SSA Framework: Balancing Accuracy and Smoothness

• Smooth Sign Accuracy (SSA) is a forecasting framework that optimizes the trade-off between predictive accuracy and forecast smoothness using sign accuracy and autocorrelation constraints.
• It integrates mean-squared error, zero-crossing rates, and lag-one autocorrelation to tailor forecasts for both stationary and non-stationary time series.
• SSA’s Lagrangian-based optimization enables controlled filter design applied in areas like business cycle estimation and economic trend extraction.

The Smooth Sign Accuracy (SSA) framework is a principled approach to forecasting that addresses the trade-off between predictive accuracy and forecast smoothness, termed the accuracy-smoothness (AS) dilemma. By integrating sign accuracy, mean-squared error (MSE), and smoothness—quantified via the rate of zero crossings—the SSA framework generalizes traditional prediction criteria, offering a versatile mechanism suitable for stationary and non-stationary time series analysis. SSA’s mathematical structure enables applications ranging from filter design to business cycle estimation, supporting controlled monotonicity and curvature in the forecast paths (Wildi, 10 Jan 2026).

1. Mathematical Principles of the SSA Criterion

For a stationary zero-mean process, the SSA framework formalizes the predictor’s objectives and constraints as follows. Let $\epsilon_t$ denote i.i.d. noise, and consider the target

$$z_{t+\delta} = \sum_{k=-\infty}^{\infty} \gamma_k \epsilon_{t+\delta-k}$$

and the $L$-lag, one-sided predictor

$y_t = \sum_{k=0}^{L-1} b_k \epsilon_{t-k}.$

Define $$\gamma_\delta = (\gamma_\delta, \ldots, \gamma_{\delta+L-1})'$$ and $b = (b_0, \ldots, b_{L-1})'$. SSA optimizes $b$ to maximize sign accuracy, measured by the correlation between $y_t$ and $z_{t+\delta}$:

$$\rho(y, z) = \frac{b' \gamma_\delta}{\sqrt{(b'b)(\gamma_\delta' \gamma_\delta)}},$$

with the induced sign accuracy

$$SA(y_t) = P(y_t z_{t+\delta} > 0) = 0.5 + \frac{1}{\pi}\arcsin(\rho(y, z)).$$

Smoothness is controlled by the expected duration between zero-crossings (holding time):

$HT(y) = \frac{\pi}{\arccos (\rho(y))},$

where $\rho(y) = \text{Corr}(y_t, y_{t-1})$ is the lag-one autocorrelation of $y_t$.

The core “primal” SSA optimization is:

$$\begin{aligned} &\max_{b \in \mathbb{R}^L} \quad b' \gamma_\delta \ &\text{subject to} \quad b' M b = l \rho_1, \quad b'b = l, \end{aligned}$$

where $M$ is the lag-one autocorrelation matrix (with off-diagonal $(k, k+1)$ and $(k+1, k)$ entries $1/2$), $l > 0$ is a scaling parameter (set $l=1$ without loss of generality), and $\rho_1$ is the user-specified target smoothness. The smoothness constraint enforces a specific lag-one autocorrelation, controlling the rate of sign-flips in the forecast.

2. Interpretation of Accuracy, Smoothness, and Trade-off

The SSA criterion directly encodes the relationship among accuracy, smoothness, and the temporal regularity of the predictor. The maximization target $b' \gamma_\delta$ is monotonically equivalent to forecast accuracy through its correlation with the target; $b' M b / (b'b) = \rho(y)$ quantifies smoothness via lag-one autocorrelation; the scale normalization $b'b = l$ ensures invariance to amplification.

The sole hyperparameter, $\rho_1 \in [-1,1]$, determines the locus on the AS trade-off:

• $\rho_1$ set to the MSE-optimal value $\rho_\text{MSE}$ recovers the unconstrained minimum-MSE predictor.
• $\rho_1 > \rho_\text{MSE}$ enforces greater smoothness than the MSE solution (yielding low-pass-like forecasts).
• $\rho_1 < \rho_\text{MSE}$ increases oscillation (yielding high-pass-like forecasts).

No additional weights (e.g., $\alpha$, $\beta$ combining MSE and smoothness) appear in the formalism. The smoothness constraint suffices to generate the Pareto-optimal family.

