The question
Industry portfolios sit closer to the real economy than the market index does. Oil and metals feel input-price shocks first, retailers see shifts in consumer demand first, and banks see credit conditions first. If that sector-level news reaches the aggregate market only gradually, lagged industry returns should help forecast the market. That is the idea behind Hong, Torous and Valkanov (2007), who found that 14 of 34 U.S. industries predicted next-month market returns over 1946–2002.
The finding has not aged well. Tse (2015) extended the sample to 2013, used 48 industries, and found only one to seven with significant predictive power, along with stronger predictability running the other way, from the market to industries. So do industries really help forecast the market, or was the original result a feature of one sample? We argue that a large part of the answer lies in how the market return is modelled.
Why the signal can hide
A monthly market return mixes two very different things. One is a slow-moving valuation component, tied to earnings and dividend yields, that is highly persistent and accounts for most of the average return. The other is the short-run repricing of the index, which is close to unpredictable from one month to the next and accounts for almost all of the variance. A standard predictive regression asks a single coefficient on a lagged industry return to explain both at once.
The economics suggests where industry information should show up. In the gradual-diffusion model of Hong and Stein (1999), investors hold different pieces of information and do not fully infer what others know from prices, so news is incorporated into some prices before others. Applied here, sector news is priced first in industry portfolios and only later in the index. Lagged industry returns should therefore be informative about the short-run repricing of the market, not about its persistent valuation component. Forecasting the components separately gives that information a way through.
What we do
Data. We use monthly value-weighted returns on the Fama–French 49 industry portfolios from Kenneth French’s data library, 1926:07–2024:12, keeping the 40 industries with complete coverage. Returns are measured in excess of the one-month Treasury bill rate. Aggregate prices, dividends and earnings come from Amit Goyal’s data library.
Three ways to forecast. For each industry we compare three forecasts of next month’s market excess return.
- Direct regression (DR). The textbook approach: regress the market excess return on the lagged industry return.
- Return-component decomposition (RCD). Following Campbell and Shiller (1988) and Lin et al. (2025), we split the market return into a fundamental component, built from the earnings and dividend yields, and an unexpected capital-gain component. The fundamental part is forecast by its current value; only the capital-gain part is forecast with the lagged industry return. The two pieces are then added back together.
- Sum of the parts (SOP). Following Ferreira and Santa-Clara (2011), the return is built from the dividend–price ratio, earnings growth (a 20-year moving average) and growth in the price–earnings multiple. Industries are used only to forecast multiple growth, and the risk-free rate is subtracted at the end.
Evaluation. Everything is estimated recursively, using only information available on each forecast date. The expanding window starts with 234 months, so the out-of-sample period runs from 1946 to 2024. Each forecast is compared with the recursively updated historical mean using the Campbell–Thompson out-of-sample R², with Clark–West tests, a more conservative moving-block bootstrap, and multiple-testing corrections across the 40 industries. We also report versions with a Campbell–Thompson shrinkage adjustment that keeps forecasts economically sensible (RCDs and SOPs), pool the industry forecasts with eight combination schemes, and measure economic value for a mean–variance investor with a risk aversion of five who can hold between 0% and 150% in equities.
What we find
Finding 1. Direct regressions confirm the sceptics.
In sample, 11 of the 40 industries clear the 0.5% R² bar that Campbell and Thompson (2008) regard as economically meaningful, led by electrical equipment at 1.54%. Out of sample, the direct regressions mostly fail to beat the historical mean: most out-of-sample R² values are negative, and the bootstrap finds a significant improvement for none of the 40 industries. On this evidence alone, one would side with Tse (2015) and conclude that industries add little.
Finding 2. Component-based forecasts recover the signal.
Applied to the same industries, the component-based designs produce positive out-of-sample R² for most of them, typically around 1%, where the direct regressions range from about +0.3% down to −1.7%. Books (1.73% under SOP), banks, retail and chips are among the strongest. The moving-block bootstrap, which is stricter than Clark–West, finds significant gains for 4 industries under RCD, 13 under RCDs, 11 under SOP and 14 under SOPs, against none under the direct regression.
Notes: Campbell–Thompson out-of-sample R² of the equal-weighted combination of the 40 industry forecasts, relative to the historical mean, 1946–2024. All four component-based values are significant at the 1% level (Clark–West); the direct-regression value is not. Source: Table 4.
