Classification vs Regression Targets for Trading

Two researchers can build a model on the same data, using the same features, and end up with completely different pipelines — one predicting a number, one predicting a label — because they never agreed on what the forecast was for. Getting this choice wrong does not show up as a crash. It shows up months later as a portfolio that trades the right direction at the wrong size, or a signal that looks great on one metric and useless on the one that pays the bills.

The problem, before any symbols

Say you want to forecast next week's return on a stock. There are two honest ways to state that as a machine-learning problem.

Regression: predict the number itself. The target is a continuous return, something like +1.8% or −0.4%. The model is graded on how close its number lands to the truth.

Classification: predict a bucket. The target is "up" or "down" — or "up more than 1%", "flat", "down more than 1%" if you want three buckets. The model is graded on how often it picks the right bucket.

These are not two views of the same forecast. They optimise for different things, they fail in different ways, and they push the rest of the research pipeline in different directions.

What each one actually buys you

A regression target keeps the magnitude information. If the true move was +5% and your model said +0.3%, that is a much bigger miss than the true move being +0.6% and your model saying +0.3% — and regression's loss function (typically squared or absolute error) reflects that. This matters a great deal if you size positions by the strength of the forecast: a model that only ever tells you "up" or "down" gives you no way to distinguish a high-conviction trade from a coin flip.

A classification target throws magnitude away on purpose. It asks a cruder question — right side or wrong side? — and in exchange it is far more robust to the outliers that plague financial returns. One earnings-day gap of +40% in your training set can drag a regression model's coefficients around badly, because squared error punishes big misses enormously; a classifier just sees one more "up" observation, no different in weight from a +0.5% day. If your trading rule genuinely only cares about direction — you take a fixed-size position either long or short — classification is asking the model the question you actually need answered, and not paying an accuracy cost to also get an answer to a question you don't need (the exact magnitude).

	Regression target	Classification target
What's predicted	A continuous return	A direction or bucket
Sensitive to outliers	Yes, often badly	Much less
Preserves conviction/sizing information	Yes	Only if you keep predicted probabilities
Natural loss function	Squared or absolute error	Log-loss, cross-entropy
Matches a fixed-size directional bet	Poorly — over-engineered for the use case	Well
Matches a signal that scales position by strength	Well	Poorly, unless probabilities are used as a proxy

The trap in the middle

The compromise that looks free — turn a classifier's predicted probability into a pseudo-magnitude, and use it for sizing — is not free. A probability of 0.62 that the stock goes up is not the same object as an expected return of +0.4%, even though people use them interchangeably. Probabilities are bounded, roughly comparable across stocks, and calibrated (if at all) to hit rate, not to the dollar payoff of being right. Treating "0.62" as if it were a return forecast smuggles a regression assumption back into a classification model that was never trained to make it.

The safest way to decide is to write out the trading rule first — in words, before any model exists — and ask what number it needs. If the rule is "go long or flat, same size every time," you need a label. If the rule is "scale the position with the strength of the signal," you need a number, and a classifier's probability is at best a stand-in for one, not a substitute.