WorksPhysics and Astronomy
Persistent structural regimes in Bitcoin transactions relate to market cycles
Abstract
Bitcoin's blockchain is a public record of every on-chain transaction, while the price of bitcoin is set by trading on external markets. We ask whether on-chain activity alone reveals regimes associated with market trends, and whether this association remains stable over time. We summarise each week of blockchain activity from 2012 to 2026 using four measures of transaction structure and fit a two-state hidden semi-Markov model without price data. The sole market-informed constraint is an asymmetry in regime persistence, motivated by the observation that Bitcoin's price increases tend to outlast its declines. During earlier market cycles, the inferred regimes exhibit distinct transaction structures and align with periods of rising and falling prices more closely than random sequences with the same regime-duration distributions. The ongoing downturn, however, breaks this correspondence: the blockchain remains in the longer-lived regime, while its transaction structure resembles neither previous downturns nor previous upturns. Our findings support a relationship between on-chain activity and market dynamics, while highlighting its potentially time-dependent nature.
Introduction
Bitcoin offers an unusual setting for studying the relation between economic activity and market dynamics. Its blockchain provides a public historical record of transactions, while BTC is simultaneously traded in external markets where its price is formed. The transaction record makes it possible to study how activity on the network changes over time rather than relying only on market variables such as price and volume. Previous work has shown that the Bitcoin transaction network has a non-trivial and evolving structure [1, 2], and that changes in this structure can accompany substantial movements in the Bitcoin price [3, 4, 5, 6]. This raises the question of whether on-chain activity itself can be organised into persistent regimes, and how such regimes relate to the broader evolution of the Bitcoin market.
Regime-based descriptions have been used extensively to capture changes in cryptocurrency market behaviour. Markov-switching models reveal distinct volatility regimes in Bitcoin returns [7], while hidden Markov models have been used to identify unobserved bullish, stable, and bearish price regimes [8] and broader positive and negative phases across cryptocurrency markets [9]. In these approaches, however, the latent states are inferred primarily from prices, returns, or volatility. A different question is whether persistent regimes can instead be identified in the transaction activity recorded on the blockchain, and compared with the market without including price in the observation vector. This distinction makes it possible to ask not only whether blockchain and market dynamics are related, but also whether that relation remains stable over successive market cycles.
Here we study whether Bitcoin on-chain activity can be described by a small number of persistent latent regimes and whether their relation to the market remains stable over time. The observed asymmetry between longer periods of overall price increase and shorter periods of overall decline (Figure 1) motivates a two-timescale persistence hypothesis, which we represent with a two-state hidden semi-Markov model (HSMM) fitted to multivariate transaction-structural observables. Market price is not included in the model but it motivates the duration constraints and the choice between two solutions. Our question is therefore whether the transaction record admits a persistent decomposition related to the market, not whether a uniquely preferred segmentation emerges without market-informed choices. We assess the selected chronology against a duration-matched random benchmark and compare its transaction structure with price-defined phases. Over the earlier market cycles, the selected regimes correspond substantially to periods of price increase and decline (Figure 2) and exhibit distinct transaction profiles (Figure 3). This correspondence is absent in the most recent downturn, where the blockchain remains in the more persistent regime and its configuration differs from previous periods of decline. Whether this departure from the historical correspondence is temporary or reflects a more persistent change in the relation between transaction structure and market dynamics remains open.
Figure 1. Temporal asymmetry of Bitcoin price dynamics. a, Bitcoin spot price [USD] on a logarithmic scale, with alternating Rising and Falling legs defined by the major price peaks and troughs. Triangles mark the extrema and labels give the corresponding leg durations in weeks. The first Rising leg (*) is truncated by the sample boundary and the final Falling leg (+) is ongoing. b, Local price trend , the 30-day log return of the price. Positive and negative values are shown separately, with vertical lines marking the extrema from panel a. The inset gives the distribution of ; 57.0% of valid days have positive trend and 42.9% negative trend.
