Raw, Excess or Residual Return as the Target
Predicting a stock's raw return mostly means predicting the market's return, since the market moves most stocks together — excess and residual targets strip that shared component out so the model has to earn its keep on the part that's actually stock-specific.
Prerequisites: Choosing What to Forecast
On an average day, most of any single stock's return is explained by the market moving, not by anything specific to that company. Train a model to predict raw return and it will happily learn to predict the market instead — which is a much easier, higher-variance target, and one that has nothing to do with stock-picking skill. Choosing between raw, excess, and residual return as the modelling target is really a choice about what shared component to strip out before the model gets to work.
Three levels of stripping
Raw return is the return exactly as observed — no adjustment. It's dominated by the market factor, so a model trained on it will mostly end up being a disguised market-timing model, whether or not that was the intent.
Excess return subtracts a benchmark, usually the market or a relevant index, from the raw return. This removes the market's contribution but leaves sector, style, and other systematic tilts still mixed in with the stock-specific part.
Residual return goes further, subtracting out the return explained by a full factor model — market, sector, size, value, momentum, and whatever else is deemed systematic — leaving only the piece of the move a factor model can't already explain. This is the cleanest target for a signal that's meant to be genuine stock-specific alpha, because it removes every return component the researcher isn't trying to forecast.
The more you strip out before modelling, the smaller and noisier the remaining target gets, but the more directly the model's job matches what a market-neutral, factor-neutral strategy actually needs. Use raw or excess return when the strategy itself will carry factor exposure; use residual return when the strategy is meant to be pure stock selection.
A worked example
On a day the market is up 1.2%, a stock returns 1.9%. Its raw return is 1.9%. Its excess return over the market is 1.9% − 1.2% = 0.7%. Suppose a factor model further attributes 0.4 percentage points of that 0.7% to the stock's sector rallying that day; the residual return — the part the factor model can't explain — is 0.7% − 0.4% = 0.3%. A model trained on raw return would have most of its variance explained by the 1.2% market move, drowning out the 0.3% that's genuinely specific to the company. A model trained on residual return sees only that 0.3%, which is smaller and noisier to predict, but it's the only piece a market-neutral, sector-neutral strategy can actually capture as alpha.
Whatever target is chosen, use the same factor model at prediction time that was used to construct the target — a mismatch between the model that built the target and the model used to neutralise the live portfolio silently leaks factor exposure back in.
Related concepts
Practice in interviews
Further reading
- Grinold & Kahn, Active Portfolio Management (ch. 3, forecasting)