How we build win probability models

Model estimates and market-derived references

A market-derived fair-value reference can be built by devigging sportsbook lines — averaging Pinnacle, FanDuel, and DraftKings odds. This gives you a consensus probability that tracks the market by construction. It can be useful for line shopping, but it is not an independent model view — because the output already agrees with the market.

The released game-state architectures described below use features such as score differential, clock, period, ratings and sport-specific inputs; some include a pregame probability prior. The API contract labels the exposed edge model_pre_blend: it is taken before the bot's market-blend step. That label does not establish that every upstream feature in every current model version is independent of market information. Inspect the named model, feature set and prediction basis in each dated evaluation before making an independence claim.

Model quality must be compared on matched games, prediction times and model versions. The dated public benchmarks show selected evaluations; their scores do not establish current all-sport performance. A separately evaluated model can provide another reference for comparison with market prices; its input independence and trading value must be assessed for that model and cohort. See /clv-evidence for the audit and retraction of our former outcome-conditioned CLV claim, /validation for the exported backtest snapshot, and /results for filtered public-ledger outcomes and evidence coverage.

Data pipeline

We scrape ESPN's play-by-play API across the core sports we model. Recorded states contain score, period, clock and other source fields. Row counts depend on source availability and the collection run; missing intervals remain missing.

Sports: NBA, NCAAMB, NCAAWB, CFB, NFL, NHL, MLB

Data is stored as Apache Parquet files. One row = one game state (score, period, clock, ESPN WP, outcome label).

Module 1 of our course teaches historical collection for NBA, NCAAMB, NHL, NFL and CFB; production collectors and supported API sports can differ.

Feature engineering

The documented feature families below vary by sport and model version. We deliberately keep the feature set small — overfitting to noise destroys trading value.

Model architecture

We use sport-specific models — no one-size-fits-all approach. The sections below distinguish the current live platform from adjacent research and course material.

Public API coverage and documented model design

The models cover NBA, WNBA, NCAAMB, NCAAWB, CFB, NFL, NHL, MLB with sport-specific pipelines. Calibration and overlays depend on the model and serving path. Raw API model estimates and the bot's calibrated, market-blended trading values can differ; the API/bot disclosure describes that distinction.

Basketball & Football (NBA, NCAAMB, NCAAWB, CFB, NFL)

Split-Phase XGBoost with 16 features including team offensive/defensive ratings (ORtg, DRtg), pace, momentum (scoring runs over last 2 and 5 minutes), Elo ratings, and interaction terms (score_diff × time_fraction). Evaluate each released model with its own training dates, chronological splits and held-out cohort; static architecture descriptions are not a current quality report.

Injury overlay: The documented NBA overlay uses ESPN injury status with a 10-minute cache and configured player impact weights. These weights are modeling heuristics, not measured causal injury effects or proof that the upstream status is fresh. When a player is OUT, the model subtracts their impact; QUESTIONABLE halves the adjustment. Total cap: ±15%.

Model selection uses a hybrid criterion: near-best Brier score AND best trading value (c/trade) on a held-out backtest window. This is a selection criterion; it does not guarantee better live P&L or prevent overfitting to the backtest.

Hockey (NHL)

XGBoost with 17 features including all basketball features plus hockey-specific metrics: power play %, penalty kill %, save %, faceoff win %, and penalty minutes differential. Inspect the corresponding model report for its dated test sample and baseline; these features do not establish a current performance level.

Injury overlay: The documented ESPN skater-status overlay excludes goalies, which use a separate adjustment. Player weights are configured heuristics; status availability and freshness depend on the source.

Goalie & shot overlay: Starting goalie detection adjusts for goalie quality. Expected goals (xG), Corsi, and power play state provide real-time shot quality signals. Combined cap: ±12%.

Baseball (MLB)

Ensemble model with starting pitcher ERA, WHIP, K/9 as base features. A model report must state the test dates and sample size for any Brier or ECE result. Three post-model overlays stack on top of the base prediction:

• Run expectancy: Tango run-expectancy matrix (outs × runners × batting side) provides granular in-game state beyond the score.
• Park factors: 30-team FanGraphs 3-year rolling park factors (Coors Field = 1.25, highest).
• Bullpen quality: 30-team ERA, time-scaled from 0% (early innings) to 100% (late innings) as starters hand off to relievers. Capped at ±5%.

Soccer (Bundesliga, EPL, La Liga, Serie A, Ligue 1)

The published April 2026 soccer incident describes a live-trading freeze (force-shadowing). The XGBoost model trained on 2,949 matches had silently vanished from the server in April 2026, degrading the live bot to a Poisson fallback that measures worse than its own entry prices. A retrained challenger failed its promotion gates, and the response included artifact-presence monitoring. Current status belongs in the repair scorecard and changelog. Pregame research continues on the Poisson + XGBoost stack.

Esports (CS2, LoL)

Both esports run models on the platform; live order flow varies with each sport's current risk status. The published August 2026 record describes a CS2 sizing pause and a LoL revert to v1 after v2 lost two head-to-head evaluations (see below). Check the repair scorecard and changelog for subsequent status.

