📈 economics

The Accuracy Smoothness Dilemma in Prediction: a Novel Multivariate M-SSA Forecast Approach

This paper introduces a novel Multivariate M-SSA forecasting approach that resolves the accuracy-smoothness dilemma by extending the Smooth Sign Accuracy framework to incorporate cross-sectional information, thereby balancing sign accuracy, mean squared error, and sign change frequency across multiple time series for applications in forecasting, nowcasting, and smoothing.

Original authors: Marc Wildi

Published 2026-02-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Marc Wildi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the weather for next week. You have two main goals, but they often fight against each other:

Accuracy: You want your prediction to hit the exact temperature every single day.
Smoothness: You want your prediction to feel natural, not jumping wildly from "freezing" to "boiling" every hour.

If you focus only on Accuracy, your forecast might look like a jagged, chaotic scribble. It might be mathematically "correct" on average, but it's so noisy that you can't trust it. If you focus only on Smoothness, your forecast might be a calm, straight line, but it will completely miss the actual storms and heatwaves.

This is the "Accuracy-Smoothness Dilemma."

The Old Way: The "Perfect" but Noisy Crystal Ball

Traditional forecasting methods (like the standard "MSE" or Mean Squared Error) are obsessed with Accuracy. They try to minimize the distance between their guess and the real number.

Think of this like a driver who is so focused on staying exactly in the center of the lane that they are constantly jerking the steering wheel left and right to correct tiny deviations. They are technically "on track," but the ride is terrifyingly bumpy. In economics, this is bad because it creates "false alarms"—predicting a recession when the economy is actually just having a bad Tuesday.

The New Way: The "M-SSA" Compass

The paper introduces a new method called M-SSA (Multivariate Smooth Sign Accuracy). Instead of just trying to hit the exact number, M-SSA asks a different question: "How often does my prediction flip-flop?"

It introduces a concept called Holding Time (HT).

Low Holding Time: Your prediction flips from positive to negative (or up to down) constantly. It's jittery.
High Holding Time: Your prediction stays in one direction for a long time. It's smooth and stable.

The Analogy: The Hiker vs. The Ant

The Old Method (MSE) is like an Ant. It takes tiny, frantic steps to stay exactly on a specific leaf. If the wind blows the leaf, the ant scrambles instantly. It's very precise but very jittery.
The New Method (M-SSA) is like a Hiker. The hiker wants to get to the destination (Accuracy), but they also want to walk on a steady path. They don't care about every single pebble; they care about the general direction. If the path wiggles a bit, the hiker keeps walking smoothly rather than jumping around.

The "Multivariate" Magic: Using the Whole Team

The paper's biggest innovation is making this work for multiple data sources at once (Multivariate).

Imagine you are a coach trying to predict the performance of a sports team.

Old Way: You look at the star player's stats alone. You get a prediction, but it's noisy because one player can have a bad day.
M-SSA Way: You look at the whole team. You see that the star player is struggling, but the defense is strong, and the weather is good. M-SSA combines all these signals.

It uses the "team" to smooth out the noise of the "individual." If the star player's data is jittery, the coach uses the steady rhythm of the defense to calm the prediction down, without losing the ability to predict the final score.

Real-World Examples from the Paper

Forecasting (Predicting the Future):
- Scenario: Predicting next month's industrial production.
- Result: The old method (MSE) might predict a sudden crash next week, then a sudden boom the week after. The M-SSA method says, "Hold on, the trend is actually steady." It reduces the "false alarms" of sudden crashes that don't actually happen.
Smoothing (Cleaning Up the Past):
- Scenario: Looking at historical economic data to find the "real" trend, removing the noise of daily fluctuations.
- Result: Traditional filters (like the famous HP filter) are like a heavy-handed editor who cuts out too much detail. M-SSA is like a skilled editor who removes the typos (noise) but keeps the story's flow (the business cycle) intact. It ensures the "story" of the economy doesn't flip-flop unnecessarily.
Signal Extraction (Finding the Needle in the Haystack):
- Scenario: Using a "Leading Indicator" (like a survey of business confidence) to predict the main economy.
- Result: The paper shows that by combining the main data with the leading indicator, M-SSA can predict a recession two months earlier than standard methods, but without the jittery noise. It's like hearing a distant rumble of thunder (the leading indicator) and knowing a storm is coming, rather than waiting for the rain to hit your face.

