V85 Model LabCan a model beat the crowd at Swedish trotting?

I built a system that estimates every horse's chance of winning in ATG's V85, compares it with how the crowd bets, and looks for the races where the crowd is wrong. It is tested like a trading model, with no peeking at the future and an uncertainty range on every number.

Open the Model Lab →

What I found

  1. The crowd is hard to beat. A model that sees only form trails the crowd's betting percentages: its gain is −10.9 %.
  2. But it has a habit. It bets too much on favourites. Correcting that gives a small, measurable edge on V85.
  3. The backtest is optimistic, because it uses the final betting percentages. So every round from 10 October 2026 is a sealed live test, and from November each prediction is published as a hash before the race.
Cumulative gain over the crowdLine chart of the running total of Model (with crowd %) gain over the crowd's betting percentages, in nats, from 2025-10-11 to 2026-10-07; it ends at +9.3 after 1,038 races. Nov 25 Feb 26 May 26 Aug 26 0 2 4 6 8 10 12 Cumulative gain (nats) +9.3
Gain against the crowd, added up race by race over the year. A line that keeps rising is a real edge; a random walk is luck.
CalibrationPredicted win chance against the share of horses that actually won, per probability bin, for Model (with crowd %) and The crowd (betting %), with 95 % ranges. Both follow the diagonal closely. 0% 10% 20% 30% 40% 50% 60% 70% Predicted win chance 0% 10% 20% 30% 40% 50% 60% 70% 80% Share that won Perfect The crowd (betting %) Model (with crowd %)
Calibration: when the model says 20 %, does the horse win about 20 % of the time?
Gain vs the crowd per modelAverage gain over the crowd per race with 95 % ranges, one dot per model: Model (with crowd %) +0.009 (range -0.002 to +0.021); Rank model (with crowd %) -0.023 (range -0.042 to -0.004); Form-only model -0.191 (range -0.233 to -0.147); Form-only rank model -0.198 (range -0.241 to -0.153). −0.2 −0.1 0.0 Gain vs the crowd (nats per race) Form-only rank model Form-only model Rank model (with crowd %) Model (with crowd %)
Model against the crowd, model by model, with 95 % ranges. Left off to keep the scale readable: Pure chance (−0.64).

The Model Lab

The Model Lab overview page: headline numbers, pipeline picture and cumulative gain chart

How it's built

Public race data becomes point-in-time features, a horse model gives calibrated win chances, these are compared with the crowd's betting, and a pool engine suggests a system or a skip. A walk-forward backtest and an experiment registry check every step.