Figure 2. a, Schematic of the two-state hidden semi-Markov model. Each week (1008 blocks) is represented by a four-dimensional observed feature vector . Each episode has a latent state and a duration drawn from the state-dependent distribution . During each state episode, successive observations are generated according to . Self-transitions are excluded, so persistence is represented by the explicit duration distributions. b, Daily Bitcoin all-time-high ratio shown in grey, together with its trailing 60-day mean in black and the decoded Long and Short episodes L1–L4 and S1–S3. The initial unshaded interval precedes the first week used by the model. The latest price decline occurs while the blockchain remains in L4. c, Annualised endpoint log-price slope for each decoded episode. All Long episodes have positive slopes and all Short episodes have negative slopes. d, Chance agreement. Grey histogram: Cohen’s between the market phases before F4 and each of 20000 random state sequences with episode durations drawn from the model’s duration distributions. Vertical line: of the HSMM states. Dark bars: random sequences that agree at least as well as the HSMM.
Figure 3. State anatomy and the changing on-chain signature of market phases. a, Distribution of the four transformed and scaled on-chain features in the decoded Long and Short states. Points denote individual weeks, boxes the interquartile range with median, whiskers the 10th–90th percentiles, and labelled open circles the medians of individual episodes – and –. b, Median feature values in the externally defined Rising (–) and Falling (–) market phases. The ongoing phase is outlined. c, Normalised coefficients of the linear-discriminant direction (LD1) fitted to historical Rising and Falling observations. Fan-out and consolidation carry the largest positive weights, while transaction entropy enters with negative weight. d, Historical Rising and Falling observations projected onto LD1 and the leading principal component of the feature variation remaining after removal of LD1. The discriminant and residual component are fitted using – and –; is excluded from the fit and projected afterwards. Marginal densities and contours show the historical Rising and Falling distributions. Small squares show the weeks, coloured by date from light (October 2025) to dark (August 2026). Diamonds mark the pooled Long- and Short-state medians, and the outlined square the median of .
Materials and Methods
Market benchmark and temporal hypothesis
Bitcoin’s market history has commonly been described in terms of multi-year boom–bust cycles occurring on a time scale comparable to the approximately four-year reward halving schedule [10, 11]. Although the correspondence between halvings and market turning points is neither exact nor necessarily causal [10, 12], the repeated cycle-scale structure motivates a systematic description of Bitcoin’s market phases.
For the market-side analysis, we use price data [13, 14, 15]. The daily price is taken as the final available close of each UTC day, over the January 2012–August 2026 observation period.
We define a retrospective market benchmark by dividing the price history into alternating Rising and Falling phases. The phases are separated by the major peaks and troughs of the price, which we locate using the halving dates (28 November 2012, 9 July 2016, 11 May 2020 and 20 April 2024). After each halving, the peak is the highest daily close in the following two years, and the trough is the lowest daily close between that peak and the next halving. Each halving cycle therefore contains one peak and one trough. A Rising phase runs from a trough to the next peak, and a Falling phase from a peak to the next trough. This gives four Rising phases, R1–R4, and four Falling phases, F1–F4 (Figure 1a). R1 begins at the start of the data, so its true start is unknown. F4 begins at the last peak, in October 2025, and is still ongoing because the next halving has not yet occurred. We use these phases in all later comparisons. They describe past market history and are not real-time signals.
The resulting benchmark exhibits a clear temporal asymmetry between Rising and Falling market phases. Excluding the sample-truncated R1 and the ongoing F4, the completed Rising phases R2–R4 persist for approximately 150–152 weeks, whereas the completed Falling phases F1–F3 last only 52–58 weeks (Figure 1a). Thus, at the cycle scale, Rising phases last much longer than Falling phases.
A complementary local measure shows a similar imbalance at shorter time scales. With the closing price on day , we use the 30-day log return , which is positive when the price is higher than 30 days earlier. Across the full sample, is positive on 57.0% of days and negative on 42.9% (Figure 1b), so upward local trends occupy more of the observation period than downward ones. Together with the asymmetry observed in the market-phase durations, this motivates a latent-state description with one relatively long and one relatively short persistence scale.
On-chain representation and feature preprocessing
To explore persistent regimes in Bitcoin transaction activity, we use blockchain data from January 2012 to August 2026, obtained from Google BigQuery [16]. Following previous works that describe each block by its numbers of transactions, inputs and outputs [5, 6], we compute four quantities for every block.