The Bottom Line

The paper argues that in economics and forecasting, being "smooth" is just as important as being "accurate."

If your prediction is too noisy, decision-makers (like central bankers or CEOs) will panic and make bad choices based on false signals. M-SSA gives them a tool to balance the two: it keeps the prediction accurate enough to be useful, but smooth enough to be trustworthy. It's the difference between a chaotic, jittery GPS that recalculates every second, and a reliable navigation system that guides you steadily to your destination.

1. Problem Statement

The paper addresses a fundamental trade-off in time series forecasting and signal extraction known as the Accuracy-Smoothness (AS) Dilemma.

The Conflict: Conventional forecasting methods, typically optimized to minimize Mean Squared Error (MSE), prioritize accuracy (tracking the true level of a series) often at the expense of smoothness. This can result in "noisy" predictors with excessive spurious sign changes (zero-crossings), which are undesirable in economic contexts (e.g., false alarms regarding business cycle turning points).
The Limitation of Existing Methods: Traditional smoothing techniques (like the Hodrick-Prescott filter) or trilemma-based approaches often lack a unified framework to simultaneously optimize for sign accuracy, MSE, and the frequency of sign changes (holding time) in a multivariate setting.
The Goal: To develop a framework that allows users to explicitly balance the trade-off between tracking accuracy and the smoothness of the predictor (measured by the duration between zero-crossings) across multiple time series.

2. Methodology: The M-SSA Framework

The author proposes the Multivariate Smooth Sign Accuracy (M-SSA) approach, an extension of the univariate SSA framework (Wildi, 2024, 2026).

Core Concepts

Smooth Sign Accuracy (SSA): Instead of minimizing MSE alone, the method maximizes the correlation between the predictor and the target (or the MSE predictor) subject to a constraint on the Holding Time (HT).
Holding Time (HT): Defined as the expected duration between consecutive zero-crossings of the predictor. Under Gaussian assumptions, HT is mathematically linked to the first-order autocorrelation (ACF) of the predictor: $ht = \pi / \arccos(\rho)$ .
The Trade-off: By constraining the HT (via the first-order ACF), the user can control the "smoothness" of the predictor. A higher HT constraint forces fewer zero-crossings (smoother), while a lower HT allows for more responsiveness (less smooth).

Mathematical Formulation

Univariate Case: The problem is formulated as maximizing the target correlation $\rho(y, z)$ subject to a fixed first-order ACF $\rho(y) = \rho_1$ and a length constraint.
Multivariate Extension (M-SSA):
- Let $x_t$ be a multivariate stationary process and $z_t$ the target.
- The predictor $y_t$ is a linear combination of past values using filter matrices $B_k$ .
- The optimization problem for the $i$ -th target is:
  $\max_{b_i} \gamma'_{i\cdot\delta} \tilde{I} b_i$
  Subject to:
  $b'_i \tilde{M} b_i = \rho_i \quad (\text{HT Constraint})$
  $b'_i \tilde{I} b_i = 1 \quad (\text{Length Constraint})$
- Here, $\tilde{I}$ and $\tilde{M}$ are Kronecker products involving the covariance matrix $\Sigma$ and the autocovariance generating matrix $M$ .
Solution: The paper derives a one-parameter solution using spectral decomposition. The optimal filter weights are expressed as a function of a Lagrange multiplier $\nu_i$ $ν_{i}$ . The solution traces an efficient frontier linking accuracy (target correlation) and smoothness (HT).
- Primal Form: Maximize accuracy subject to a smoothness constraint.
- Dual Form: Maximize smoothness (minimize zero-crossings) subject to an accuracy constraint.