CS2: 15 features including round differential, Elo, map-specific win rates (per team per map), recent form, head-to-head record, and Elo momentum. Economy data from bo3.gg and HLTV scorebot provides equipment value and consecutive-loss tracking as post-model overlays. The documented CS2 entry configuration used underdogs (20–42c) and a 20c+ minimum modeled gap after a cohort audit. Those are configuration choices, not proof that future underdog entries will outperform.

LoL: v1 analytical model — Elo ratings plus a binomial series model for BO1/BO3/BO5. The v2 XGBoost override (20 features: gold differential, kills, towers, dragons, barons, inhibitor advantage, and more) was reverted in August 2026 after two independent evaluations on 110 fire-anchored matches both favored v1 (Brier 0.1692 vs 0.2077). That dated comparison supports the reported revert; it does not establish the status or performance of later versions.

Tennis (ATP, WTA)

Hierarchical analytical model: player-specific serve win rates (by surface) → point probability → game probability → set probability → match probability. Serve rates are computed from 13,174 ATP matches (2020–2024) using the Sackmann dataset, giving us player-specific rates for hard, clay, and grass surfaces instead of tour averages. The documented Elo ratings are surface-aware; player coverage depends on the rating-file version. The documented activation threshold is 50 resolved trades; reaching it does not establish adequate out-of-sample calibration.

Calibration, measurement and risk controls

The documented bot pipeline can apply calibration and overlays before its measurement and risk checks. Applicability varies by sport, venue and serving path; the raw API model output is not identical to this chain:

1. Post-hoc isotonic calibration: Fitted on held-out calibration season. Evaluate the fitted map on an unseen cohort with stated bins and sample counts. A low fitting-set error does not guarantee held-out calibration, and a model estimate of 70% is not a promise of a 70% win rate in a new sample.
2. Live rolling recalibration: Automatically refits an isotonic correction every 25 resolved trades per sport, using a 500-sample rolling buffer persisted to disk. This is a documented refresh mechanism, not proof that every drift is detected or corrected; each resulting map still needs an independent evaluation.
3. Closing-line value (CLV) audit: Admitted live fills are measured against our captured closing price where one exists; the scorecard publishes the current measured count and coverage rather than freezing an old percentage into this methodology page. The winner/loser split we used to cite here was retracted 2026-07-21 as a measurement artifact; the honest unconditioned aggregate (share of fills beating the close and mean CLV per trade) is published per sport at /clv. The four-failure-mode taxonomy (bad model / bad blend / bad timing / bad CLV label) is published at /clv/repair alongside the explicit reactivation criteria for any sport currently paused.
4. Dynamic position sizing: Rolling 50-trade performance per sport determines a size multiplier (0.25x–1.5x). Hot sports (>55% WR + positive P&L) get sized up; cold sports (<40% WR) get sized down. Recomputed every 10 minutes.
5. Meta-model edge filter (research, shadow mode only): A Gradient Boosting classifier trained on resolved trades was proposed to separate real edges from noise. It has never been trusted to block live trades — it runs in shadow mode as a research instrument. The "filters 28% of losing trades" figure we once cited was not reproducible and has been removed.

Spread/Total

Spread and total models require a different setup from moneyline markets: regression on remaining margin or total, then a distributional layer to convert that forecast into cover/over probabilities.

Backtesting methodology

Our backtests are designed to avoid the common mistakes that inflate results.

• Time-split validation: Train on season N−1, test on season N. Feature timestamps and label availability must also be checked; the split alone does not prove there is no leakage.
• Real market prices where available: Where direct market snapshot history exists, backtests use recorded market prices. Where coverage is incomplete, we call out any proxy assumptions explicitly on the validation page.
• Deduplication: One entry per (game_id, side, score, period). No duplicate signals from the same game state.
• Subsampling where needed: Historical replay backtests can downsample state rows to better match how the live system would realistically observe and act on the game.
• Hold to settlement: The hold-to-settlement backtests assume settlement at game end and omit exit timing or stop-loss rules. This is a study assumption, not a promise about all live positions.
• Net of fees: Results are reported net of the fee and slippage assumptions used by the active backtest export, with the exact assumptions disclosed on the validation page.
• Sport-specific entry filters: Filters like score differential, game state, or pregame-consensus agreement are applied only where validation shows they improve robustness.
• Execution windows: Entry timing constraints vary by sport and market. We only keep them when out-of-sample testing shows they improve execution quality.

What we tried that doesn't work

We believe in showing failures alongside successes.

• Sell-the-top / fade spikes: Negative EV. The tested configurations did not show reliable reversal after the modeled costs.
• Compression sniping: Negative on historical data, adverse selection live.
• Martingale / double-down on dips: Catastrophic — doubles position into adverse moves.
• Early NBA taker configs: Earlier versions underperformed until pacing/context features and stricter entry filters were added. Old model versions are not treated as current proof.
• Spread/Total mean-reversion exits: The tested exit rules underperformed; hold-to-settlement was the comparison used in those tests.
• ESPN live WP as model input: In the documented game-state experiments this input was rejected because it can make the evaluation circular. Some architectures use a pregame probability prior; this historical finding does not certify feature independence for every current release.
• Conformal prediction intervals: Tested and rejected — added complexity without improving trading value.