Epsom Racecourse© Photo Healy Racing
The Epsom Derby takes place this Saturday afternoon and with AI (Artificial Intelligence) all the rage at the moment we thought it would be fun to see what predictions a computer would come up with for the big race.
In this project, we asked the computer boffins at irishracing.com to collate all the historical data we have on previous Epsom Derby winners and utilised machine learning techniques to identify the most likely winner among the current contenders for the 2023 Derby.
The Science
To effectively analyse the data, we applied feature encoding and scaling techniques. Categorical features were encoded using the LabelEncoder, assigning unique numerical labels to each category. Numerical features were standardised using the StandardScaler, ensuring all features were on a similar scale for accurate comparison.
To establish a baseline profile for the typical Epsom Derby winner, we calculated the median values of the historical winners’ encoded and scaled feature set. This allowed us to identify the most representative combination of attributes that characterised the typical winner.
To determine the similarity between each historical winner and the most typical winner, we employed cosine similarity. This metric enabled us to measure the similarity in feature space, providing a quantifiable measure of how closely each historical winner matched the most typical winner.
Weighting the data
Recognizing that certain features carry more significance in determining the winner, we assigned weights to the important elements within the data. These weights were based on domain knowledge and expert insights. Race distance of previous performances, most recent race performance, breeding, key Trial races, betting position were given higher weights to reflect their greater impact on the prediction.
Determining the most likely winner
Next, we compared the current contenders for the 2023 Derby against the most typical winner using the calculated similarity scores. Each contender’s feature set was encoded and scaled following the same methodology used for historical winners. The similarity scores were adjusted by multiplying them with the respective weights of the important features.
To identify the most likely winner from the current contenders, we aggregated the adjusted similarity scores for each contender. The median value was chosen as the summary statistic for the overall similarity scores. By considering both the feature similarity and the weighted importance of features, we obtained a comprehensive measure of each contender’s likelihood of winning the Epsom Derby.
Results and Prediction
Based on our analysis, the contender with the highest overall similarity score was identified as the most likely winner of the 2023 Epsom Derby.
Auguste Rodin winning the Vertem Futurity Stakes at Doncaster as a 2-y-o© Photo Healy Racing
This prediction suggests that Auguste Rodin would most likely win the Epsom Derby according to our ML model. A cosine similarity score of 0.72470 suggests a relatively high degree of similarity between the most similar contender and the average historical winner. The closer the score is to 1, the more similar the contender is to the historical winners in terms of their features used in the analysis.
Auguste Rodin is trained in Ireland by Aidan O'Brien who has already trained the winner of the big race a remarkable eight times. Auguste Rodin has been towards the head of the ante post betting markets for this year's Derby following a stellar two-year-old campaign which saw the son of Japanese Sire Deep Impact step up from a Maiden victory at Naas in July to land a Group 2 contest on Irish Champions Weekend in September before hitting the mark at Group 1 level in the Futurity Stakes at Doncaster in October.
Auguste Rodin flopped on his seasonal debut when sent off favourite in the English 2000 Guineas at Newmarket over one mile, but O'Brien hasn't lost faith in his star and remains confident that the colt has the talent to win an Epsom Derby.
Top Three Predictions
Here is a table showing the top three predictions according to our computer generated experiment.
