How RoliBot NBA works

Data & seasons

Regular-season team games are pulled season-by-season via LeagueGameFinder. The code automatically includes every season from 2018-19 through the current NBA season string (e.g. 2025-26), so the merged training frame always grows with the calendar — not a fixed "2018–2025" cap.

Richer runs may embed a per-season audit under pipeline.season_pull (older JSON). The v5 CLI JSON is slimmer: slate date, injury snapshot, accuracy guardrails, and parlay blocks are at the root — the dashboard reads both shapes.

Model training

• Features: rolling offensive/defensive proxies, rest and back-to-back flags, win-streak and momentum differentials, home/away splits merged into one row per game.
• Split: strictly time-ordered 80% train / 20% holdout (no random shuffle leakage).
• Ensemble (v5): four calibrated models — XGBoost, Random Forest, Gradient Boosting, Logistic Regression — blended at 45% / 30% / 15% / 10%.
• Features (v5): ~65 rolling team signals including multi-window momentum (3 / 5 / 10 games) and variance-style cues so unstable teams are distinguishable from steady ones.
• Cache: fingerprinted training data; cache hit skips full retrain (see JSON model.cache_hit).
• Evaluation: when the JSON includes them, holdout accuracy, log-loss, and calibration bins appear on the dashboard; v5 runs may omit the full evaluation block.
• LLM layer (optional): with ANTHROPIC_API_KEY, short per-game narratives can augment the rules-based output; without a key the pipeline still runs.

Props, parlays, injuries

Props use PlayerGameLog for the current NBA season (with a previous-season merge early in the year when sample size is thin). Hit rates blend recent games heavier than older ones; opponent defensive context nudges scoring lines. Injuries: ESPN-style feed with tighter name matching; Out, Doubtful, GTD, Questionable, and similar statuses are excluded from props. Traded players: recent game logs must align with the roster team or the player is skipped. Activity: very stale game activity is filtered so long-term IR doesn't slip through. Same-game parlays (SGP) are listed separately; hit probabilities still multiply legs (books price correlation).

Live calibration (graded picks)

After the morning grading job fills graded_picks, this chart compares each model-confidence bucket to realized hit rates. Points hugging the diagonal mean stated probabilities matched outcomes — a sanity check before trusting Kelly-style sizing.

How to use the dashboard

1. Set your bankroll assumptions in the predictor config to match what you actually risk.
2. Treat STRONG / GOOD sides plus printed Kelly amounts as templates, not guarantees.
3. Safe prop parlays prioritize hit rate; risky ladders are for tiny stakes only.
4. Hot-streak legs emphasize recent form vs season baseline — still statistical, not narrative.
5. Re-run or deploy the API after the date rolls so the slate matches "tonight."
6. Invalidate or delete the model cache when you want a forced full retrain on fresh history.
7. Never bet more than you can afford to lose — this stack is research and entertainment.