Methodology

Every model is first built and tested against 10 seasons of real historical match data across 24 leagues worldwide. This establishes a baseline: does the approach show a genuine statistical edge over a long, real sample - not a handful of lucky matches.

Rather than testing a model only on the data it was built with, we run walk-forward validation: the model is evaluated exclusively on data that came after its training window, simulating what it would have actually seen live at the time. A model that only performs well in-sample never survives this step.

Once live, every signal is published before its match starts and its real outcome is recorded afterward - continuing the same out-of-sample discipline indefinitely. This is what separates a verifiable live record from a backtest that could theoretically be curve-fit after the fact.

Every public signal stores the exact moment it was published and the match's kickoff time. A signal is only ever included in the public record if it was published strictly before kickoff - enforced automatically, not by policy. Once published, a signal's match, odds and timing can never be edited - only its final result is added once known.

Stake sizing responds to model confidence and live drift from backtested expectations - never to whether a particular bet is about to win or lose (which is unknowable in advance). This keeps the risk system honest: it manages exposure, it does not chase outcomes.

Publishing the exact trigger conditions, filters or signals behind a model lets that edge be copied or arbitraged away - which would hurt every user relying on it. So we publish everything about outcomes (timestamps, odds, results, ROI, drawdown) and nothing about the "why" behind a specific pick. See the full, continuously-updated record at edgefinder.tech/results.