SIG forecasts whether each stock symbol in the daily universe will go up or down on the next trading day.

Features

A feature is a derived number computed from a symbol's price and volume history: things like recent momentum, intraday volatility, where the close sits inside the day's range, or how the morning's volume compares to the rest of the day. Tetsuo defines about fifty of them. They're what a forecasting model actually learns from, rather than the raw price bars.

Per-Symbol Feature Selection

Each stock behaves differently, so the same fifty features are not equally useful for every symbol. SIG runs a configurator per symbol that walks the feature catalogue one at a time and keeps the ones that improve directional accuracy on held-out data. The winning subset gets saved as that symbol's configuration and is what SIG actually trains and forecasts on going forward.

Direction vs Magnitude

SIG only answers the easier question: which way is the stock going from T[1]@1100:T[2]@1100 (the time range between when we buy at 1100 ET the next trading day, and the time when we sell at 1100 ET the trading day after that) -- based on data available at T[0]@1700 (the closing time of the current closed trading day).

The harder question of how much it will move is left to MAG, the magnitude forecaster downstream.

Splitting these tasks allows specialization for each aspect we care about: direction and magnitude. This lets each side optimise for its own objective. SIG can be confident about direction without committing to a size, and MAG only spends compute on symbols SIG already thinks will go up, focused purely on how much it will change.

Model

SIG uses a gradient-boosted tree classifier trained per symbol on TDM's daily-wide derivative dataset. The trained model outputs a direction (UP or DOWN) plus a confidence score and a classifier probability that downstream stages use to decide whether to act on the call.