3. Extensions to Dependent and Integrated Series

The SSA methodology is preserved for non-white, stationary, or integrated data:

• Dependent stationary series: For $x_t \approx \Xi\epsilon_t$ (finite Wold MA), recast $b_\epsilon = \Xi b_x$ and $\gamma_{\Xi\delta} = \Xi \gamma_{x\delta}$, and solve the SSA criterion for $b_\epsilon$.
• I(1) extension (maximal-monotone-SSA): For an I(1) process $\tilde{x}_t$, with $x_t = \Delta \tilde{x}_t$, estimate the MSE filter on $\tilde{z}_{t+\delta}$, impose the cointegration $\sum b_{xk} = \sum \gamma_{MSE, k} = \Gamma(0)$, and constrain the ACF of $\Delta y_t$.
• I(2) extension (lowest-curvature-SSA): Further impose $\sum k b_{xk} = \dot{\Gamma}(0)$, optimize for the ACF of $\Delta^2 y_t$.

Adjustments for cointegration align predictions with monotonicity or trend smoothness constraints, important in econometric filtering and business cycle estimation.

4. Optimization Algorithm

The SSA optimization employs a Lagrangian formulation:

$$\mathcal{L}(b, \lambda_1, \lambda_2) = b' \gamma_\delta - \lambda_1 (b'b - 1) - \lambda_2 (b'Mb - \rho_1).$$

Setting the gradient to zero yields the linear system:

$\gamma_\delta = 2\lambda_1 b + 2\lambda_2 M b.$

Diagonalization in the $M$ eigenbasis supplies a closed-form representation for $b(\nu)$ parameterized by a scalar dual variable $\nu$ related to the Lagrange multipliers:

$$b(\nu) = D \sum_{i=1}^L \frac{w_i}{2\lambda_i - \nu} v_i, \quad w_i = v_i' \gamma_\delta,$$

where $(\lambda_i, v_i)$ are the eigenpairs of $M$. The unique $\nu$ enforcing $\rho(y(\nu)) = \rho_1$ is found by root-finding, exploiting strict monotonicity. Each evaluation is $O(L)$ and convergence is logarithmic in tolerance.

5. Empirical Performance and Business Cycle Applications

Empirical application in business cycle nowcasting demonstrates the SSA filter’s adaptability. Using the two-sided HP(1600) trend as the acausal target, the one-sided SSA filter achieves targeted trade-offs:

	Target	MSE	SSA(0.97,0)	SSA(0.80,0)
$\rho(y, z)$	1.000	0.733	0.717	0.716
$SA(y)$	1.000	0.762	0.754	0.754
$ACF_1(y)$	0.996	0.926	0.970	0.800
$HT(y)$	34.32	8.14	12.79	4.88

Setting $\rho_1=0.97 > \rho_{MSE}$ yields a low-pass, smoother forecast; $\rho_1=0.80 < \rho_{MSE}$ accentuates oscillations. In I(1)–SSA filtering of US Industrial Production (with $\rho_1$ chosen to match HP-concurrent smoothness, $HT \approx 18.2$), the SSA filter lowers MSE by approximately 25% relative to the HP-concurrent filter (Wildi, 10 Jan 2026).

6. Theoretical Guarantees and Statistical Properties

SSA predictors exhibit uniqueness and finite-sample existence within the feasible range of $\rho_1$ (Corollary 2.4). The closed-form, one-parameter solution provides an explicit time-domain filter with AR(2) recursion and boundary conditions (Theorem 2.1). SSA admits a dual interpretation: it yields the most accurate predictor for a given smoothness, or the smoothest possible for a given accuracy (Theorem 2.7).

Given the target vector estimated from data, the SSA filter weights $\hat{b}(\nu)$ are multivariate normal, with mean $D N^{-1} \mu_\gamma$ and covariance $D^2 N^{-1} \Sigma_\gamma N^{-1}$, where $N = 2M - \nu I$ (Wildi, 10 Jan 2026). Robustness studies demonstrate that, even under heavy-tailed noise with degrees of freedom as low as $4$, empirical holding times of SSA predictors deviate by no more than 5% from Gaussian theoretical values.

7. Context and Relationship to Established Frameworks

The SSA framework generalizes classical MSE-based design (e.g., Wiener–Kolmogorov filters). By controlling the predictor autocorrelation, SSA unifies the selection of low-pass and high-pass forecast filters and enables monotonicity and curvature constraints central to economic latent trend extraction (e.g., alternative HP filtering). Its structural flexibility and rigorous optimality properties distinguish it from penalty-weighted smoothing approaches, offering unique leverage over the interplay between sign directionality and temporal regularity in forecasting (Wildi, 10 Jan 2026).