For block , let and be the total numbers of inputs and outputs of its transactions, and its number of transactions. We define
where only prevents division by zero. The input–output imbalance lies between −1 and 1: it is positive when the block has more inputs than outputs and negative when it has more outputs than inputs. The consolidation and the fan-out measure the average numbers of inputs and outputs per transaction, on a logarithmic scale. The fourth quantity is the transaction entropy
where is the total amount of bitcoin in the outputs of transaction in block . The entropy is high when value is spread evenly over many transactions and low when a few transactions carry most of it.
These quantities record only how many inputs and outputs transactions have and how much bitcoin they move. They do not tell why a transaction was made, for example whether it is a trade, a payment, or a transfer between a user’s own addresses.
Values change strongly from block to block, so we group the blocks into consecutive sets of 1008 blocks and take the median of each quantity within each set. A block is added about every ten minutes, so each set covers about one week (6.7 days on average). From now on we will refer to these sets simply as weeks; they are the time step of the whole analysis. We write for the median of quantity in week .
Between 2012 and 2026, the typical values of imbalance, fan-out and entropy change considerably. A model fitted directly to these values could simply separate early years from later years, instead of finding states that recur. We therefore replace each of these three values by its percentile rank in the recent past: is the fraction of the previous 156 weeks (about three years) in which quantity was lower than in week . For example, means that the value in week is higher than in 90% of the previous three years. Formally,
where is the set of these previous weeks, equals 1 when its condition holds and 0 otherwise. Near the start of the data, fewer than 156 previous weeks exist and we use all of them. We require at least 39 previous weeks (about nine months), so the first 39 weeks serve only as history and 760 weeks remain for the model. Consolidation changes much less over the long run, so we use its weekly median directly.
Finally, we put the four series on a common scale. For each of them we subtract its median over the 760 weeks and divide by its interquartile range, which gives the model inputs . A value of zero therefore means a typical week: for consolidation, a typical level; for imbalance, fan-out and entropy, a typical percentile rank, not a typical raw value. For each week , the model observes the four inputs
shown in Figure S1.
Figure S1. Final on-chain features used by the HSMM. Time series of the four model inputs over the 760 weeks used by the model. Input–output imbalance, fan-out and transaction entropy are trailing percentile ranks. All four are then centred on their full-sample medians and divided by their interquartile ranges.
Hidden semi-Markov model
The market analysis (Figure 1) shows that the characteristic duration of Rising phases is substantially longer than that of Falling phases, motivating the consideration of two distinct persistence scales in the latent blockchain dynamics. Consequently, we model the temporal evolution of the on-chain observation sequence using a two-state hidden semi-Markov model (HSMM). A standard hidden Markov model (HMM) represents state persistence through repeated self-transitions, implying a geometric distribution of state dwell times [17, 18]. This is restrictive for the present problem because regime duration is itself part of the hypothesis. An HSMM instead assigns an explicit duration distribution to each latent state, allowing persistence to be modelled separately from the transition process [18]. We therefore consider two states, denoted Long () and Short (), which represent blockchain regimes with relatively long and short persistence, respectively. These labels refer to temporal persistence, not to Rising and Falling market phases or to trading positions. Market price is not included in the HSMM observation vector; within the model, it serves only to motivate the asymmetric duration structure. The state-dependent feature distributions and decoded transition dates are inferred from the on-chain observations for each fit. The model structure is summarised in Figure 2a.
Let denote the latent state in week . Conditional on , the four model inputs are modelled by a diagonal Gaussian emission distribution,
so that each state has its own mean and variance for each input. The diagonal covariance is a simplifying assumption: given the state, the inputs are treated as independent, although they are built from the same transaction counts and can therefore be correlated.
Temporal persistence is represented separately through state-specific duration distributions. An episode is a period of consecutive weeks in the same state, and its duration is the number of weeks it contains. For an episode in state , we take a normal random variable with mean and standard deviation . Each duration in an allowed range then receives the probability that rounds to ,
rescaled so that these probabilities sum to one. Durations outside have probability zero.
