How Did We Do?

Paste a club or tournament pairs results page — the whole page is fine, junk and all. Every pair's actual score is compared with what the Bridge Ratings site predicts against this exact field (ratings updated through 08/15/26), with a percentile to show how unusual the result was. Nothing you paste leaves this page.

Reading it. Predicted is the expected percentage against exactly the pairs it was ranked against — in a Mitchell each direction is its own competition; a Howell comprises one field. Percentile (personal) is about THAT PAIR, not the field: where the actual result landed among the outcomes the model considered likely for that particular pair against that particular field: 90% means better than nine sessions in ten, 50% is dead typical. It is read from the measured spread of real results (not an assumed bell curve), and it accounts for field size, movement, boards, and rating coverage.

Unrated players (marked *) are given the average rating of the rated players in the same field. Measured across recent sessions, unrated players actually run about 0.4 percentage points weaker than that average, so a pair with an unrated player is very slightly flattered by its prediction. Robots are the exception: they use the robot's true rating (one shared rating for the club's robots, shown in parentheses) — robots appear on the ratings page but are not ranked.

Is this for real? Checked against 9,343 pair-sessions played after the ratings were frozen: actual results landed uniformly across the predicted percentile bands (largest band off by 0.7% on 10-point bands) — the model neither exaggerates nor undersells how unusual results are.

methodology

For each pair, the predicted percentage is the average over the other pairs in its ranking group of a logistic win-probability in the two pairs' combined ratings — the same arithmetic as the main page's pair predictor, which makes the prediction self-normalising: predictions in a group average exactly 50, like the real percentages do.

The percentile comes from the empirical distribution of (actual − predicted) over 114,742 pair-sessions from 2025-06-01 to 2026-07-25. Because a session's luck depends on how it is measured, the residual is first standardised by a fitted scale — a log-linear model in field size, boards, movement (Howell fields are noisier at the same size), and whether the pair carried unrated players (fitted by heteroscedastic-normal maximum likelihood). The standardised residual is then ranked against a 999-quantile table of the same standardised residuals.

Out-of-sample check: on 9,343 pair-sessions dated after the ratings file was built (so none of them could influence it), the percentiles came out uniform — every 10-point band held 9.3–10.5% of results (expected 10%), Kolmogorov–Smirnov 0.007 against a 5% critical value of 0.014, and the extreme tails matched (0.93% of results below the 1st percentile, 1.18% above the 99th). The same uniformity holds separately for Mitchells, Howells, small and large fields, and fully-rated versus partly-rated pairs.

If something looks wrong please contact me; I'm trying to make this as useful as possible.

Doug Couchman is a bridge player and instructor, and a private tutor specializing in graduate admissions exams (LSAT, MCAT, GMAT, and GRE). Click here for more information.