Over 3.5 goals happens in roughly 30% of matches. An honestly calibrated model therefore almost never rates it above 50% — so a binary pick column read "Under" on 99% of rows. True, but useless.
The calibration itself was excellent: across 30,582 matches our stated over-3.5 probabilities averaged 29.6% against an actual rate of 30.3%, tracking bucket by bucket. The failure was the display convention, not the model. We tested variants that widened the probability spread to force more Over calls — every one graded worse. You cannot make a 30% event more predictable by shouting.
The fix is a three-way call: Over at 45%+ (those land ~47% — genuinely strong spots), Under below 28%, and an explicit no-call between, where the percentage is shown but no pick is made. We also tested the popular belief that goals are rising season on season: across our whole dataset the average has been flat at ~2.75 per game since 2022 — the 2023-24 winter spike was a one-off, not a trend.
Method: walk-forward testing — every prediction is made using only information available before kick-off, then graded against what happened. Studies run June-July 2026 on the site's own dataset. Questions: contact us.