (−0.36, −0.23)−0.33 (−0.41, −0.24)
ceiling209253146179

Table S15. Table S15. Table 5 of the main text on the whole archive: posterior median and 95 per cent interval of the tail shape, and the posterior median of the endpoint at the reference bodyweight in 2026 with, in parentheses, the posterior probability that it lies within ten per cent of the local record of the current era. The last row uses no body, with the point estimate of the endpoint.

The attempt and declaration parameters and the athlete correlation barely change, and neither do the men’s tails, which are bounded on both datasets with shapes within 0.02 of each other. The whole archive places the men’s ceilings at 110 kilograms of bodyweight four and five kilograms higher. The time trend of the location is much smaller on the whole archive, 0.04 to 0.06 kilograms a year for men against 0.32 and 0.36 on the current era and 0.35 and 0.45 for women against 0.60 and 0.79, so a single linear trend over the whole period does not describe the rise within the current era. For women the whole archive gives nearly the same shapes, −0.22 against −0.23 in the snatch and −0.21 against −0.18 in the clean and jerk, with a wider interval for the clean and jerk, a probability of a finite ceiling of 1.00 in the snatch and 0.88 in the clean and jerk, and ceilings at 86 kilograms of 157 and 196 kilograms, against 143 and 187 on the current era. At that bodyweight the heaviest women’s lifts of all years, 138 kilograms in the snatch and 166 in the clean and jerk, were made in 2009 and 2012, and the model carries the tail they imply to 2026 through the trend in the location. The women’s clean and jerk tail on the whole archive depends more on the anchor. At τ=0.01\tau= 0.01 the posterior mode of its shape is −0.03-0.03 and that of its endpoint 336 kilograms, against 196 and 198 at τ=0.02\tau= 0.02 and 0.03. At the lower-threshold mode of the women’s snatch the endpoint lies between 153 and 157 kilograms at the three anchors.

On the whole archive the record book with the athlete-days exchangeable gives men’s ceilings of 212 and 256 kilograms, two and six kilograms above the full model, and with the athlete factor 227 and 276, with the tail still bounded. The ordering of the layers is the same as on the current era, and the athlete factor again widens the record-book tail. The copula gives ρ=0.96\rho= 0.96 for men and 0.97 for women, with ψ=0.15\psi= 0.15 for both. The chains of the final model on the whole archive mix as on the current era for men, with R^≤1.06\hat{R} \le1.06, and more slowly for women. In the women’s clean and jerk R^\hat{R} reaches 1.10 and the endpoint has an effective sample size of 83. The women’s snatch has the two regions of § 16. A joint set of four chains, run as there, gives 0.95 of the mass, with a standard error of 0.02, to the lower-threshold region, where the mixture components reach R^=1.23\hat{R} = 1.23 and the endpoint has an effective sample size of 138. The higher-threshold region holds 0.05 of the mass. Only two chains spent long stretches in it and they disagree within it, with R^\hat{R} of 1.5 for the shape, so its 939 draws among the 17,423 pooled ones are the least reliable part of the whole-archive posterior. Without them the median ceiling of the women’s snatch at 86 kilograms moves from 157.2 to 156.9 kilograms and the probability of a finite ceiling from 0.996 to 1.

The copula stage

The joint likelihood of the paired athlete-days was evaluated on the grid

ρ∈{0.88,0.90,0.92,0.93,0.94,0.95,0.96,0.97,0.98},ψ∈{0,0.05,0.10,0.15,0.20,0.25,0.30,0.40,0.50}\begin{aligned} \rho&\in\{0.88, 0.90, 0.92, 0.93, 0.94, 0.95, 0.96, 0.97, 0.98\}, \\ \psi&\in\{0, 0.05, 0.10, 0.15, 0.20, 0.25, 0.30, 0.40, 0.50\} \end{aligned}