Handling Non-Stationarity and Dependencies

The framework accommodates autocorrelated processes by utilizing the Wold decomposition (MA inversion) of the data-generating process. The target filter and the MA inversion are combined via convolution to derive the optimal coefficients.

3. Key Contributions

Multivariate Generalization: Extends the SSA framework to multivariate systems, allowing the integration of cross-sectional information (e.g., leading indicators) to improve the AS performance of forecasts.
Explicit AS Trade-off Management: Provides a rigorous mathematical mechanism to customize benchmark predictors (like MSE or HP filters) by shifting them along the efficiency frontier. Users can choose a predictor that is "smoother" than the MSE benchmark or "more accurate" than a standard smoother for a given HT.
Dual Interpretation: Establishes that M-SSA predictors are the smoothest possible solutions for a given level of tracking accuracy, distinguishing them from traditional curvature-based smoothers (like HP).
Unified Framework: Demonstrates that forecasting, nowcasting (real-time signal extraction), and smoothing are all special cases of the M-SSA criterion, differing only in the specification of the target filter $\Gamma_k$ (causal vs. acausal).

4. Results and Applications

The paper validates the M-SSA approach through three distinct application domains:

A. Forecasting (VAR(1) Example)

Setup: Applied to a bivariate VAR(1) process.
Findings: The M-SSA predictor successfully adapted to specific constraints. For one series, it "unsmoothed" the MSE benchmark (introducing more zero-crossings to match a lower HT constraint), while for the other, it "smoothed" the MSE benchmark (reducing zero-crossings).
Outcome: The method successfully manipulated the holding time while maintaining high target correlation, proving the flexibility of the efficient frontier.

B. Smoothing (Comparison with HP Filter)

Setup: Compared M-SSA smoothers against the classic Hodrick-Prescott (HP) filter on White Noise and VAR(1) data.
Findings:
- When matched for Holding Time, M-SSA achieved higher target correlations (lower MSE) than the HP filter.
- When matched for target correlation, M-SSA achieved longer Holding Times (smoother) than HP.
- While HP minimizes curvature (second-order differences), M-SSA minimizes zero-crossings. The paper argues that for business cycle analysis, controlling sign changes is often more relevant than minimizing curvature.

C. Signal Extraction (Nowcasting with Leading Indicators)

Setup: Nowcasting US Industrial Production (INDPRO) using the Composite Leading Indicator (CLI).
Comparison: Compared Univariate HP, Univariate SSA, Bivariate MSE, and Bivariate M-SSA.
Key Findings:
- Smoothness: The Bivariate M-SSA produced the smoothest nowcast (highest sample HT: ~24.5 vs. ~15.5 for HP-C).
- Accuracy: M-SSA achieved the second-highest target correlation, outperforming the univariate SSA and HP.
- Leading Indicator Effect: The inclusion of the CLI in the multivariate design eliminated the lag observed in the univariate SSA. The M-SSA nowcast was coincident with the classic HP filter but significantly smoother, effectively utilizing the leading information to reduce noise without sacrificing timeliness.
- Decomposition: The analysis showed that during crises, the CLI component drove the M-SSA signal, demonstrating the model's ability to dynamically weight leading indicators.

5. Significance and Conclusion

Theoretical Impact: M-SSA resolves the accuracy-smoothness dilemma by providing a continuous, parametric frontier rather than a single optimal point (like MSE). It formalizes the relationship between sign accuracy, MSE, and zero-crossing frequency.
Practical Utility: The approach offers a flexible tool for economists and analysts who need to balance the risk of false alarms (noise) against the need for timely detection of turning points.
Implementation: The author provides an open-source R package (R-package-SSA-Predictor) that implements these algorithms, allowing for the customization of standard filters (HP, Hamilton, BK) to meet specific AS requirements.
Conclusion: The M-SSA framework represents a significant advancement in predictive modeling, particularly for business cycle analysis, by prioritizing the sign accuracy and holding time of predictors, which are critical for decision-making in alternating growth and contraction phases.