Data table
| Mean-combination forecast | |
|---|---|
| Direct regression (DR) | −0.11 |
| Return-component decomposition (RCD) | 0.92 |
| RCD with shrinkage | 1.02 |
| Sum of the parts (SOP) | 1.30 |
| SOP with shrinkage | 1.29 |
Pooling tells the same story. Every combination of the direct-regression forecasts is negative or negligible. Combinations of the component-based forecasts are positive and significant under almost every scheme, reaching about 1.3% for SOP, with a peak of 1.49%. Pooled forecasts do not beat the best single industry, but they remove the need to pick one in advance.
Finding 3. The gains are economically meaningful.
An R² of 1% sounds small, but for monthly market returns it is not. For our mean–variance investor, timing the market with direct industry forecasts adds little relative to the historical-mean strategy, and loses value for some industries. The component-based forecasts raise the certainty-equivalent gain to roughly 2–3% a year for most industries before costs, and the gains stay positive for almost all industries after a deliberately conservative 25 basis points per transaction. A simple long-only switching rule based on SOPs forecasts earns 7.39% a year with a Sharpe ratio of 0.55, against 6.94% and 0.47 for buy-and-hold, and 37 of the 40 industry strategies significantly beat the historical-mean timing benchmark.
When it works, and when it does not
The gains are not constant over time. Splitting the out-of-sample period, the component-based designs are strong in 1946–1975 (a mean SOP R² of 2.38%) and in 2001–2024 (1.24%, with 39 of the 40 industries positive), but weaker in 1976–2000. They are also concentrated in recessions: the mean SOP R² is 4.09% in NBER recession months, against 0.07% in expansions, and all 40 industries are positive in recessions. Unlike the direct regressions, though, the component-based forecasts remain positive for most industries in expansions as well.
- NBER recessions
- Expansions
Notes: Forecast errors are classified using the lagged NBER recession indicator: 124 recession and 824 expansion months. Source: Table 7, Panel B.
Data table
| NBER recessions | Expansions | |
|---|---|---|
| DR | 2.31 | −1.29 |
| RCD | 3.65 | −0.25 |
| RCD with shrinkage | 3.23 | 0.12 |
| SOP | 4.09 | 0.07 |
| SOP with shrinkage | 3.51 | 0.36 |
Other checks leave the conclusion unchanged. Moving the start of the out-of-sample period from 1946 to 1950 or 1955 barely changes the SOPs results (a mean R² of 1.04–1.05%, with all 40 industries significant). Using the coarser Fama–French 30-industry classification gives similar results. At a two-month horizon the component-based forecasts still beat the benchmark on average; at three months the performance is weaker.
Takeaways
- Industry information has not gone away. Whether you find it depends on how the market return is modelled: a direct regression buries it, a component-based design recovers it.
- The signal sits in short-run repricing. That is where gradual information diffusion says it should be, and it helps reconcile Hong, Torous and Valkanov (2007) with Tse (2015).
- Sum of the parts is the strongest design. RCD and SOP agree, so the result does not hinge on one decomposition.
- Expect time variation. Predictability is strongest in recessions and in the earliest and most recent decades. It is real, but it is not constant.
A natural next step is to use machine-learning methods that draw on the whole cross-section of industries at once, rather than one industry at a time.
Key references
- Campbell, J. Y., and Shiller, R. J. (1988). The dividend-price ratio and expectations of future dividends and discount factors. Review of Financial Studies, 1(3), 195–228.
- Campbell, J. Y., and Thompson, S. B. (2008). Predicting excess stock returns out of sample: Can anything beat the historical average? Review of Financial Studies, 21(4), 1509–1531.
- Clark, T. E., and West, K. D. (2007). Approximately normal tests for equal predictive accuracy in nested models. Journal of Econometrics, 138(1), 291–311.
- Ferreira, M. A., and Santa-Clara, P. (2011). Forecasting stock market returns: The sum of the parts is more than the whole. Journal of Financial Economics, 100(3), 514–537.
- Hong, H., and Stein, J. C. (1999). A unified theory of underreaction, momentum trading, and overreaction in asset markets. Journal of Finance, 54(6), 2143–2184.
- Hong, H., Torous, W., and Valkanov, R. (2007). Do industries lead stock markets? Journal of Financial Economics, 83(2), 367–396.
- Tse, Y. (2015). Do industries lead stock markets? A reexamination. Journal of Empirical Finance, 34, 195–203.
- Welch, I., and Goyal, A. (2008). A comprehensive look at the empirical performance of equity premium prediction. Review of Financial Studies, 21(4), 1455–1508.