Public work on the S&P 500 ETF under a shared standard: causal features, walk-forward evaluation, baselines, and a willingness to stop when the edge is not there. The live production line is short-horizon realized volatility. Experiments that fail those standards are documented as null results and stay out of production, including future work that does not clear the bar.
End goal: a unified SPY market-state system built from modular models that each answer a different question (how large is next-horizon vol? is the session range expanding? which regime? which way?), sharing data and ops discipline, shown together with honest track records. Not one black-box “all of SPY” model.
What “suite” means: one data spine (SPY OHLCV, later bars; optional VIX), one ops spine (registry, candidate vs active, promotion gates, health, ledgers), one product surface (panels for vol, range, regime, direction, track). Shared culture: walk-forward only, label-native baselines, freeze vs promote separated, clean fails published.
What it does not mean: the same head predicting vol and direction; reopening the live vol freeze to chase a sibling problem; or treating research pass as live promote.
Only the vol line is production-frozen and live. Everything else is research, planned, or parked until it clears its own bars.
Walk-forward forecasts of next-five-trading-day annualized realized volatility for SPY, compared to naive persistence and GARCH(1,1). Architecture is frozen (monthly LightGBM, nine baseline features). Monthly weight updates, when run, are gated; the modeling ladder is not reopened for routine ops. This is the system that remains active.
GraphForge is a standalone product (any workflow): encode product law as typed state graphs and verify pipelines. SPYForge is the folder inside the SPY research repo that applies it here (freeze, candidate-only retrain, promote gates, agent ACL). Simulate board verifies that application. Not forecast skill. Production stays on vol pipeline scripts until an explicit cutover.
Walk-forward classification of whether the next SPY session range exceeds a trailing ten-session mean. Research architecture freeze of LightGBM trial 22 (locked OOS n=1,905: Brier 0.21378 vs climate 0.2454; Spearman 0.164). Paper hold-out after 2026-07-31 is 32 scored origins (Brier 0.2448 vs climate 0.2148); not a live badge. Promotion gate cannot write live. Vol freeze is unchanged.
Walk-forward three-way classification of whether the next SPY session range lands in the
low, mid, or high tercile of the trailing 252 sessions. Research architecture freeze of
multinomial logistic 00_logistic_regime_21d (locked OOS n=1,938: Brier 0.5677
vs climate 0.6658; accuracy 0.532 vs lagged 0.524). Not a live badge. Vol freeze is unchanged.
We tested a strict causal daily-bar model that predicts whether SPY will close higher than it opened, using only information available at the open. Out-of-sample (2019-2026) neither logistic nor tree models beat the always-up baseline on accuracy or Brier score. A weak ranking of open-to-close returns exists, but it does not survive realistic transaction costs.
Negative results are part of the research process. The same evaluation standards apply across experiments: causal features, baselines, walk-forward evaluation, and an explicit stop when the bar is not cleared.
Martial Systems LLC. For investor or partnership inquiries, email martialsys@gmail.com.
