Real Data, Scientific Models, Real Results

Human vs Machine Learning

Introduction

I have built prediction models using Python to forecast horse racing outcomes based on public data from the Jockey Club. The models I used include Neural Networks, Logistic Regression, and Random Forest. However, the prediction accuracy is often below 50%. I was wondering if that’s the best result we can get from forecasting or prediction prespective.

To address this, I set up a website at https://nyp.pythonanywhere.com/racecard/?1 and invited horse racing fans to submit their choices. For each race, they can submit 3 selected horses.

The competition

In the latest match at the Hong Kong Jockey Club, there were 5 human participants whose performance I tracked. I was amazed to find that some participants did exceptionally well. For instance, a website member with the username “ML” consistently and accurately picked 2 out of 3 horses for Place selection.

The final results showed that humans won the competition with 52% accuracy, while Neural Network Regression came in 2nd with an accuracy rate of 48%.

Conclusion

There must be some factors lacking in the machine learning data that give human judgment an advantage. This is the area I am going to explore further. I hope that my models will eventually outperform humans in the long run.