Model

Ensemble Meta-Learner

The second-layer classifier. Reads every teammate's buy-probability as a feature and produces the final daily yes-or-no that shows up on the homepage.

Accuracy51%

AlgorithmLogisticRegression · C=1 · l2 · 6 stacked features

DataPer-day buy-probabilities from every teammate base model: christa_proba, ryan_proba, cameron_proba, hrijoo_proba, marcus_proba, ai_proba

What the model pays attention to

marcus_proba61%

christa_proba17%

hrijoo_proba9%

ai_proba7%

cameron_proba3%

ryan_proba3%

Approach

Pulled aligned historical data from every upstream source (EIA WTI daily, Alpha Vantage WTI, NOAA Gulf SST, EIA weekly US gasoline stocks, Yahoo CL=F, Yahoo RB=F), replayed each teammate's base model over the 2022-2026 overlap window, and stacked their per-day buy-probabilities into a training set. Label was forward-looking: did WTI finish higher 7 trading days later? Held out the most recent 20% as a time-ordered test split. Trained logistic regression on purpose — the fitted coefficients stay readable as 'how much trust this teammate earned,' which a deeper model would have buried at this data size. Six stacked features, one per teammate base model. Test accuracy 50.8% on 187 held-out days, beating always-buy (49.7%) and majority-vote (47.6%). Marcus's weight came out biggest by a wide margin; Ryan's is slightly negative, meaning his signal is most useful inverted. Full training is in scripts/retrain_ensemble.py; ensemble/build_prediction.py loads the .joblib every morning.

Features

christa_probaderived

BUY probability from Christa Irakoze's WTI gas-timing model (logistic regression over EIA daily crude oil spot prices).

ryan_probaderived

BUY probability from Ryan Wolff's WTI crude oil signal (gradient boosting over Alpha Vantage WTI history).

cameron_probaderived

BUY probability from Cameron Anthony's Gulf SST storm-risk model (gradient boosting over NOAA OISST sea surface temperatures).

hrijoo_probaderived

BUY probability from Hrijoo Mullick's gasoline inventory signal (gradient boosting over EIA weekly US stocks).

marcus_probaderived

BUY probability from Marcus Holder's WTI futures signal (gradient boosting over Yahoo CL=F daily closes, 15 engineered features).

ai_probaderived

BUY probability from the AI-Created RBOB futures baseline (logistic regression over Yahoo RB=F daily closes — an LLM-chosen model included for head-to-head comparison against the human-chosen signals).

Built by

John WheelerLinkedIn

Team lead on the project — built the stacked ensemble that combines every teammate's classifier into the site's final YES/NO, plus the daily pipeline, the site, and the R2 data layer.

Trained on

2022-04-06 to 2025-07-07 (745 rows). Test window 2025-07-08 to 2026-04-07 (187 rows).

All models