The Probability Receipts: What 77,457 Starters Say About CHV

Across 11,058 races, CHV's pre-race win probabilities closely matched what actually happened within 11 probability groups.

CHV Analysis

• Sample: 11,058 races and 77,457 Thoroughbred starters
• Test: Compare each horse's pre-race modeled win probability with the actual win rate of horses in the same probability group
• Result: The weighted mean absolute gap across the 11 groups was only 0.39 percentage points
• Consistency: Eight of the 11 groups finished within 0.75 percentage points of the model, and every group finished within 1.2 points
• Largest absolute gap: The 40–<45% group won 43.44% of the time versus 42.29% modeled, a difference of +1.15 points
• Largest model overestimate: The 25–<30% group won 26.22% of the time versus 27.33% modeled, a difference of −1.11 points

This is the first article in a CHV series showing the statistical receipts behind the model. This installment is strictly about probability. Future articles will examine risk and price.

Here are the predictions. Here is what happened.

Here are 11,058 races and 77,457 starters. Here are the model's win-probability predictions made before the races occurred. Here is what happened. Now show the calibration.

That last step matters. A probability model should not be judged by one winner, one losing streak, or one memorable longshot. It should be judged by whether horses assigned similar chances win at similar rates over a large sample.

If the model gives a group of horses an average 20% chance to win, roughly 20 of every 100 horses in that group should win. If the actual rate is far above or below 20%, the model is miscalibrated in that range. If modeled and actual rates stay close across the probability spectrum, the model is doing the job it claims to do: estimating win probability.

What the calibration shows

The analysis compares the average modeled W% in each group with the group's actual win rate. The perfect-calibration line represents exact agreement. A result above that line means horses won more often than modeled; a result below it means they won less often.

The CHV results remain very close to perfect calibration from the lowest-probability horses through the 50%+ group. The weighted mean absolute gap across the 11 groups was just 0.39 percentage points. Eight groups finished within 0.75 points, and every group finished within 1.2 points.

Where the model misses the most

The largest absolute gap is in the 40–<45% group. Those horses carried a 42.29% average modeled probability and won 43.44% of their starts. The model was low by 1.15 percentage points.

The 25–<30% group is almost as far away in the other direction. Its 27.33% average modeled probability produced a 26.22% actual win rate, so the model was high by 1.11 points. This is the model's largest overestimate and applies to a larger sample of 5,191 starters.

Both gaps deserve attention as the model continues to develop. They also need context. Even the largest difference across all 11 groups is only about 1.15 percentage points. At 35–<40%, the gap narrows to just 0.03 points. The 50%+ group is separated by only 0.14 points.

That is what useful calibration looks like: consistent agreement between forecast probabilities and observed results across a large sample.

Why probability comes first

Every CHV letter grade, adjusted morning line, Value reading, and wagering decision begins with the same question: How likely is this horse to win?

Probability creates the foundation. A horse with a 30% win probability should be treated differently from one with a 5% probability, even before price enters the discussion. The model's W% gives CHV users a common scale for comparing horses across fields, tracks, distances, surfaces, and race types.

The calibration results support using that number as an estimate rather than a decorative score. Horses grouped around 5%, 20%, 40%, or 50%+ won at rates that closely matched those expectations.

The practical lesson is simple: read W% as a probability. It is not a guarantee that a horse will win today. A 40% horse still loses most individual races. Over many comparable starters, however, the result should move toward the modeled rate.

What this analysis does—and does not—show

This analysis evaluates probability calibration. It answers whether CHV's pre-race W% estimates align with actual outcomes after horses are grouped by their modeled probability.

It does not answer whether every well-calibrated horse is a good bet. Profit depends on the price available relative to the probability. It also does not describe the risk created by uncertainty, field composition, or the way probabilities are distributed within an individual race.

Those are separate questions, and they deserve separate evidence. The next articles in this series will examine risk and then price.

Related CHV resources

• Learn how modeled W% becomes a quick visual signal in CHV Letter Grades.
• Review factor and field definitions in the CHV Glossary.
• Watch the workflow in the CHV Tutorial Videos.
• Read race analysis and model research on CHV Insights and in the CHV Substack archive.

Continue your CHV journey

Use the Charting Horse Value app to see modeled W%, letter grades, factor rankings, and race-level analysis for today's cards. Full app access is included with a monthly or annual paid subscription.