Supervised signal classification

Instead of a fixed indicator rule, a supervised model takes a set of features, such as momentum, volatility, and valuation, and classifies whether the next period's return is attractive. Trees and neural nets are used because they can capture interactions that a linear rule misses. The published comparisons still find familiar features, especially momentum, liquidity, and volatility, doing much of the work.

Not investment advice. Nothing here is a recommendation to buy or sell anything. Backtests and historical studies are not guarantees, and a rule that worked in one sample can fail in the next. Trading can lose money. Links are for education, not endorsements. Do your own research.

Rules

Define the label first, for example the sign of next month's excess return, using only information known before that month. Train on a rolling window, purge the boundary so the label does not overlap the features, and test on the following window. Constrain the model, or the tree will memorize the sample. Size positions from the predicted rank, with a hard risk limit the model cannot override.

Where AI fits

This is the strategy. The extra AI layer is an LLM that writes the feature pipeline and the evaluation report, including a comparison with a simple momentum baseline. If the model cannot beat that baseline out of sample, after costs, it does not belong in production. AutoML that maximizes in-sample accuracy will happily do the wrong thing.

Markets

A broad stock universe, or liquid futures, with point-in-time data. Survivorship-biased price histories make any classifier look smarter than it is.

Timeframes

Monthly or weekly predictions. Daily classification is possible and much easier to overfit.

Risks

Leakage, from using the whole sample to pick features or from a label that overlaps the training row, is the usual failure. Flexible models fit noise. A high classification accuracy can still lose money after costs and after a few bad months in a row.

← Back to all strategies