The Long/Short hypothesis enters through these duration distributions. The standard deviation and the allowed range are set in advance. For the Long state, weeks and ; for the Short state, weeks and . The mean is fitted to the data, which gives weeks and weeks. Because durations outside are excluded, the mean and standard deviation of the resulting durations can differ from and .
Self-transitions between consecutive episodes are excluded, following strict explicit-duration formulations [19].
Persistence is therefore represented by Eq. (6) rather than by repeated state self-transitions; with two states, successive episodes alternate between and . The model is initialised by -means clustering [20] and fitted using an EM-style forward–backward procedure, with posterior state occupancies used to update the Gaussian parameters and posterior episode-duration counts used to update and . The decoded chronology is obtained by explicit-duration Viterbi inference [18].
We repeat the fit from twenty initialisations with different random seeds. The runs converge to two distinct decoded segmentations: B1 is reached in 11 runs and contains eight episodes, whereas B2 is reached in nine runs and contains seven (Figure S2). B1 has the higher marginal likelihood. We retain B2 for the market comparison because its states show a clearer correspondence with the market phases defined above. The null test described below (Chance agreement) accounts for this selection.
Figure S2. Dependence of the HSMM solution on initialisation. a, Twenty model initialisations converge to two distinct decoded state sequences. Basin B1 is recovered in 11 of 20 starts and contains eight episodes, while basin B2 is recovered in 9 of 20 starts and contains seven episodes. b, Annualised endpoint log-price slopes for the decoded episodes of each solution, calculated as in Figure 2c and grouped by latent state. In B2, all Long episodes have positive slopes and all Short episodes have negative slopes, whereas both states exhibit positive slopes in B1. Although B1 has the higher likelihood, B2 shows a clearer correspondence between the inferred blockchain regimes and historical market direction.
Comparison with the market phases
We compare the decoded states with the market phases in four ways. First, we measure the agreement between the state (Short or Long; Figure 2b) and the market direction (Falling or Rising; Figure 1a) over the 713 weeks before , as the fraction of weeks in which they match and as Cohen’s [21]. Cohen’s corrects this fraction for the agreement expected by chance, given how often each label occurs: means perfect agreement, agreement at chance level, and less agreement than chance. We compare with the null model described in the next subsection. Second, for each episode we compute the annualised endpoint log-price slope, , where and are the prices at the start and end of the episode and is the elapsed time in years; this is the mean log return per year over the episode (Figure 2c; both fitted solutions in Figure S2). Third, we take the median of each feature within each market phase, which compares the phases with the transaction structure without the HSMM (Figure 3b). Fourth, we summarise the four-dimensional configuration with two orthogonal axes fitted to the historical weeks only (– and –), with excluded from the axis fitting. This exclusion does not extend to the full-sample preprocessing or HSMM fitting. The first axis, LD1, is obtained by linear discriminant analysis (LDA) of Rising versus Falling weeks [22]: it is the weighted sum of the four inputs that best separates the two groups. Only the ratios between the weights matter, so we scale them so that their squares sum to one and choose their sign so that Falling weeks have higher LD1 values; the weights are shown in Figure 3c. The second axis, residual PC1, is the leading principal component [23] of the historical weeks after their LD1 component has been removed. All weeks, including those of , are then projected on these fixed axes (Figure 3d; the phase in detail in Figure S3). LD1 is supervised by the historical market labels and is used here to describe their contrast, not to establish out-of-sample classification performance.
Figure S3. Within-phase evolution of on-chain transaction structure during R3. a, Bitcoin all-time-high ratio during the externally defined R3 Rising phase (2018-12-15 to 2021-11-10). The daily ratio is shown in grey and its trailing 60-day mean in black. Although R3 forms one broad Rising phase, it contains substantial internal variation, including the pronounced March 2020 drawdown. Dashed vertical lines divide the interval into three consecutive equal-count chronological partitions, labelled 1–3. The solid vertical line marks the HSMM transition from L3 to S3 on 2021-03-07. b, The same R3 observations projected onto the LD1–Residual-PC1 coordinates of Figure 3d. The 118 weeks decoded as L3 are shown as circles and the 36 weeks decoded as S3 as squares. Colour indicates time from early to late. Light-grey points show the remaining historical observations, while diamonds mark the pooled Long- and Short-state medians. Open circles labelled 1–3 show the centroids of the three chronological partitions defined in panel a, with arrows indicating their temporal ordering. The distribution shifts through R3 towards the Short-associated side of LD1, while the blockchain transition to S3 occurs before the end of the price-defined Rising phase.
