tony spencer
Colt
AbilityPlus V7
Building an Evidence-Based Horse Racing Rating System
Author: Tony
Version: 1.0
Foreword
For over twenty years I have built horse racing ratings and betting models. Like many people, I started with traditional handicapping methods, but over time I realised that many accepted ideas in racing are based more on opinion than evidence.
The aim of AbilityPlus is different.
Rather than trying to imitate the betting market or reproduce existing ratings, the objective is to estimate each horse's true chance of winning from historical evidence using machine learning and statistical modelling.
Everything described in this document has been tested using historical data. Many ideas worked. Many more failed. Those failures are just as important because they explain why the model eventually evolved into its current form.
Chapter 1
The Question
The obvious question seems simple:
Unfortunately, this turns out to be one of the most difficult machine learning problems.
A typical handicap contains between 8 and 16 runners.
Only one horse wins.
Every other runner loses.
That immediately creates an extremely unbalanced dataset.
For example...
| Result | Frequency |
|---|---|
| Winner | 1 |
| Losers | 7–15 |
Machine learning generally performs better when classes are more balanced.
This led to the first major research question.
Should the model predict winners directly?
Or should it predict something else that is easier to learn?
Chapter 2
Early Experiments
This chapter can contain exactly the material you've already been writing:
- Classification
- Regression
- Pounds beaten
- Lengths beaten
- Converting lengths into pounds
- Problems with eased horses
- Outliers
- Capping
- Winsorisation
- Using field averages
- Median versus mean
- Removing abnormal performances
Include examples of approaches that didn't improve the model. That shows the reader how the final design was reached.
Chapter 3
Building the Database
Explain the scale of the work.
For example:
- More than 1.1 million runners
- Nearly 200 engineered variables
- Over 15 years of racing history
- Multiple daily automated pipelines
- Historical reconstruction of races
- Continuous validation
Chapter 4
Feature Engineering
This chapter could easily be 15 pages on its own.
Break it into sections:
- Ability
- Form
- Speed
- Pace
- Sectionals
- Class
- Distance
- Going
- Draw
- Trainer
- Jockey
- Headgear
- Improvement
- Adaptive Momentum
- Breakthrough Detection
- Development Profiles
- Race-relative features
For every feature explain:
Chapter 5
Machine Learning
Explain why you selected CatBoost.
Discuss:
- categorical variables
- missing values
- interactions
- overfitting
- walk-forward testing
- blind testing
Chapter 6
Calibration
One of the strongest parts of the project.
Explain:
Raw Rating
↓
Probability
↓
Calibration
↓
100% Book
↓
Fair Odds
↓
Value
This is where AbilityPlus moves beyond being just a ratings system and becomes an odds compiler.
Chapter 7
Daily Production
Describe the complete daily workflow from declarations to final fair odds, showing how every stage builds on the previous one.
Chapter 8
Continuous Development
One of my favourite sections.
Describe how you never assume a new idea is better. Every change has to earn its place through blind testing.
Examples:
- V3 → V4
- V4 → V5
- V5 → V6
- V6 → V7
Each version should include:
- what changed,
- why it changed,
- whether it improved the results.
Chapter 9
Lessons Learned
This is where you can include the practical lessons from the project.
For example:
- Good ideas often fail when tested.
- More variables do not automatically produce a better model.
- Calibration is just as important as prediction.
- Honest validation is more valuable than impressive backtests.
- The market is a benchmark, not a target.
Final Chapter
The Journey Continues
Finish by making it clear that AbilityPlus is not a finished product. It's an evolving research project where every improvement must be supported by evidence before it becomes part of the production model.