get_model_vs_market

com.olympus-bets/olympus-bets-analytics · Verify this server

Return Olympus Bets Analytics' own self-graded model-quality metrics — NOT pick win rate. This is a different question than "did our picks win money?" (see get_performance_summary / get_track_record for that). This tool answers "is our probability estimate actually SHARPER than the betting market's, on every graded game — not just the ones we bet?" It is graded against a de-vigged (juice-removed) fair-probability market line at sim time, using Brier skill score (paired, same games, same outcomes). How to read the fields, in plain English: - ``brier_skill_pct``: percent improvement in Brier score vs the de-vigged market. POSITIVE = our model is sharper than the market. NEGATIVE = the market is sharper than us. Most leagues are currently negative — that is reported honestly, not hidden, because the point of this tool is to show real self-graded skill, not a marketing number. - ``model_weight_star`` (w*): the blend weight (0.0-1.0) our model earned in a model+market blend that minimizes log-loss. 0.0 means "defer entirely to the market's number"; 1.0 means "our number alone is already optimal." This is fit empirically per league/window, not asserted. - ``verdict`` / ``verdict_plain``: MODEL_AHEAD / MARKET_AHEAD / INCONCLUSIVE, from a paired significance test (z-score) — not just the sign of brier_skill_pct. - ``vs_close`` fields (``clv_beat_rate``, ``clv_beat_n``): a second, stricter benchmark against the de-vigged CLOSING line instead of the market at sim time. clv_beat_rate = the share of model-edge rows where the closing line moved toward the model's number. Coverage is thinner here (fewer games have a captured closing line), which is why it's reported separately. - ``n`` / ``reliable``: sample size behind each cell. Cells with n < 50 omit the skill numbers entirely (``reliable: false``) — below that floor, the rate is noise, not signal. Windows: ``30d`` (most current, smallest sample) and ``90d`` (steadier, larger sample). Use 90d as the primary read; use 30d to see if something is actively shifting. Freshness: the underlying file rebuilds daily (~12:50 UTC). If it is stale (>36h old), this tool returns ``{"status": "updating", ...}`` instead of presenting old numbers as current — never treat a missing ``windows`` key as "no skill data," check ``status`` first. Args: league: Optional league filter (e.g. "MLB", "NHL"). Omit for all leagues covered by the scoreboard (NBA, NHL, MLB, SOCCER, WNBA, TENNIS, LOL, CS2, GOLF, WC — CFB/NFL/CBB not yet in-season/covered). Returns: ``{status, generated_at, benchmark, close_benchmark, sample_floor_n, windows: {"30d": {...}, "90d": {...}}}`` where each window has ``overall`` (blended-across-leagues cell) and ``by_league`` (list of per-league cells, each carrying its own ``league`` code).