Label Construction for Supervised Trading Models

Fixed horizon returns, triple barrier labels, and meta-labeling each teach a model a different task, and the wrong choice undermines the rest.

Label construction for supervised trading models is the step that decides what the model is actually trying to predict, and the choice matters as much as the choice of algorithm, because a fixed horizon return label, a triple barrier label, and a meta-labeling setup each teach the model a different task even when trained on identical features. Getting this step wrong produces a model that optimizes a target nobody actually trades against.

The simplest label is the return over a fixed number of days or bars following the observation, for example the return over the next five trading days. It is easy to compute and easy to explain, and it is also the label most likely to teach the model something a real strategy cannot use, because it ignores how the position would actually be managed, including any stop loss or profit target that would have closed the trade earlier.

A fixed horizon label also treats a move that happens on day one the same as a move that happens on day five, even though a strategy with a stop loss would have exited on day one and never experienced the day five outcome at all. This mismatch between the label and the trading rule it is meant to support is a common, quiet source of a model that backtests well on the label but poorly once wrapped in an actual execution rule.

The triple barrier method, described in detail in quantitative finance research on this exact labeling problem, sets three exit conditions at the moment a position would be opened: an upper barrier representing a profit target, a lower barrier representing a stop loss, and a vertical barrier representing a maximum holding period. Whichever barrier the price path touches first determines the label, a win if the upper barrier is hit first, a loss if the lower barrier is hit first, and a smaller, time based outcome if the vertical barrier is reached without touching either price barrier.

This label matches the actual trading rule far more closely than a fixed horizon return does, because it accounts for the fact that a real position would have been closed the moment either barrier was touched. The tradeoff is more design choices: how wide to set the barriers, typically scaled to a measure of recent volatility so they adjust to current conditions rather than staying fixed in absolute terms, and how long to allow before the vertical barrier forces an exit.