at five posterior draws of the marginal parameters of each lift, with the athlete factor switched off, and the posterior over the grid averaged across the draws. For men the posterior mass sits on ρ=0.95\rho= 0.95 and 0.960.96 with ψ=0.15\psi= 0.15 or 0.100.10, and for women on ρ=0.97\rho= 0.97 with ψ=0.10\psi= 0.10 or 0.050.05. The Gaussian copula (ψ=0\psi= 0) has posterior probability 5×10−125 \times10^{-12} for men and 10−510^{-5} for women. The grid is coarse by design, since each evaluation integrates the two capabilities of 10,174 or 9,511 athlete-days over the product of the two quadrature grids. The return level of the men’s total at 110 kilograms of bodyweight is 439.7 kilograms (438.1, 442.0) at a horizon of one in a thousand athlete-days and 445.8 (442.7, 449.0) at one in ten thousand, against 440.1 and 446.1 for the sum of the two marginal levels on the same 40 posterior draws. For women at 86 kilograms the figures are 299.6 against 300.0 and 310.4 against 310.7. Computed draw by draw, the median deficit of the total at one in ten thousand is 0.3 kilograms (−0.4, 0.8) for men and 0.4 (−1.6, 2.4) for women.

Proof of Proposition 1

Proof of Proposition 1. Drop the superscript ll, write primes for quantities under TΔθlT_\Delta\theta_l and put z′=eΔzz' = e^\Delta z. Since μl\mu_l and ωl\omega_l are multiplied by eΔe^\Delta and GlG_l is unchanged, gi′(z′)=e−Δgi(z)g'_i(z') = e^{-\Delta}g_i(z) and Gi′(z′)=Gi(z)G'_i(z') = G_i(z), so the athlete term κ\kappa in (12) is unchanged. The margin log⁡(z′/wij)\log(z'/w_{ij}) and the truncation point log⁡{z′/(wi,j−1+1)}\log\{z'/(w_{i,j-1}+1)\} exceed their values at zz by Δ\Delta, as do the margin modes, while the margin scales are unchanged, so every factor dkd_k and DkD_k in (11) is unchanged. Finally, ηij′(z′)=ηij(z)+Δ/τ1+δj′−δj\eta'_{ij}(z') = \eta_{ij}(z) + \Delta/\tau_1 + \delta'_j - \delta_j, which equals ηij(z)\eta_{ij}(z) for j=2,3j = 2, 3 and ηi1(z)+Δ/τ1\eta_{i1}(z) + \Delta/\tau_1 for j=1j = 1. Changing the variable of the inner integral in (12) from z′z' to zz, the factor eΔe^\Delta in dz′\mathrm{d}z' cancels the factor e−Δe^{-\Delta} in gi′g'_i.

