SILO ALG

Picks which proprietary forecasting pipeline MAG runs. This is a MAG-wide setting because the choice is ingrained into every forecast AND every configurator walk — the configurator must measure accuracy under the same regime the forecaster will run, otherwise the saved accuracy does not represent what is actually produced. Changes take effect on the next forecast or configurate kickoff; in-flight jobs continue under the regime they started with.

SILO ALG 1 is the two-point compose method. The model is trained to predict the raw closing price two trading days ahead. At inference the runner makes two point-predictions — one for T+1, one for T+2 — and the reported magnitude is the percent difference between them. This is the preserved proprietary baseline. Susceptible to compounding error because the reported magnitude is a difference of two model outputs.

SILO ALG 2 is the direct-magnitude method. Each row of the history is massaged so its training target IS the percent change from that row's T+1 close to its T+2 close, and the model emits a single point-prediction at inference that IS the reported magnitude. No difference of two predictions, no compounding error. Higher accuracy in practice.

Training History

Maximum number of most-recent trading-day rows MAG keeps from TDM's daily-wide derivative CSV before training and forecasting. TDM may store far more history than is optimal for a per-symbol LightGBM fit; this cap trims the fetched dataframe to its last N rows before the Forecaster sees it. Applied symmetrically by the forecaster and the configurator so the configurator's walk measures accuracy under the same window the forecaster trains on. Shorten for faster recency-biased fits; lengthen if the data warrants more context.

Minimum SIG Confidence

Minimum SIG confidence (the aggregate walking score that gauges feature-set coherence at the time SIG saved each symbol's configuration) a row in SIG's latest full report must meet for MAG's bulk Forecast All Symbols routine to include the symbol in its work set. 0 = no filter (every UP-direction row passes). 100 = pass only perfect-confidence rows. The bulk configurator does NOT apply this filter — it walks the same inventory SIG's bulk configurator walks, so MAG and SIG stay configurated against the same symbol set.

Minimum SIG Probability

Minimum SIG classifier probability of the printed direction (new in this generation versus the prior MAG architecture, which only filtered on configurator accuracy) a row in SIG's latest full report must meet for MAG's bulk Forecast All Symbols routine to include the symbol. 0 = no filter; 100 = pass only perfect-probability rows. Applied alongside the minimum confidence above — both thresholds must be cleared for a symbol to enter the forecast work set. Same scope as the confidence minimum: applies to the bulk forecast, not the bulk configurate.