How we predict

Versi pulls structured match data from Sofascore — fixtures, lineups, formations, suspensions, recent form. We layer on weather forecasts from OpenWeather (temperature, wind, precipitation at kickoff), referee tendencies from our own historical dataset, and live closing odds from major bookmakers. Each input is timestamped, version-controlled, and passed to the model with explicit confidence weights — so we know which signals drove every prediction.

Each prediction runs through a multi-stage pipeline. Stage 1: feature extraction normalizes raw match data into structured signals (form deltas, xG ratios, head-to-head priors). Stage 2: context retrieval pulls historical analogues — past matches between similar opponents under similar conditions. Stage 3: LLM reasoning is where Versi weighs the signals, debates the markets, and writes the analysis. Stage 4: post-processing validates the output, checks for contradictions, and assigns a confidence score before the prediction goes live.

After every match in our dataset, the actual result is logged against the prediction. Per-market accuracy tables (1X2, totals, BTTS, Asian handicap) feed back into the prompt template — if Versi systematically over-predicts goals in a specific league, we adjust the system prompt. Every two weeks we run a regression check against a holdout set; predictions older than 6 months are excluded from the live model so it stays current with how teams and managers evolve.

We publish hit rates live on /statistics, broken down by market, league, and confidence bucket. Aggregate is meaningless without context — what matters is, say, 78% on high-confidence Asian handicap calls in the Premier League across 600+ matches. We don't backfill; predictions are timestamped before kickoff and result evaluation runs on Sofascore-reported finals. If a number drops, you'll see it move on the dashboard before we hide it.

Predictions are probabilistic, not financial advice. A 75% confidence call still loses 1 in 4 times — bankroll discipline matters more than any individual pick. Don't chase losses, don't bet money you can't afford to lose, and never confuse our analysis with a guarantee. If betting stops being fun, take a break. For support: BeGambleAware (begambleaware.org) and GamCare (gamcare.org.uk) offer free, confidential help — please reach out.