Hermes NBA 1.1: a dedicated probability engine for basketball matchups
Hermes NBA 1.1 is the first Hermes-family model behind WhoWins.ai matchup percentages. It was designed around a simple product goal: show a clear pregame probability while keeping the explanation layer separate and transparent.
What Hermes NBA 1.1 does
Hermes NBA 1.1 produces the prediction percentages used on NBA matchups. It is built for pregame winner prediction, not live betting, player props, totals, or spread-specific advice. On WhoWins.ai, Hermes supplies the probability layer while cortex-rubric can still supply the plain-language matchup analysis.
Why market signal matters
The strongest research result was practical: clean market features outperformed the richer basketball-only feature set available in the current dataset. Sportsbook markets compress a large amount of public information, including team quality, injuries, rest, expected scoring environment, and bookmaker adjustments. Hermes NBA 1.1 treats that market signal as a serious input rather than pretending it does not exist.
How the model was evaluated
The research pipeline used leakage-safe pregame features from historical NBA games, then enriched the odds-era rows with historical sportsbook snapshots. Candidate models included logistic regression, ExtraTrees, HistGradientBoosting, LightGBM, paired team-history LSTMs, and lean ensembles. The best observed result was a lean market ensemble with 0.6927 test AUC, 0.6601 test accuracy, and 0.6850 test precision.
What did not win yet
Sequence models were useful as research branches, but they did not beat the simpler market-focused models with the current data. The best ensemble included an LSTM candidate, but the optimizer assigned it zero weight. That is a useful constraint: Hermes should stay understandable and production-friendly until more odds-era and player-state data justifies more complexity.
What users should take away
Hermes percentages should be read as model probabilities, not guarantees. A 64% prediction means the model sees one side as meaningfully stronger than a coin flip under the available pregame information. It does not remove uncertainty, and it should be compared against the market, the AI summary, and your own view of the matchup.
Keep reading