References

  1. [1]Karin Aarssen and Laurens de Haan. On the maximal life span of humans. Mathematical Population Studies, 4:259–281, 1994.pubmed.ncbi.nlm.nih.gov/12318734
  2. [2]M. B. Adam and J. A. Tawn. Bivariate extreme analysis of Olympic swimming data. Journal of Statistical Theory and Practice, 6:510–523, 2012.DOI
  3. [3]Erling B. Andersen. Asymptotic properties of conditional maximum-likelihood estimators. Journal of the Royal Statistical Society, Series B, 32:283–301, 1970.
  4. [4]Cibele N. Behrens, Hedibert F. Lopes, and Dani Gamerman. Bayesian analysis of extreme events with threshold estimation. Statistical Modelling, 4:227–244, 2004.DOI
  5. [5]Jan Beirlant, Armelle Guillou, Goedele Dierckx, and Amélie Fils-Villetard. Estimation of the extreme value index and extreme quantiles under random censoring. Extremes, 10:151–174, 2007.DOI
  6. [6]Geoffroy Berthelot, Valérie Thibault, Muriel Tafflet, Sylvie Escolano, Nour El Helou, Xavier Jouven, Olivier Hermine, and Jean-François Toussaint. The citius end: world records progression announces the completion of a brief ultra-physiological quest. PLoS ONE, 3:e1552, 2008.
  7. [7]Stuart Coles. An Introduction to Statistical Modeling of Extreme Values. Springer, London, 2001.DOI
  8. [8]Stuart Coles, Janet Heffernan, and Jonathan Tawn. Dependence measures for extreme value analyses. Extremes, 2:339–365, 1999.DOI
  9. [9]John S. Croucher. An analysis of world weightlifting records. Research Quarterly for Exercise and Sport, 55:285–288, 1984.DOI
  10. [10]A. C. Davison and R. L. Smith. Models for exceedances over high thresholds. Journal of the Royal Statistical Society, Series B, 52:393–442, 1990.DOI
  11. [12]John H. J. Einmahl and Yi He. Accurate estimates of ultimate 100-meter records. Extremes, 29:671–685, 2026.arxiv.org/abs/2502.04085
  12. [13]John H. J. Einmahl and Jan R. Magnus. Records in athletics through extreme-value theory. Journal of the American Statistical Association, 103:1382–1391, 2008.
  13. [14]John H. J. Einmahl and Sander G. W. R. Smeets. Ultimate 100-m world records through extreme-value theory. Statistica Neerlandica, 65:32–42, 2011.DOI
  14. [15]John H. J. Einmahl, Amélie Fils-Villetard, and Armelle Guillou. Statistics of extremes under random censoring. Bernoulli, 14:207–227, 2008.
  15. [16]Lincoln E. Ford, A. J. Detterline, K. K. Ho, and W. Cao. Gender- and height-related limits of muscle strength in world weightlifting champions. Journal of Applied Physiology, 89:1061–1064, 2000.
  16. [17]Isabel Fraga Alves, Cláudia Neves, and Pedro Rosário. A general estimator for the right endpoint with an application to supercentenarian women’s records. Extremes, 20:199–237, 2017.
  17. [18]Jozo Grgic, Bruno Lazinica, Brad J. Schoenfeld, and Zeljko Pedisic. Test-retest reliability of the one-repetition maximum (1RM) strength assessment: a systematic review. Sports Medicine – Open, 6:31, 2020.DOI
  18. [19]Jim E. Griffin, Laurentiu C. Hinoveanu, and James G. Hopker. Bayesian modelling of elite sporting performance with large databases. Journal of Quantitative Analysis in Sports, 18:253–267, 2022.DOI
  19. [20]Jim E. Griffin, Maria-Zafeiria Spyropoulou, and James G. Hopker. Modelling between- and within-season trajectories in elite athletic performance data. Journal of the Royal Statistical Society, Series C, 2026. Article qlag002.arxiv.org/abs/2405.17214
  20. [21]Paul Gustafson. Bayesian Inference for Partially Identified Models: Exploring the Limits of Limited Data. Chapman and Hall/CRC, Boca Raton, 2015.
  21. [22]Janet E. Heffernan and Jonathan A. Tawn. A conditional approach for multivariate extreme values. Journal of the Royal Statistical Society, Series B, 66:497–546, 2004.
  22. [23]Shane T. Jensen, Kenneth E. Shirley, and Abraham J. Wyner. Bayesball: a Bayesian hierarchical model for evaluating fielding in major league baseball. The Annals of Applied Statistics, 3:491–520, 2009.DOI
