Forecasting Returns vs Forecasting Risk
Predicting where a price is going and predicting how much it might move are different problems with different levels of difficulty — risk forecasts are typically far more reliable than return forecasts, and a research process should treat them differently.
Prerequisites: Choosing What to Forecast
Ask a hundred researchers to forecast next month's return for a stock and their predictions will scatter widely, most of them close to noise. Ask the same hundred to forecast next month's volatility and their predictions will cluster much more tightly, most of them reasonably close to the truth. Returns and risk are both "forecasts," but they are not equally hard, and treating them as the same kind of problem is a common source of misplaced confidence.
Why risk is the easier target
Returns are close to unpredictable at short horizons because if they weren't, the predictable part would already have been traded away by someone. Volatility, by contrast, is strongly autocorrelated — a stock that has been moving a lot recently tends to keep moving a lot for a while, a pattern known as volatility clustering, and it holds up remarkably consistently across markets and time periods. A simple forecast like "next month's volatility will be close to a weighted average of the last few months' realized volatility" already captures most of the predictable variation. No comparably simple rule works nearly as well for returns.
This asymmetry has a direct practical consequence: a research desk should expect real, useful skill in risk forecasting, and should be far more skeptical of claimed skill in return forecasting of similar apparent strength — an that would be a strong result for volatility would be an implausibly, suspiciously strong result for returns.
Risk forecasting and return forecasting sit at different points on the difficulty spectrum. A single portfolio process can and should use a high-confidence risk forecast (position sizing, portfolio volatility targeting) alongside a low-confidence, heavily shrunk return forecast (the actual trading signal) — conflating the two invites over-betting on the shakier of the two.
A worked example
A researcher backtests a volatility forecast (a simple exponentially-weighted average of squared past returns) against realized next-month volatility across 500 names and finds a correlation of 0.65 — strong, and consistent with what's typically seen in practice. The same researcher backtests a return forecast built from a comparably simple model and finds a correlation of 0.65 with next-month realized return. The second number should trigger immediate suspicion rather than celebration — a return-forecast correlation that high is far outside what any known, real signal achieves, and is a strong hint of a look-ahead leak in the pipeline, not a breakthrough.
When comparing two forecasts' apparent skill, first check whether they're forecasting the same kind of target. An "impressive" return next to a "modest" volatility may mean the return number is broken, not that the volatility model is weak.
Related concepts
Practice in interviews
Further reading
- Grinold & Kahn, Active Portfolio Management (ch. 3 and ch. 5)