MAG forecasts how much each stock symbol will move on the next trading day for the names SIG already calls upward.

What's Being Predicted

The output is a percent magnitude. SIG already answered which way the stock is going, so MAG fills in the size of the expected move. Together the two forecasts give the rest of the pipeline a directional call plus a number to rank candidates against each other.

The Candidate Set

MAG only runs against symbols SIG has called UP and whose SIG forecast confidence and probability both clear configurable minimums. The intent is to keep MAG from spending compute on calls that are not actionable with confidence anyway.

Per-Symbol Feature Selection

Like SIG, MAG runs a per-symbol configurator that walks the feature catalogue and keeps the features that improve a magnitude-accuracy score on held-out data. The winning subset gets saved as that symbol's MAG configuration and is what MAG actually trains on for that symbol's future forecasts.

Model

MAG uses a gradient-boosted regressor trained per symbol on TDM's daily-wide derivative dataset. The operator picks between two algorithm variants on the Configuration page: the original method that predicts price levels and reports the difference between two predictions as the magnitude, or a newer method that trains the model directly on percent change so the single point-prediction IS the magnitude.