Chance agreement
The HSMM states and the market phases both last for many weeks. Two such slowly changing sequences can overlap over long stretches purely by chance, so comparing them with labels drawn at random week by week would make the agreement look more significant than it is. Instead, we generate 20000 random sequences of Long and Short episodes whose durations are drawn from the model’s duration distributions. Each random sequence covers the same 760 weeks as the data. Long and Short episodes alternate, and the length of each episode is drawn from or [Eq. (6)], with the fitted values and . Each sequence starts at a random point inside its first episode, so its transitions fall on random dates. No blockchain data are used. For each random sequence we compute with the market phases over the 713 weeks before $F4 ({xref}Fig. 2d). The $p$-value is the fraction of random sequences with $\kappa$ at least as large as that of the HSMM. Since B2 was chosen between two fitted solutions, we also report the $p$-value corrected for this choice, $2p$ (Bonferroni correction {citep}[24]`), which holds whatever the relation between the two solutions. This test takes the duration settings as given; it does not repeat the model fit or the choice between solutions.
Results
Decoded chronology and market correspondence
The selected segmentation resolves the on-chain sequence into seven persistent episodes, , , , , , , , spanning September 2012 to August 2026. Over most of the sample, the decoded blockchain chronology follows the main Bitcoin market phases (Figure 2b): Short episodes occur predominantly during Falling phases, whereas Long episodes occur predominantly during Rising phases. The same relation is visible at the episode scale. The annualised endpoint log-price slope is positive for all four Long episodes and negative for all three Short episodes (Figure 2c). The transition dates do not coincide exactly, showing that the blockchain and price segmentations are related but not identical. The clearest exception occurs in the ongoing 2025–2026 Falling phase, during which every week so far is assigned to .
Before , the decoded state matches the market direction in 88% of the 713 weeks, with Cohen’s . Only 1.9% of the 20000 random sequences in the duration-matched benchmark agree with the market at least as well (; Figure 2d). Because B2 was chosen from two fitted solutions, we also correct for this choice, which gives .
Transaction structure of the latent regimes
The two states differ clearly in transaction structure (Figure 3a). Compared with the Long state, the Short state has higher consolidation, input–output imbalance and fan-out, and lower transaction entropy. Consolidation is compared with the whole sample: in Short weeks, transactions have more inputs than in a typical week. The other three features are percentile ranks, which compare each week with the preceding three years. A positive imbalance in the Short state therefore means a higher imbalance than usual at that time, not more inputs than outputs; the same applies to fan-out and entropy. The open circles in Figure 3a, which mark the medians of individual episodes, show that this difference is not due to a single period. It is clearest for consolidation, whose median is below zero in all four Long episodes and above zero in all three Short episodes; the other features vary more between episodes. In Short episodes, transactions therefore merge more previously received coins and create more outputs than usual, while value is concentrated in fewer transactions. An earlier study linked similar changes in the input–output structure of blocks to periods of market
stress [5].
Price-defined phases and the ongoing downturn
The market phases, defined from the price alone as in Figure 1a, allow the same comparison without the HSMM (Figure 3b). In each of the three completed Falling phases, consolidation, input–output imbalance and fan-out are positive. The Rising phases are less uniform: consolidation remains close to or below zero, whereas imbalance, fan-out and entropy change sign from one phase to the next. This mirrors the HSMM result (Figure 3a): the Falling phases share the signature of the Short state, the Rising phases are as variable as the Long episodes, and consolidation is again the most consistent feature.
fig:3c condenses this pattern into a single direction. LDA, fitted to the weeks of the completed phases – and – with left out, finds the weighted combination of the four features, LD1, that best separates Rising from Falling weeks (see Materials and methods). Its weights restate what Figure 3b shows: fan-out and consolidation, positive in every Falling phase, carry the largest positive weights, and entropy, low in two of the three Falling phases, enters with a negative weight. Imbalance is also positive in every Falling phase but receives little weight: it is built from the same input and output counts as consolidation and fan-out, and adds little information beyond them. Figure 3d shows every week on LD1 and on residual PC1, the direction of largest remaining variation, both fitted to the historical weeks alone. The historical Rising and Falling weeks form two overlapping but distinct distributions, and the Long- and Short-state medians of the HSMM (blue and red diamonds) lie inside them, respectively.
The ongoing Falling phase breaks the historical pattern. Its consolidation, imbalance and fan-out are all negative, opposite in sign to every earlier Falling phase, and its low entropy is shared with and (Figure 3b). In Figure 3d, the early weeks (light squares) fall within the historical Rising distribution, near the Long-state median, consistent with the blockchain remaining in . Later weeks (darker squares) drift towards lower values on both axes, and many of the most recent ones lie below every historical week along residual PC1. The current downturn therefore lacks the transaction structure of past declines without matching past Rising phases. Earlier mismatches between the price and blockchain chronologies were offsets in timing, such as the gradual shift from to within (Figure S3); the present one involves a feature configuration unlike that of any earlier phase.
Discussion and conclusions
This study asked whether Bitcoin transaction structure is organised into persistent regimes and whether their relation to the market remains stable over time. Over the historical market cycles, the transaction record admits a persistent decomposition closely related to the market. Unlike regime models fitted to prices, returns or volatility [7, 8, 9], the HSMM observes transaction structure alone, yet its Long and Short regimes follow the Rising and Falling phases of the market. Three complementary observations support this correspondence: the decoded chronology (Figure 2b), the signs of the episode price slopes (Figure 2c), and a transaction structure shared by the Short state and the price-defined Falling phases, which emerges even without the model (Figure 3a,Figure 3b). Random sequences drawn from the same duration distributions rarely reach a comparable agreement with the market (Figure 2d).
The Short state admits a direct interpretation in terms of how bitcoin is moved: transactions combine more previously received coins and create more outputs than usual, while value is concentrated in fewer transactions. This profile is consistent with earlier evidence that on-chain activity varies with market conditions. Javarone et al. reported changes in input–output structure around periods of market stress [5], and Di Antonio et al. found that an entropy measure of block structure dropped sharply as Bitcoin first gained value [6]. Bovet et al. found that blockchain activity becomes more heterogeneous during price declines [4]. Hoang and Baur associated negative price shocks and elevated volatility with transfers from private wallets to exchanges [25], a mechanism that may contribute to the Short-state profile. Our analysis does not identify the actors or motives behind individual transactions; rather, it shows that a small set of aggregate features combines into persistent configurations that recur across market cycles.
The ongoing downturn departs from this pattern in a way that differs from earlier discrepancies. Before , the two chronologies disagreed only on timing: during , for instance, the weeks shifted gradually towards the Short side of , and the transition to preceded the end of the Rising phase (Figure S3). In , by contrast, the blockchain remains in while the market declines; consolidation, imbalance and fan-out are negative, unlike in every earlier Falling phase (Figure 3b); and the later weeks move away from both historical distributions (Figure 3d). The transaction structure itself has changed, not merely the timing of the transitions. So far, therefore, the relation between transaction structure and market direction has not remained stable. As is still unfolding, whether this departure is transient or lasting remains open.
Several developments in the Bitcoin ecosystem may have weakened the historical link. On the blockchain side, Ordinals and inscriptions introduced new uses of block space and have been shown to affect transaction fees, block-space demand and overall network activity [26, 27]. Layer-2 systems such as the Lightning Network allow payments to be made without each being recorded on-chain, and their adoption has been associated with reduced congestion of the base layer [28]. Both developments alter which activity reaches the blockchain, and hence what our features capture. The market has evolved as well: derivatives contribute substantially to Bitcoin price discovery [29], and recent work identifies an important role for spot Bitcoin ETFs [30]. As the ecosystem matures, the blockchain and the market need not remain as closely coupled as in earlier cycles. These mechanisms are plausible, but the cited studies do not examine the features used here, and our analysis cannot attribute the behaviour of to any of them.
Some limitations should be noted. The historical record contains only three completed Falling phases, and consecutive weeks are serially dependent, so the correspondence rests on a small number of independent repetitions. Market history informs two modelling choices, the allowed durations of the two states and the selection of B2 over B1, which attains the higher likelihood (Figure S2); our result is therefore the existence of a market-related decomposition of the transaction record, not a unique segmentation. The model is also deliberately simple, with two states and Gaussian emissions that only approximate the bounded percentile ranks. Finally, the analysis is descriptive: it neither tests causal relations between blockchain activity and price, for which dedicated methods exist [31, 32, 33, 34] and have been applied to Bitcoin [4, 5], nor provides a trading signal.
Within these limits, our results show that Bitcoin’s public transaction record carries a persistent, market-related structure that can be identified without including prices in the model. On-chain activity therefore offers a view of market regimes that complements those based on prices, returns and volatility. The correspondence, however, has not carried over to the current downturn, which suggests that the relation between Bitcoin’s blockchain and its market evolves with the ecosystem rather than remaining fixed.
Data Availability
Block and transaction data are available from the Google BigQuery public dataset bigquery-public-data.crypto_bitcoin [16]. Daily prices are Bitstamp BTC/USD closes for 2012–2024 [13, 14] and Binance BTC/USDT closes for 2025–2026 [15].
Credit Authorship Contribution Statement
BADS: Conceptualization, Formal analysis, Software, Writing – original draft. GDA: Conceptualization, Formal analysis, Supervision, Writing – review & editing. MAJ: Conceptualization, Supervision, Writing – review & editing.
Declaration of Competing Interest
The authors declare no competing interests.
Declaration of Generative AI and AI-Assisted Technologies in the Manuscript Preparation Process
During the preparation of this work the authors used large language models, including Claude (Anthropic) and ChatGPT (OpenAI), to revise the manuscript text and to check and extend the analysis code. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.
Acknowledgments
BADS gratefully acknowledges financial support from the Exponential Science Foundation.
References
- [1]D. Kondor, M. Pósfai, I. Csabai, and G. Vattay, Do the rich get richer? An empirical analysis of the Bitcoin transaction network, PLOS ONE 9, e86197 (2014).
- [2]N. Vallarano, C. J. Tessone, and T. Squartini, Bitcoin transaction networks: An overview of recent results, Frontiers in Physics 8, 286 (2020).DOI
- [3]D. Kondor, I. Csabai, J. Szüle, M. Pósfai, and G. Vattay, Inferring the interplay between network structure and market effects in Bitcoin, New Journal of Physics 16, 125003 (2014).
- [4]A. Bovet, C. Campajola, F. Mottes, V. Restocchi, N. Vallarano, T. Squartini, and C. J. Tessone, The evolving liaisons between the transaction networks of Bitcoin and its price dynamics, in Proceedings of Blockchain Kaigi 2022 (BCKK22), JPS Conference Proceedings, Vol. 40 (The Physical Society of Japan, 2023) p. 011002.
- [5]M. A. Javarone, G. Di Antonio, G. V. Vinci, R. Cristodaro, C. J. Tessone, and L. Pietronero, Disorder unleashes panic in Bitcoin dynamics, Journal of Physics: Complexity 4, 045002 (2023).
- [6]G. Di Antonio, G. V. Vinci, L. Pietronero, and M. A. Javarone, The rise of rationality in blockchain dynamics, New Journal of Physics 25, 123042 (2023).
- [7]D. Ardia, K. Bluteau, and M. Rüede, Regime changes in Bitcoin GARCH volatility dynamics, Finance Research Letters 29, 266 (2019).
- [8]P. Giudici and I. Abu Hashish, A hidden Markov model to detect regime changes in cryptoasset markets, Quality and Reliability Engineering International 36, 2057 (2020).
- [9]F. Pennoni, F. Bartolucci, G. Forte, and F. Ametrano, Exploring the dependencies among main cryptocurrency log-returns: A hidden Markov model, Economic Notes 51, e12193 (2022).
- [10]J. Y.-L. Chan, S. W. Phoong, S. Y. Phoong, W. K. Cheng, and Y.-L. Chen, The Bitcoin halving cycle volatility dynamics and safe haven-hedge properties: A MSGARCH approach, Mathematics 11, 698 (2023).
