Choosing the Prediction Target: Return, Rank or Direction

Two researchers both build a model on the same features and the same universe of stocks, and get very different results — not because one's model is better, but because one trained it to predict next week's raw return and the other trained it to predict which stocks will rank in the top 20% next week. Those are genuinely different prediction problems that happen to share the same input data, and picking the wrong one for what you actually plan to do with the output (build a portfolio, trade a single name, size a position) can waste a perfectly good model on the wrong question.

Three common target choices

• Raw forward return — predict the actual percentage return over some horizon. Most information-rich if you get it right, but the hardest target: returns are noisy, fat-tailed, and dominated by a small number of extreme days that can distort a model trained with a standard loss function.
• Cross-sectional rank — predict where a stock will fall relative to other stocks in the same period, not its absolute return. This throws away information about magnitude but sidesteps a lot of noise that affects the whole market at once (a broad market rally lifts every stock's raw return without saying much about which stock is relatively better), and it maps naturally onto long-short portfolio construction, which only cares about relative ordering anyway.
• Direction (up/down) — predict a binary sign. The simplest target, easiest to model with standard classification tools, but it throws away magnitude entirely: a model that correctly calls the direction of a 0.1% move and a 10% move gets equal credit, even though only one of those is worth trading.

Matching the target to the use case

The right choice depends on what the prediction feeds into. A long-short, dollar-neutral portfolio that ranks stocks and takes the top and bottom deciles cares only about relative ordering — a rank target is a natural fit, and it's more robust to market-wide noise than trying to predict raw returns directly. A single-name directional trade with a fixed position size might genuinely only need the sign right. A strategy that sizes positions proportionally to conviction, on the other hand, needs magnitude information that a pure rank or direction target throws away — it needs something closer to the raw (or risk-adjusted) return.

Worked example: three targets from the same data

Suppose next-month returns for five stocks are $+8\%, +1\%, -2\%, +3\%, -6\%$. As a raw return target, the model tries to predict these five numbers directly — hard, because a single large mover like $+8\%$ can dominate a squared-error loss. As a rank target, the same data becomes ranks 1 through 5 (best to worst): stock A (rank 1), D (rank 2), B (rank 3), C (rank 4), E (rank 5) — the model only needs to get the relative ordering right, which is a more forgiving problem, especially if the whole market moved together that month and every raw number was inflated by a common factor. As a direction target, it collapses further to just $+, +, -, +, -$ — the model must correctly call three "up" and two "down," discarding entirely the information that A's move was four times larger than D's.

Explore how classifier flexibility trades off against overfitting above — a direction target sets up exactly this kind of classification problem, while a rank or raw-return target instead asks for a continuous ordering or number, a different kind of question for the same underlying data.

What this means in practice

Rank targets are the default choice for most cross-sectional equity alpha research because they're naturally robust to market-wide moves and map directly onto long-short portfolio construction; raw or risk-adjusted return targets are preferred when magnitude genuinely drives position sizing; directional targets are mostly used for simple, single-instrument signals where a fixed-size trade is the plan.