  23. [24]Alexander Kolliari-Turner, Brian Oliver, Giscard Lima, John P. Mills, Guan Wang, Yannis Pitsiladis, and Fergus M. Guppy. Doping practices in international weightlifting: analysis of sanctioned athletes/support personnel from 2008 to 2019 and retesting of samples from the 2008 and 2012 Olympic Games. Sports Medicine – Open, 7:4, 2021.
  24. [25]Anthony W. Ledford and Jonathan A. Tawn. Statistics for near independence in multivariate extreme values. Biometrika, 83:169–187, 1996.DOI
  25. [26]Manuele Leonelli. How exceptional was the Big Three era? Extremes and persistence in men’s professional tennis. arXiv:2608.27362, 2026.arxiv.org/abs/2608.27362
  26. [27]M. H. Lietzke. Relation between weightlifting totals and body weight. Science, 124:486–487, 1956.pubmed.ncbi.nlm.nih.gov/13360275
  27. [28]Roderick J. A. Little and Donald B. Rubin. Statistical Analysis with Missing Data. Wiley, Hoboken, NJ, third edition, 2019.
  28. [29]Charles F. Manski. Partial Identification of Probability Distributions. Springer, New York, 2003.DOI
  29. [30]Alexander J. McNeil and Rüdiger Frey. Estimation of tail-related risk measures for heteroscedastic financial time series: an extreme value approach. Journal of Empirical Finance, 7:271–300, 2000.
  30. [31]Robert J. Mislevy and Hua-Hua Chang. Does adaptive testing violate local independence? Psychometrika, 65:149–156, 2000.
  31. [32]Paul-Hieu V. Nguyen, James M. Smoliga, Benton Lindaman, and Sameer K. Deshpande. Quantifying the limits of human athletic performance: a Bayesian analysis of elite decathletes. Journal of Quantitative Analysis in Sports, 2026. doi: 10.1515/jqas-2026-0026.DOI
  32. [33]Georg Rasch. Probabilistic Models for Some Intelligence and Attainment Tests. Danmarks Paedagogiske Institut, Copenhagen, 1960.
  33. [34]M. E. Robinson and J. A. Tawn. Statistics for exceptional athletics records. Journal of the Royal Statistical Society, Series C, 44:499–511, 1995.DOI
  34. [35]Donald B. Rubin. Inference and missing data. Biometrika, 63:581–592, 1976.
  35. [36]Carl Scarrott and Anna MacDonald. A review of extreme value threshold estimation and uncertainty quantification. REVSTAT Statistical Journal, 10:33–60, 2012.
  36. [37]Roy G. Sinclair. Normalizing the performances of athletes in Olympic weightlifting. Canadian Journal of Applied Sport Sciences, 10:94–98, 1985.pubmed.ncbi.nlm.nih.gov/4017158
  37. [38]Harry Spearing, Jonathan Tawn, David Irons, Tim Paulden, and Grace Bennett. Ranking, and other properties, of elite swimmers using extreme value theory. Journal of the Royal Statistical Society, Series A, 184:368–395, 2021.DOI
  38. [39]Jess Spearing, Jonathan A. Tawn, David Irons, and Tim Paulden. A framework for statistical modelling of the extremes of longitudinal data, applied to elite swimming. arXiv:2306.12419, 2023.arxiv.org/abs/2306.12419
  39. [40]Alec G. Stephenson and Jonathan A. Tawn. Determining the best track performances of all time using a conceptual population model for athletics records. Journal of Quantitative Analysis in Sports, 9:67–76, 2013.DOI
  40. [42]S. Kyle Travis, Michael C. Zourdos, and Caleb D. Bazyl er. Weight selection attempts of elite classic powerlifters. Perceptual and Motor Skills, 128:507–521, 2021.DOI
  41. [43]Aki Vehtari, Andrew Gelman, Daniel Simpson, Bob Carpenter, and Paul-Christian Bür kner. Rank-normalization, folding, and localization: an improved R̂ for assessing convergence of MCMC. Bayesian Analysis, 16(2):667–718, 2021.DOI
  42. [44]Jennifer L. Wadsworth, Jonathan A. Tawn, Anthony C. Davison, and Daniel M. Elton. Modelling across extremal dependence classes. Journal of the Royal Statistical Society, Series B, 79:149–175, 2017.arxiv.org/abs/1408.5060
  43. [45]Julien Worms and Rym Worms. New estimators of the extreme value index under random right censoring, for heavy-tailed distributions. Extremes, 17:337–358, 2014.DOI

Paper details

vlnc:W_gcvkkpd8
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents