betaFeedback

WorksEconomics, Econometrics and Finance

Zero Crossings on the Accuracy-Smoothness Frontier: Maximal Monotone and Minimal Inflection Non Stationary Predictors

Abstract

The Smooth Sign Accuracy (SSA) framework resolves the smoothness–accuracy dilemma for stationary predictors by balancing mean squared error (MSE) against the frequency of sign changes. We generalize it to non-stationary integrated processes (I-SSA), deriving MSE-optimal predictors constrained by the expected duration between turning points (I(1)) or inflection points (I(2)). The resulting maximal-monotone and minimal-inflection predictors, filters, and smoothers customize classic benchmarks along a new efficient Accuracy–Smoothness (AS) frontier. Tracking the identity (the original data) along this frontier defines the I-SSA trend, a novel smoothing concept that, unlike classical filters, imposes no artificial structure on the data-generating process and is directly interpretable through explicit control of turning- or inflection-point frequency. The optimization is a quadratically constrained quadratic program (QCQP) that reduces to a Generalized Trust Region Subproblem (GTRS) with hidden convexity. On its feasible, invertible region the problem is convex with a well-defined global optimum, and an exact closed-form expression delivers the efficient AS frontier. Under fairly general assumptions, finite-sample solutions converge to an asymptotic I-SSA predictor, devoid of integration boundary perturbations. The framework unifies mean squared error on non-stationary levels, and sign accuracy, zero-crossing frequency, and target correlation on stationary differences.

Introduction

Forecasting requires balancing conflicting objectives. The SSA framework [43, 44] targets accuracy (estimating future levels) and smoothness (suppressing spurious noise). We propose I-SSA, extending SSA to non-stationary integrated series and letting users balance accuracy (A) and smoothness (S) along a novel, efficient AS-frontier. We focus on univariate, homoscedastic series (the multivariate stationary case is treated in [45]). I-SSA overcomes the limitations of the frequency-domain forecast trilemma [28], offering better interpretability, numerical efficiency, and robustness against overfitting.

Traditional smoothing methods include curvature-penalized approaches [42, 15, 17]—despite recent critiques [13, 33]—splines [9, 46], trend extraction [14, 8], and spectral filtering [3, 36, 31]. Since our approach emphasizes turning and inflection points (TP/IP), a natural benchmark is the HP filter, a special case of Whittaker–Henderson smoothing, still widely used in business-cycle analysis and crisis prediction [7]; we complement the analysis with the boosted HP filter [33] (Phillips and Shi 2024). Broader comparisons are available in our R package.

While the HP filter minimizes curvature, SSA defines smoothness via the expected duration between sign changes (zero crossings, ZC) of a stationary, zero-mean predictor, building on Rice [40]. I-SSA extends this to integrated processes by targeting ZCs in stationary differences while minimizing MSE at the non-stationary levels. Economically, sign changes in growth rates mark transitions between expansions and contractions, so capturing business cycles requires sufficiently long intervals between crossings. This mean duration, the holding time (HT) [43], is explicitly regulated in I-SSA for the growth (I(1)) or acceleration (I(2)) of the predictor via the rate of TPs or IPs.

I-SSA defines a novel Accuracy-Smoothness (AS) frontier: no other linear predictor of the same, or shorter length can improve both MSE on non-stationary levels (accuracy) and ZC rate in stationary differences (smoothness). Within this family, I-SSA is characterized by the fewest growth sign-changes, which justifies the terms maximal monotone (I(1)) and minimal inflection (I(2)).2 The classic MSE predictor is one point on this frontier, whereas I-SSA covers it in full. We also give a closed-form link between HT and MSE along the frontier, extending the stationary AS dilemma [43, 44 b] to integrated processes.

Extending SSA to the non-stationary setting raises technical caveats. The cointegration constraint between the acausal target and the causal predictor, which ensures finite MSE on levels, introduces asymmetries in the Toeplitz structure of the system matrices. Permanent ‘unit-root’ memory prevents convergence to circulants, so the system matrices do not commute and a fixed Fourier eigenbasis is unavailable. We therefore use a generalized eigenvalue decomposition of a matrix pencil, as in cointegration [19, 21] and canonical correlation analysis [4], but with a structure originating in the optimization literature on the generalized trust-region subproblem (GTRS; [30, 1]) because the Hessian is generally indefinite. The pencil separately encodes the non-stationary level and stationary autocorrelation operators, which the generalized decomposition diagonalizes simultaneously. This yields a closed-form, MSE-optimal predictor constrained by the expected duration between turning points, to our knowledge not previously explored for integrated processes. Although the formulation matches the stationary case, the geometry differs fundamentally: the generalized eigenvectors are non-orthogonal and depend on the HT, and the acausal target and causal predictor rely on different spectral bases, leading to a counterfactual spectral decomposition (details below).

Formally, the optimization is a generalized trust-region subproblem (GTRS) with hidden convexity. If the classic MSE predictor is known, I-SSA reduces to a convolution with it that modifies its ZC rate while preserving MSE optimality; we call this transformation customization. For this problem, the region of invertibility and strict monotonicity reduce prediction to a one-dimensional search with certification of the global optimum. Under fairly general assumptions, finite-length solutions converge to a well-structured infinite-length I-SSA problem, free of finite-length boundary perturbations.

I-SSA customization means the projection of existing benchmark predictors onto the non-stationary AS frontier. In the applications, we customize the HP filter and its boosted variant, introducing the maximal monotone HP trend nowcast, obtained by targeting the acausal two-sided (boosted) HP filter in I-SSA. This reconciles growth sign analysis, in differences, with optimal level tracking. With the cursor on the frontier set to replicate the benchmark’s TP rate, the I-SSA trend achieves lower MSE, notably at TPs, with faster response (smaller retardation) and no over- or undershooting. All examples are replicable with our open-source R package (https://github.com/wiaidp/R-package-SSA-Predictor), which also extends customization to Hamilton’s regression filter ([13]), the Baxter–King filter ([3]), and a refined Beveridge–Nelson decomposition ([20]).

Replacing the two-sided HP target by the identity, i.e., the original data, defines the I-SSA trend: minimal MSE subject to maximal monotonicity (I(1)) or minimal inflection (I(2)). This novel smoothing principle imposes minimal structural assumptions on the data generating process (DGP), consistent with the typical spectral shape of economic time series ([12]), and prescribes no functional form (polynomial, spline, minimum curvature). TPs or IPs emerge from the interaction between the ‘neutral’ (amorphous-memory) HT constraint and optimal tracking, which supports efficiency and interpretability. Relative to the classic Whittaker–Henderson smoother, the I-SSA trend yields a lower nowcast MSE at directional changes, confirming the findings for the HP customization. Moreover, the concurrent I-SSA trend distributes TPs more evenly and gives clearer signals, avoiding clusters of spurious TPs around directional changes.

The remainder of the paper is organized as follows. § 2 reviews the SSA criterion ([43, 44 b]); § 3 extends it to integrated processes through I-SSA; § 4 applies customization to business cycle nowcasting and the I-SSA trend to data smoothing; § 5 concludes.

SSA

The following summarizes the stationary Smooth Sign Accuracy framework of Wildi (2024, 2026a,b), which provides the foundation for I-SSA.

Consider a zero-mean stationary series xtx_t (non-zero-mean processes are centered), xt=∑i=0∞ξiϵt−ix_t = \sum_{i=0}^{\infty} \xi_i\epsilon_{t-i}, with ξ0=1\xi_0 = 1. The one-sided sequence ξ∞:=(ξ0,ξ1,…)′\xi_{\infty} := (\xi_0,\xi_1,\ldots)' is square summable and corresponds to the weights in the (purely non-deterministic) Wold decomposition of xtx_t, as detailed by Brockwell and Davis (1993). Let zt=∑k=−∞∞γkxt−kz_t = \sum_{k=-\infty}^{\infty} \gamma_k x_{t-k} be another stationary process and assume we want to predict zt+δz_{t+\delta}, the prediction target, with δ∈Z\delta\in\mathbb{Z}. We construct a length-LL causal predictor yt=∑k=0L−1bkxt−ky_t = \sum_{k=0}^{L-1} b_k x_{t-k} for zt+δz_{t+\delta}, termed forecasting, nowcasting, or backcasting as δ>0\delta> 0, =0= 0, or <0< 0.3

For example, with xt=ϵtx_t = \epsilon_t, γ0=1\gamma_0 = 1, γ1=0.5\gamma_1 = 0.5, zt=ϵt+0.5ϵt−1z_t = \epsilon_t + 0.5\epsilon_{t-1} is an MA(1); one-step forecasting (δ=1\delta= 1) gives the MSE predictor yt=0.5ϵty_t = 0.5\epsilon_t, i.e., b0=0.5b_0 = 0.5, bk=0b_k = 0 otherwise.

The MSE predictor tracks the level optimally but can be ‘noisy’: here yt=0.5ϵty_t = 0.5\epsilon_t is WN while zt+1z_{t+1} is more persistent. Similarly, in real-time signal extraction the one-sided concurrent filter is markedly noisier than its acausal, two-sided counterpart (see examples below). To address this, Wildi [43] proposes to control the HT, the mean duration between consecutive zero-crossings, of the predictor.

Specifically, let Ξ‾\overline{\Xi} denote the L×LL \times L matrix with ii-th row Ξ‾i:=(ξi−1,ξi−2,…,ξ0,0L−i)\overline{\Xi}_i := (\xi_{i-1},\xi_{i-2},\ldots,\xi_0,0_{L-i}), i=1,…,Li = 1,\ldots,L, where 0L−i0_{L-i} is a zero vector of length L−iL-i. Let xt:=(xt,…,xt−(L−1))′\mathbf{x}_t := (x_t,\ldots,x_{t-(L-1)})' and ϵt:=(ϵt,…,ϵt−(L−1))′\boldsymbol{\epsilon}_t := (\epsilon_t,\ldots,\epsilon_{t-(L-1)})'. Additionally, we define bϵ:=Ξ‾bx\mathbf{b}_{\epsilon} := \overline{\Xi}\mathbf{b}_x. Consequently, we obtain:

yt=bx′xt≈(Ξ‾bx)′ϵt=bϵ′ϵt,y_t = \mathbf{b}_x'\mathbf{x}_t \approx(\overline{\Xi}\mathbf{b}_x)'\boldsymbol{\epsilon}_t = \mathbf{b}_{\epsilon}'\boldsymbol{\epsilon}_t,

where the approximation via the finite MA inversion of xtx_t holds if the filter coefficients decay to zero sufficiently rapidly or, equivalently, if LL is sufficiently large (exact results can be derived but are omitted here). Wildi [43, 44] proposes the following SSA criterion

min⁡bϵE[(zt+δ−yt)2](1)\min_{\mathbf b_\epsilon} E[(z_{t+\delta}-y_t)^2] \tag*{(1)}
bϵ′Mbϵ−ρ1bϵ′bϵ=0,(2)\mathbf b_\epsilon^\prime\mathbf M\mathbf b_\epsilon-\rho_1\mathbf b_\epsilon^\prime\mathbf b_\epsilon=0, \tag*{(2)}

where the MA approximation holds at the stationary point of the Lagrangian (see details below). The L×LL \times L autocovariance generating matrix M\mathbf M,

M=(00.5000⋯0000.500.500⋯000⋮⋮⋮⋮⋮⋱⋮⋮⋮00000⋯0.500.500000⋯00.50),\mathbf M= \begin{pmatrix} 0 & 0.5 & 0 & 0 & 0 & \cdots& 0 & 0 & 0\\ 0.5 & 0 & 0.5 & 0 & 0 & \cdots& 0 & 0 & 0\\ \vdots& \vdots& \vdots& \vdots& \vdots& \ddots& \vdots& \vdots& \vdots\\ 0 & 0 & 0 & 0 & 0 & \cdots& 0.5 & 0 & 0.5\\ 0 & 0 & 0 & 0 & 0 & \cdots& 0 & 0.5 & 0 \end{pmatrix},

satisfies bϵ′Mbϵ=∑k=1L−1bϵ,k−1bϵ,k\mathbf b_\epsilon^\prime\mathbf M\mathbf b_\epsilon=\sum_{k=1}^{L-1}b_{\epsilon,k-1}b_{\epsilon,k}, the first-order autocovariance of the predictor, assuming standardized white noise. The constraint (2) imposes a first-order autocorrelation (ACF(1)) of ρ1\rho_1 and is feasible if ∣ρ1∣<ρmax⁡(L)|\rho_1|<\rho_{\max}(L), the largest eigenvalue of M\mathbf M determining the Rayleigh quotient. Since ∣ρ1∣<ρmax⁡(L)|\rho_1|<\rho_{\max}(L), the matrix M−ρ1I\mathbf M-\rho_1\mathbf I is indefinite, with both positive and negative eigenvalues, which gives the SSA problem its GTRS structure.

Let ht(y∣bϵ)ht(y\mid\mathbf b_\epsilon) denote the (stationarity-invariant) expected duration between consecutive zero-crossings of yty_t, the HT.4 Assuming Gaussianity,

ht(y∣bϵ)=πarccos⁡(ρ(y)),(3)ht(y\mid\mathbf b_\epsilon)=\frac{\pi}{\arccos(\rho(y))}, \tag*{(3)}

linking HT and ACF(1) (Wildi [43]); this relation is fairly robust to non-Gaussianity (central limit theorem, see Wildi [45]). Since arccos is monotonic on ]−1,1[] -1,1[, the criterion becomes

min⁡bϵ E[(zt+δ−yt)2]subject to ht(y∣bϵ)=ht1\begin{aligned} \min_{\mathbf b_\epsilon}\ &E[(z_{t+\delta}-y_t)^2]\\ \text{subject to }&ht(y\mid\mathbf b_\epsilon)=ht_1 \end{aligned}

Specifying ht1>htMSEht_1>ht_{MSE} (ht1<htMSEht_1<ht_{MSE}), the HT of the MSE predictor, yields a smoother (unsmoother) predictor with fewer (more) ZCs. The resulting SSA predictor minimizes MSE subject to the ZC rate; dually, it minimizes the ZC rate for given MSE (efficient frontier). The MSE predictor is recovered by setting ht1=htMSEht_1=ht_{MSE}. Wildi [43, 45] proposes variants of the SSA criterion in which the MSE objective (1) is replaced by either the target correlation, i.e., the correlation between the predictor yty_t and the target zt+δz_{t+\delta}, or the sign accuracy, i.e., the probability that predictor and target have the same sign (assuming Gaussianity). Since all variants are equivalent, SSA unifies MSE, target correlation, sign accuracy, and the rate of zero-crossings within a single interpretable framework.

I-SSA

References

  1. [1]Adachi, S., Iwata, S., Nakatsukasa, Y., and Takeda, A. (2017). Solving the trust-region sub-problem by a generalized eigenvalue problem. SIAM Journal on Optimization, 27(1).
  2. [3]Baxter, M. and R.G. King (1999). Measuring business cycles: Approximate band-pass filters for economic time series. Review of Economics and Statistics 81, 575–593.
  3. [4]Box, G. E. P., and Tiao, G. C. (1977). A canonical analysis of multiple time series. Biometrika, 64(2).DOI
  4. [6]Cornea-Madeira, Adriana (2017). The Explicit Formula for the Hodrick-Prescott Filter in Finite Sample. Review of economics and statistics, 314-318.
  5. [7]Drehmann, M., and Yetman, J. Which Credit Gap Is Better at Predicting Financial Crises? A Comparison of Univariate Filters. International Journal of Central Banking, 17, 4, October 2021.
  6. [8]Durbin, J. and Koopman, S. J. (2012). Time Series Analysis by State Space Methods, 2nd edn. Oxford: Oxford University Press.DOI
  7. [9]Eilers, Paul H.C. et al. (2015). Twenty years of P-splines. SORT-Statistics and Operations Research Transactions. 39,2, pp. 149-86, https://raco.cat/index.php/SORT/article/view/302258.
  8. [10]Gröchenig, K. (2010). Wiener’s lemma: theme and variations. An introduction to harmonic analysis in the style of Wiener, in Four Short Courses on Harmonic Analysis it Birkhäuser (Applied and Numerical Harmonic Analysis), 175–234.
  9. [11]Gröchenig, K. and Klotz, A. (2010). Noncommutative approximation: inverse-closed subalgebras and off-diagonal decay of matrices. Constructive Approximation. 32, 429-466.arxiv.org/abs/0904.0386
  10. [12]Granger, C. W. J. (1966). The Typical Spectral Shape of an Economic Variable. Econometrica.34, 1, https://doi.org/10.2307/1909859.
  11. [13]Hamilton, J.D. (2018). Why you should never use the Hodrick-Prescott filter. Review of Economics and Statistics 100, 831–843.
  12. [14]Harvey, A. (1989). Forecasting, structural time series models and the Kalman filter. Cambridge: Cambridge University Press.
  13. [15]Henderson, R. (1924). A New Method of Graduation. Transactions of the Actuarial Society of America 25, 29–40.
  14. [17]Hodrick, R. and Prescott, E. (1997) Postwar U.S. business cycles: an empirical investigation. Journal of Money, Credit, and Banking 29, 1–16.
  15. [19]Johansen, S. (1988). Statistical analysis of cointegration vectors. Journal of Economic Dynamics and Control, 12(2-3).
  16. [20]Kamber, G. Morley, J. Wong, B. (2024). Trend-Cycle Decomposition in the Presence of Large Shocks. CAMA Working Papers 2024-24. Centre for Applied Macroeconomic Analysis, Crawford School of Public Policy, The Australian National University, revised Aug 2024.DOI
  17. [21]Johansen, S. (1991). Estimation and hypothesis testing of cointegration vectors in Gaussian vector autoregressive models. Econometrica, 59(6).DOI
  18. [23]McElroy, T. (2008) Exact formulas for the Hodrick-Prescott filter. The Econometrics Journal 11,1, 209-217, https://www.jstor.org/stable/23116069.DOI
  19. [24]McElroy, T and Wildi, M. (2010) Signal Extraction Revision Variances as a Goodness-of-Fit Measure. Journal of Time Series Econometrics,2,1.
  20. [25]McElroy, T. (2022) Casting vector time series: algorithms for forecasting, imputation, and signal extraction. Electron. J. Statist. 16 (2), 5534-5569. DOI: 10.1214/22-EJS2068DOI
  21. [26]McElroy, T.S. and Livsey, J.A. (2022) Ecce Signum: An R Package for Multivariate Signal Extraction and Time Series Analysis arXiv:2201.02148, https://arxiv.org/abs/2201.02148
  22. [27]McElroy, T. and Wildi, M. (2016) Optimal Real-Time Filters for Linear Prediction Problems. Journal of Time Series Econometrics 8, 155-192.DOI
  23. [28]McElroy, T. and Wildi, M. (2019) The trilemma between accuracy, timeliness and smoothness in real-time signal extraction. International Journal of Forecasting , 35 (3), 1072-1084.
  24. [30]More, J. J. (1993). Generalizations of the trust region problem. Optimization Methods and Software, 2(3-4).DOI
  25. [31]Percival, D. B. and Walden, A. T. (2000) Wavelet Methods for Time Series Analysis. Cambridge: Cambridge University Press.
  26. [33]Phillips, P. C. B., and Shi, Z. (2021) Boosting: Why You Can Use the HP Filter. International Economic Review, 62(2), 521–570.
  27. [34]Mei, Z., and Phillips, P. C. B., and Shi, Z. (2024). The boosted Hodrick-Prescott filter is more general than you might think. Journal of Applied Econometrics, 39(7), 1260–1281.
  28. [35]Polik, I. and Terlaky, T. (2007) A survey of the S-lemma. SIAM Review, vol. 49, no. 3.DOI
  29. [36]Pollock, D. S. G. (2000) Trend estimation and de-trending via rational square-wave filters. Journal of Econometrics, 99(2), 317-334.DOI
  30. [38]Pong, T. K. and Wolkowicz, H. (2014) The generalized trust region subproblem. Computational Optimization and Applications, 58,2.DOI
  31. [40]Rice,S.O. (1944) Mathematical analysis of random noise. J. Bell. Syst. Tech. J 23, 282-332.
  32. [42]Whittaker, E. T. (1923) On a new method of graduation. Proceedings of the Edinburgh Mathematical Society, 41, 63-75.DOI
  33. [43]Wildi, M. (2024) Business Cycle Analysis and Zero-Crossings of Time Series: a Generalized Forecast Approach. J. Bus. Cycle Res. https://doi.org/10.1007/s41549-024-00097-5.
  34. [44]Wildi, M. (2026a) Sign Accuracy, Mean-Squared Error and the Rate of Zero Crossings: a Generalized Forecast Approach. arXiv:2601.06547. https://doi.org/10.48550/arXiv.2601.06547.
  35. [45]Wildi, M. (2026b) The Accuracy Smoothness Dilemma in Prediction: A Novel Multivariate M-SSA Forecast Approach. Journal of Time Series Analysis, https://onlinelibrary.wiley.com/doi/abs/10.1111/jtsa.70058.
  36. [46]Wood, S. N. (2017) Generalized Additive Models: An Introduction with R. 2nd edn. Boca Raton, FL: Chapman and Hall/CRC.

Paper details

vlnc:W_fczf5nvu
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.

  • Economics, Econometrics and Finance

Not recorded for this paper

View statement
Looking for related papers…
Contents