Daily retraining
GitHub workflow & logs ↗Configured for every day at 03:17 UTC, after the US market close. On non-trading days, prices retain their latest trading-session date. Each run refreshes daily prices and dated filings, retrains the supported risk models, repeats chronological validation and scores earlier issued forecasts whose outcomes have matured.
GitHub schedules can be delayed. A green completed run means the process and verification succeeded; it does not mean the model has a proven edge. A failed attempt retains the last successful forecasts with their original dates.
Models, their jobs and their limits
The daily pipeline maintains company and ETF volatility models. It does not repeat the entire architecture search or retrain archived return/foundation-model experiments.
How each model works
Price Ridge combines past volatility, returns, drawdowns and trading activity with regularization that limits large fitted weights. Financial-history Ridge also sees dated accounting ratios, their earlier changes and filing reactions. Gamma learns positive variance relative to recent volatility. TabM combines eight small neural predictions with shared parameters. Long-memory HAR summarizes volatility across short and long past windows.
Inputs are screened and transformed using training data only. The prior calendar year chooses among the financial candidates; it cannot choose using the evaluation year. Neural runs use 250 updates, a fixed compute budget rather than a claim of full convergence.
Company forecast explorer
Browse model coverage and latest next-month forecasts
Company risk breadth alongside Hydra stress
Does it improve forecasting?
Every result names its comparator. Green means lower historical forecasting loss; adjusted significance and stability are separate checks. Positive loss improvement means smaller forecast errors relative to the named control; negative means worse. An adjusted p below 0.05 passes the significance threshold, with stability checks still required. Company and fund families each adjust eight comparisons. Reusing already researched history does not create an untouched holdout.