- [11]G. B. M’bakob, Bubbles in Bitcoin and Ethereum: The role of halving in the formation of super cycles, Sustainable Futures 7, 100178 (2024).
- [12]M. Lashkaripour, Some stylized facts about Bitcoin halving, Finance Research Letters 69, 106198 (2024).
- [13]Zielak (mczielinski), Bitcoin Historical Data, Kaggle dataset, https://www.kaggle.com/datasets/mczielinski/bitcoin-historical-data (2025), licensed under CC BY-SA 4.0.
- [14]M. de Beer, bitstamp-btcusd-minute-data, https://github.com/ff137/bitstamp-btcusd-minute-data (2025), GitHub repository, accessed: 2026-03-15.
- [15]Binance, Binance Public Data: spot klines, BTCUSDT, 1 hour, https://data.binance.vision (2026).
- [16]A. Day and E. Medvedev, Introducing six new cryptocurrencies in BigQuery public datasets—and how to analyze them, Google Cloud Blog, https://cloud.google.com/blog/products/data-analytics/introducing-six-new-cryptocurrencies-in-bigquery-public-datasets—and-how-to-analyze-them (2019), Bitcoin dataset bigquery-public-data.crypto_bitcoin.
- [17]L. R. Rabiner, A tutorial on hidden Markov models and selected applications in speech recognition, Proc. IEEE 77, 257 (1989).
- [18]S.-Z. Yu, Hidden semi-Markov models, Artificial Intelligence 174, 215 (2010).DOI
- [19]M. J. Johnson and A. S. Willsky, Bayesian nonparametric hidden semi-Markov models, Journal of Machine Learning Research 14, 673 (2013).
- [20]S. P. Lloyd, Least squares quantization in PCM, IEEE Transactions on Information Theory 28, 129 (1982).DOI
- [21]J. Cohen, A coefficient of agreement for nominal scales, Educational and Psychological Measurement 20, 37 (1960).
- [22]R. A. Fisher, The use of multiple measurements in taxonomic problems, Annals of Eugenics 7, 179 (1936).
- [23]I. Jolliffe, Principal component analysis, in Wiley StatsRef: Statistics Reference Online (John Wiley & Sons, Ltd, 2014).
- [24]O. J. Dunn, Multiple comparisons among means, Journal of the American Statistical Association 56, 52 (1961).
- [25]L. T. Hoang and D. G. Baur, Loaded for bear: Bitcoin private wallets, exchange reserves and prices, Journal of Banking & Finance 144, 106622 (2022).
- [26]L. Bertucci, Bitcoin Ordinals: Determinants and impact on total transaction fees, Research in International Business and Finance 70, 102338 (2024).
- [27]A. Wiedenmann and A. Guettler, Bitcoin Ordinals and Inscriptions: An analysis of Bitcoin’s evolving network dynamics, The Journal of the British Blockchain Association 9, 1 (2026).
- [28]A. Divakaruni and P. Zimmerman, The Lightning Network: Turning Bitcoin into money, Finance Research Letters 52, 103480 (2023).
- [29]C. Alexander and D. F. Heck, Price discovery in Bitcoin: The impact of unregulated markets, Journal of Financial Stability 50, 100776 (2020).
- [30]A. Mohamad, Do Bitcoin ETFs lead price discovery following their introduction in the Bitcoin market?, Computational Economics 66, 947 (2025).
- [31]C. W. J. Granger, Investigating causal relations by econometric models and cross-spectral methods, Econometrica 37, 424 (1969).
- [32]J. Pearl, Causality: Models, Reasoning, and Inference, 2nd ed. (Cambridge University Press, Cambridge, 2009).
- [33]G. Sugihara, R. May, H. Ye, C.-h. Hsieh, E. Deyle, M. Fogarty, and S. Munch, Detecting causality in complex ecosystems, Science 338, 496 (2012).DOI
- [34]G. Di Antonio and G. V. Vinci, Nonlinear correlations underlie linear response and causality, Physical Review Research 7, L042029 (2025).arxiv.org/abs/2410.08708





