Why Stock Returns May Be Predictable Only in Brief Windows

Stock Returns May Be Predictable, but Only in Brief Windows

Investors have spent decades searching for indicators that reliably predict stock returns. Valuation ratios, interest rates, credit spreads, inflation, market volatility, and measures of economic activity have all shown promise at one time or another. The difficulty is that an indicator can appear useful in one period and become ineffective in the next.

A new study published in the Journal of Econometrics¹ argues that this instability may be a defining feature of market predictability rather than a defect in the data. Stock returns might be predictable, but only during temporary episodes created by recessions, financial shocks, policy changes, or shifts in how investors process information.

The researchers developed statistical monitoring procedures intended to detect these episodes as they emerge. Their work reframes the practical question. Instead of asking whether a variable predicts returns across an entire historical sample, investors should ask whether it has recently begun to contain useful information.

Why Permanent Market Signals Are So Difficult to Find

Traditional predictive research often estimates one relationship across decades of data. For example, an analyst might test whether dividend yields predict the equity premium, meaning the return stocks generate above a risk-free asset. This assumes that the relationship remains sufficiently stable throughout the sample.

Markets rarely cooperate. Monetary policy changes, regulations evolve, and widely discovered strategies attract capital. A signal can strengthen during a crisis and disappear once investors recognize it. A full-period model may therefore dilute a useful episode with many years in which no relationship existed.

This helps explain why financial variables can look statistically significant in one study and fail in another. The disagreement may not mean one result was necessarily wrong. The studies may have examined different market regimes.

The challenge becomes harder because many common predictors are highly persistent. Dividend yields, book-to-market ratios, and interest-rate measures can remain elevated or depressed for long periods. They may also move alongside the same shocks affecting returns. Conventional statistical tests can mistake those characteristics for predictive power.

How Real-Time Monitoring Changes the Problem

The paper proposes monitoring incoming data sequentially after establishing an initial training period with no detected predictability. Its preferred approach uses a cumulative sum, or CUSUM, process built from IVX instrumental-variable conditions.

In simpler terms, the system keeps track of whether the relationship between a candidate indicator and subsequent returns is drifting far enough from its historical baseline to warrant a signal. It is designed to work without requiring the user to know in advance whether the predictor is weakly persistent, close to a unit root, or somewhere between those cases.

The monitoring framework addresses several problems simultaneously:

• Predictive relationships may begin and end without warning.
• Repeated testing can generate false alarms even when no signal exists.
• Changes in volatility can make ordinary tests detect relationships that are not real.
• Predictors can be persistent and correlated with unexpected market returns.

The volatility adjustment is especially important. A major market shock creates larger observations, which can push a monitoring statistic across a threshold even if the underlying predictive relationship has not changed. This is related to the broader problem that familiar gauges may capture only part of the risk environment, as discussed in Securities.io’s examination of why the VIX does not fully capture stock market risk.

What the Simulations Found

The authors tested their methods across simulated markets with different predictor persistence, volatility paths, endogeneity relationships, and starting dates for predictive episodes. Several proposed CUSUM procedures maintained reasonably controlled false-positive rates across many of these conditions.

The strongest methods generally detected predictability earlier than competing rolling-window procedures, particularly when the predictor was extremely persistent. This matters because a short-lived signal loses practical value if a model identifies it only near the end of the episode.

However, no version dominated under every condition. One procedure reacted particularly quickly but produced excessive false positives after an upward volatility shift. Another was more conservative and slower. The paper’s preferred variance-adjusted CUSUM approach offered a more useful balance between detection speed and false-alarm control.

A model that signals too easily encourages costly turnover and overconfidence. A cautious model may recognize a regime only after much of the opportunity has passed. Real-time systems therefore need an explicit policy for choosing between missed signals and false alarms.

What Happened When the Model Watched the S&P 500

The researchers conducted a pseudo-real-time exercise using monthly S&P 500 data and candidate predictors covering January 1965 through December 2024. The predictors included valuation ratios, volatility measures, technical indicators, bond yields, credit-market variables, inflation, and economic activity.

Full-sample testing found little evidence that most variables predicted returns consistently. Yet the monitoring procedures identified distinct windows. Several bond and macroeconomic indicators became informative around the 1973-1975 market disruption. The term spread showed predictive content around the period leading into Black Monday. Other signals appeared around the global financial crisis, 2018, and 2022.

These results support a regime-based interpretation. Credit returns, yield spreads, liquidity, valuation, and economic slack may matter most when the market is repricing risk. They need not forecast returns during ordinary conditions to become useful during periods of stress.

The Bigger Insight Is Signal Lifecycle Management

The study’s most valuable contribution is not a list of indicators to trade. It is a framework for managing the lifecycle of an investment signal.

A production model should have an activation rule, a retirement rule, and a method for recalibrating after a detected episode. That is different from building a forecasting model once and assuming its coefficients remain valid. It treats predictability as a changing system state.

This perspective also suggests that volatility serves two roles. It may contain information about expected returns, but it can also corrupt the statistical test used to detect that information. Research covered by Securities.io has similarly found that daily returns can remain valuable for volatility forecasting even when higher-frequency datasets are available. More data does not automatically solve model instability if the relationships within that data keep changing.

For investment teams, the logical next step would be a layered system. One model would monitor whether a candidate predictor has entered an active regime. A separate forecasting layer would estimate the direction and magnitude of expected returns. Portfolio construction would then account for trading costs, drawdowns, liquidity, and position limits.

That separation is essential because statistical predictability is not the same as a profitable strategy. The paper does not calculate portfolio returns, implementation costs, taxes, or risk-adjusted performance. Its historical exercise is pseudo-real-time rather than a live deployment. Claims that the method can beat the market would go beyond the evidence.

Investing in the Financial Intelligence Layer

The study points toward growing demand for cleaner financial data, adaptive analytics, and tools capable of recognizing regime changes. One company positioned within that infrastructure is S&P Global (SPGI ).

S&P Global operates across benchmarks, ratings, commodity intelligence, and market data. Its Market Intelligence division supplies many of the datasets and analytical workflows institutions use to research companies, credit conditions, and market behavior. The company has also expanded its AI capabilities through Kensho, including tools designed to connect complex corporate records and make financial datasets easier for machines to interpret. In 2026, S&P Global described an expanded version of Kensho Link with broader entity coverage and improved recognition across complex corporate structures.

This does not mean S&P Global owns or commercializes the monitoring method described in the study. The investment connection is broader: adaptive financial models depend on extensive, standardized, timely data. As quantitative systems become more responsive to market regimes, data quality and integration become part of the competitive advantage.

Securities.io’s recent profile of MSCI and the infrastructure behind financial markets highlights the same underlying trend. Index providers increasingly operate as data and analytics companies rather than simple publishers of benchmark lists.

Market Predictability Is a Moving Target

The paper does not resolve the efficient-market debate, but it offers a more realistic way to frame it. Markets can be broadly difficult to predict while still producing brief periods in which economic and financial indicators become unusually informative.

For investors, the implication is not to chase every newly significant variable. It is to treat signals as conditional, monitor their stability, and recognize that a relationship discovered during one crisis may not survive the next expansion. The durable advantage may be less about finding one permanent forecasting formula and more about detecting when the market’s rules have temporarily changed.

References:

¹ Demetrescu, M., Schmidt, F., & Taylor, A. M. R. (2026). Real-time monitoring for stock return predictability in nonstationary volatility environments. Journal of Econometrics, 258, 106341. https://doi.org/10.1016/j.jeconom.2026.106341