sports-predictor-api.fly.dev

How many external API calls we've used today/this month.

One-stop proof + health surface: model registry, live calibration, recent backtests.

Background scheduler state + last run summary.

Manually trigger the daily model-refresh job (re-ingest intl + refit DC).

Every upcoming fixture with calibrated probabilities and derived markets. The one endpoint priced for agents (x402, USDC on Base). It is pure model output — no LLM — so it carries no per-call model cost, only compute. Free when X402_PAY_TO is unset, so an unconfigured deploy degrades to free rather than to a broken paywall.

Plain-English question -> the model's grounded answer. Fixture resolution and the prediction itself are free (heuristics + math); only the short write-up calls the LLM, so this sits behind the same rate limits and daily budget as the other paid endpoints.

Get all upcoming matches with odds across all major sports for next 7 days.

Find the safest favourites across all sports for the next 7 days. Ranks by implied probability (1/odds). No AI calls — pure odds-driven shortlist so it's fast and free. Use `/api/analyze` on any pick for deep analysis.

Top teams by Elo rating for the given sport.

Quick Elo-only 1X2 prediction (no LLM, no API calls).

Update per-sport ensemble weights. Pass weights_json as a JSON string, e.g. '{"elo": 0.5, "poisson": 0.4, "llm": 0.1}'. Components don't need to sum to 1.0 — blend() renormalises.

Fit Dixon-Coles on the requested league's full season and save the model. Run this once per season per league. The /analyze endpoint will then pick up the saved model. ~5-10 seconds per league.

Download football-data.co.uk CSVs (results + closing odds) into SQLite. leagues : comma-list (default: all big-5). seasons: comma-list like '2425,2324,2223'. Free, no quota impact. Run once, then use /historical/backtest.

Rigorous backtest on historical data WITH ROI (uses real closing odds). Compares naive/elo/poisson/market/ensembles, reporting Brier, log-loss, flat ROI, and value-bet ROI (only betting where we beat the closing line).

Replay historical matches through chosen pipelines. Query params: leagues : comma-list (e.g. 'premier_league,la_liga') pipelines : comma-list of {naive, elo, llm}. 'llm' costs Anthropic credits. matches_per_league : total fixtures pulled per league (default 200 ≈ full season) score_fraction : last X% scored; rest seed Elo (default 0.3 → 70% warmup) score_cap : optional max scored matches per league (use to limit LLM cost)

Seed Elo ratings for the 48 World Cup national teams. Run once before the tournament. Won't clobber ratings learned from real matches unless overwrite=true.

Predict an international match using national-team Elo. neutral=true (default) strips home advantage (most WC games are neutral-venue). Blends with the LLM if you also want narrative context — see /analyze.

Pull real international results (WC2022, Euro2024, Nations League, Copa, friendlies) from API-Football and update national-team Elo chronologically. This replaces hand-guessed seeds with data-driven ratings.

Full correct rebuild of international Elo: seeds → 2022-24 competitive (K=1.0) → pre-WC friendlies at ⅓ weight → logged WC results at 2.5x weight.

Fetch a full CL season's results from ESPN (Tue/Wed/Thu + final dates).

Walk-forward backtest on the harvested season: seed cross-league Elo from domestic ratings, predict every match with only prior info, score.

CL fixture prediction on the walk-forward-corrected European ratings.

Download latest football-data season CSVs for PL + La Liga and rebuild each league's Elo from all stored seasons (chronological replay).

CL-style walk-forward backtest of a domestic league season, with a Pinnacle-closing-odds market benchmark on the same matches.

Draw-aware v2 prediction for a PL / La Liga fixture. league: premier_league | la_liga. Home advantage applied.

In-season: ingest finished PL/La Liga results (ESPN, free), score the model, update league Elo. Run daily once the 2026-27 season starts.

Write predictions for upcoming fixtures, before they are played. Normally the nightly scheduler does this. Exposed for the first run of a season and for backfilling after downtime.

Record a pick for an upcoming fixture. Free, no stake, no payout. Refused at or after kickoff, and refused if you already picked this fixture: a leaderboard where standing picks can be revised proves nothing.

The exact text to sign for a pick. Served rather than built client-side so the wallet signs the same bytes the server will verify. A mismatch here is the usual way signed-payload schemes silently break.

One player's picks, scored and pending.

Standings by average Brier, against the model on the same fixtures.

Score picks whose fixtures now have results. The scheduler does this nightly; exposed for backfills.

v2 draw-aware prediction: Elo->Poisson goals bridge + calibration layer fitted on our own logged WC predictions. CAN pick the draw (v1 cannot).

(Re)fit the probability calibration layer on wc_match_log.

The daily ritual: ingest new WC results (ESPN, free + no quota), score the model's pre-match predictions (live calibration), update Elo at WC weight, and preview today/tomorrow fixtures with fresh probabilities.

Current national-team Elo ladder.

Monte Carlo the whole tournament → advance / reach-final / win-the-cup odds. groups_json: optional JSON of {"A": ["Team1", ...4], ...}. If omitted, the top-48 rated teams are auto snake-seeded into 12 groups. Provide the REAL draw once it's known for accurate outright probabilities.

The prediction log, with each row's provenance stated rather than assumed. Two kinds of row live here and the difference matters: prelogged=1 the probabilities were written BEFORE kickoff and the result was filled in afterwards. A genuine forward prediction; nothing about it can be revised in hindsight. prelogged=0 the probabilities were generated when the result was ingested. Still walk-forward -- the model had not seen that match, and Elo updates only after the row is scored -- but it is a backtest of that match, not a forward call. Every World Cup row is prelogged=0, because pre-kickoff logging did not exist while that tournament was running. Saying so is the point: a record that overstates how it was produced is worth less than a smaller honest one. Club rows from the 2026-27 season onwards are pre-logged. Public, read-only.

Reliability curve + Brier decomposition for one league-season. This is the evidence behind the calibration claim: does a stated 60% actually happen ~60% of the time, and how does that compare to the closing line on the same matches. Accuracy is deliberately not the headline — no 1X2 model picks draws, so ~27% of matches are unwinnable on that metric. Serves the report stored at fit time — a pure read. The underlying backtest replays a full season AND rebuilds elo_ratings, so it must never run on a public request path (CPU exhaustion + concurrent rebuilds racing the same ratings table).

(Re)fit the per-sport calibration alphas. Club alphas come from the walk-forward backtest (no live club log yet); the international alpha comes from logged World Cup predictions.

How well-calibrated are the AI's confidence scores? Joins logged predictions to settled bets (by match name) and bins by confidence. After ~20+ settled bets the buckets become meaningful.