Choosing the Target: Raw, Excess or Residual Returns

Before a single feature is engineered, an ML researcher building an equity model has to answer a question that gets less attention than it deserves: predict what, exactly? Next week's raw return, the return relative to the market, or the return left over after a risk model has already explained away style and sector effects? Each choice trains the model to learn something different, and the wrong choice can make an otherwise good model useless for its intended job.

Why the target changes what the model learns

A model trained to predict raw next-month returns will happily learn that "everything went up when the market went up" — a large share of its apparent predictive power can come from simply detecting market direction, which a portfolio manager building a long-only fund might want, but which a market-neutral book already hedges away and gets zero benefit from. Training on excess returns (stock return minus the market return) removes that market component. Training on residual returns — what's left after a full risk model strips out market, sector, size, value, momentum and other known factors — pushes the model even further, toward whatever the risk model can't already explain.

Think of the total return as a stack of components that a risk model can peel apart layer by layer; the choice of target decides how many layers you strip off before handing the model what's left to predict.

Worked example

A researcher has a five years of monthly stock returns and a Barra-style risk model. She trains the same gradient-boosted model three times, changing only the target.

1. Raw return target. Reported test-set IC of 0.09 — but nearly all of it evaporates when returns are decomposed, because the model was mostly detecting which months the market as a whole was up or down and assigning that to every stock.
2. Market-excess return target. IC drops to 0.05 — market timing is gone, but the model may still be partly reflecting sector rotations (it learned that whole sectors moved together, not which specific stocks within a sector were mispriced).
3. Full risk-model residual target. IC drops further to 0.03, but this smaller number is the one that survives being combined into a market-neutral, sector-neutral portfolio — the earlier two ICs were partly illusions that a neutralized portfolio could never have captured anyway.

The third model looks "worse" by raw IC, yet it is the only one whose reported skill translates into a genuinely tradeable market-neutral signal.

What this means in practice

The right target depends on how the signal will ultimately be used. A long-only fund happy to carry some market and sector exposure can reasonably train on excess returns and let a portfolio construction step handle the rest. A market-neutral book that will neutralize every factor exposure anyway should train on residual returns from the start — otherwise the backtest reports skill the live portfolio can never actually harvest, because the neutralization step removes precisely the part of the prediction that was driving the reported IC.