Value investing, which involves buying cheap stocks and selling expensive ones, has been one of the most durable ideas in empirical asset pricing. The value factor known as HML, originally introduced by Fama and French in 1992 and 1993 and sourced here directly from the Ken French Data Library, is well documented, intuitive, and historically profitable. However, as is well documented and painfully noted across the industry, value has its difficult periods, and it has underperformed for many years in recent decades. The characteristic proposed by Fama and French to proxy value is not a perfect measure. A long literature has produced many refinements and alternatives, from anchored valuation ratios in Lakonishok, Shleifer and Vishny (1994) to composite valuation signals that are now common in industry practice, as in Asness et al. (2013). Many institutional investors have therefore developed their own proprietary value proxies.
Our analysis uses the original HML construction for transparency and replicability, but the ideas we develop here, including sentiment conditioning, volatility information, and short horizon momentum cues, can be applied to any modern or proprietary value definition. Like all valuation based strategies, even the best versions remain highly cyclical, with periods of deep drawdowns driven by macro shocks or speculative surges that challenge even the most patient investors.
As a general disclaimer, we are not advocates of time series trading strategies. Even when the time series represents a long–short strategy, we are still timing the return of a portfolio through an overlay. Moreover, our core investment framework is entirely cross sectional. These results should be viewed as illustrative and as a potential complement to cross sectional strategies, rather than a standalone timing model.
To sum up, in this research note, we ask a simple question with practical importance:
Can sentiment help machine learning keep us from value’s worst periods while still capturing its long-run premium?
Our answer, based on a structured out-of-sample test, is cautiously optimistic.
Sentiment as a Cyclical Pressure Valve
A natural starting point is investor sentiment. Baker and Wurgler’s sentiment index, shown below alongside NBER recessions, captures waves of optimism and pessimism that move through markets. Periods of exuberance, such as the late 1960s, the late 1990s, and 2021, tend to compress valuation spreads and lift expensive “story” stocks.
Figure: Baker–Wurgler Sentiment with Recessions
This intuition is consistent with decades of research. High sentiment predicts lower returns for difficult-to-arbitrage stocks (Baker–Wurgler 2006), and the value spread tends to widen following sentiment peaks. When sentiment unwinds, 2000-2002, 2008-2009, or mid-2022, value typically rebounds sharply.
Given this backdrop, we explore whether changes in sentiment, combined with momentum and volatility characteristics of HML, can help predict near-term value performance.
A Random Forest Approach
We constructed a feature set of 60+ variables including:
Sentiment levels and changes over 1, 3, 6, and 12 months
HML factor momentum (lags and moving averages across horizons)
Volatility signals for HML and the equity market (1m, 3m, 6m, 12m realized vols)
Interaction terms, including sentiment × volatility
Crucially, the model is estimated fully out-of-sample: at each month, it only uses information available at that moment.
We evaluate three trading rules derived from the Random Forest predictions:
Binary Long/Flat (long HML when predicted positive)
Continuous Scaling (position proportional to predicted return)
Tercile Long/Short (long the top third of predictions, short the bottom third)
The third approach is the only explicitly long/short method, and it proves most informative.
Out-of-Sample Performance
The figure below plots cumulative returns for the strategies.
Figure: Random Forest HML Timing (OOS)
While static HML delivers a cumulative 1.82× return since inception, the machine learning strategies generate meaningfully higher returns, with the Tercile Long/Short model rising nearly 6× over the same period.
Table: Performance Table
A few points stand out:
The Tercile Long/Short strategy earns nearly twice the Sharpe ratio of static HML.
Max drawdown improves materially, consistent with avoiding the worst value crashes.
Even modest predictive accuracy (55%) yields substantial economic value—consistent with the idea that timing a cyclical factor requires only small but persistent edges.
What Drives the Predictions?
To shed light on the model’s inner workings, we examine feature importance.
Figure: Top 15 Feature Importances
The top predictors: HML_MA3, SENT × HML_vol, HML_vol_3m, MKT_vol_6m, Sentiment Change 12m, cluster around three themes:
Short-term mean reversion signals dominate
The 3-month moving average of HML returns (HML_MA3) is the single most important predictor. Value’s short-run swings tend to be sharp and mean reverting—captured cleanly by this signal.
Volatility matters, especially when combined with sentiment
HML volatility measures (1–12 months) appear repeatedly. The presence of the sentiment × HML volatility interaction suggests that volatility spikes during optimistic periods are particularly informative—often foreshadowing value reversals.
Long-horizon sentiment changes outperform sentiment levels
Sentiment trends over 6–12 months carry substantially more weight than the absolute sentiment index. This aligns with the intuition that accelerating optimism or pessimism is more predictive than level itself.
Main Takeaways
Our results suggest that machine learning can help investors navigate value’s cycles—not by forecasting fundamentals, but by exploiting behavioral regimes embedded in sentiment, momentum, and volatility.
A simple Random Forest model, using publicly available data and strictly out-of-sample evaluation, produced:
Higher returns,
Higher Sharpe ratios, and
Smaller drawdowns
than a passive exposure to the value factor.
This does not imply that timing value is easy, or that such strategies will always work. But the combination of behavioral indicators (sentiment) and technical factors (momentum & volatility) appears to contain economically meaningful information about when value is likely to shine and when caution is warranted.
For allocators willing to actively manage their factor exposures, these tools provide a useful direction for timing their cross-sectional strategies.
Curious to dig deeper? If you’d like to discuss this research further or explore potential collaborations, feel free to reach out at [email protected]
Visit our website here: www.noaxcapital.com
References
Asness, C. S., Moskowitz, T. J., & Pedersen, L. H. (2013). Value and Momentum Everywhere. Journal of Finance, 68(3), 929–985.
Baker, M., & Wurgler, J. (2006). Investor Sentiment and the Cross-Section of Stock Returns. Journal of Finance, 61(4), 1645–1680.
Baker, M., & Wurgler, J. (2007). Investor Sentiment in the Stock Market. Journal of Economic Perspectives, 21(2), 129–151.
Fama, E. F., & French, K. R. (1992). The Cross-Section of Expected Stock Returns. Journal of Finance, 47(2), 427–465.
Fama, E. F., & French, K. R. (1993). Common Risk Factors in the Returns on Stocks and Bonds. Journal of Financial Economics, 33(1), 3–56.
Lakonishok, J., Shleifer, A., & Vishny, R. W. (1994). Contrarian Investment, Extrapolation, and Risk. Journal of Finance, 49(5), 1541–1578.