DeepRegatta/Guides

Guide · Retrace · IRC

What makes a boat beat its IRC rating?

Mostly boatspeed relative to the fleet. Across 597 IRC races and 14,842 boat entries, a model trained to spot rating-beaters puts 79% of its signal on speed-family factors — average speed and VMC versus the rest of the fleet — 14% on the rating itself, and under 2% on routing. The unglamorous conclusion: before optimising the course, make the boat go at its number.

Evidence · Retrace playbook S1a
What drives beating an IRC rating — share of model signal by factor family
SpeedSpeed: 79.3% of model signal79.3%Rating (TCC)Rating (TCC): 13.6% of model signal13.6%Race contextRace context: 3.2% of model signal3.2%StartStart: 2.2% of model signal2.2%RoutingRouting: 1.7% of model signal1.7%share of aggregated driver importance · CatBoost · AUROC 0.853 ± 0.02
View as table
Factor familyShare of signal
Speed79.3%
Rating (TCC)13.6%
Race context3.2%
Start2.2%
Routing1.7%
Weather / fleet / handling0%
Aggregated driver importance by category for the model that predicts top-25% corrected finishes. Weather, fleet-position and boat-handling families carried no measurable weight in this sample.Source: Retrace playbook S1a_irc · 597 races, 14,842 boat entries · data through 16 Jul 2026.

What the evidence shows

The three strongest individual drivers are all speed measures: average speed versus the fleet, speed as a ratio of the fleet mean, and VMC — real progress toward the finish — versus the fleet. The boat's TCC comes fourth: where you sit in the rating band still matters, but far less than how close to potential you sail. Weather factors carry no measurable weight here, which fits how handicap scoring nets out conditions shared by the whole fleet.

What that means for a crew

It orders the work. Time and budget spent on the speed loop — sails, trim, helming, weight discipline — is statistically better spent than time on elaborate routing plans. Then use a race replay of your own events to find exactly where the speed went missing.

Methodology

Retrace trains one model per racing scenario — here CatBoost over 120 features per boat entry — to predict a top-25% corrected result, with grouped k-fold cross-validation so no race appears in both training and test folds. Reported AUROC is 0.853 ± 0.02 across folds; individual driver importances are aggregated into the factor families charted above. Full details are on the Retrace methodology page.

Limitations

These drivers are correlational, not causal: they say what distinguished boats that beat their rating in this sample of 597 races, not what will change your next result. Features are derived from tracker data and inherit its noise. Findings describe this sample — they are not a rule of sailing.

Open the IRC playbook on Retrace free · no signup