AI Struggles with Horse Racing Predictions: Insights from Recent Data
After a promising trifecta success, issues with the horse racing AI reveal deeper problems in data usage and prediction accuracy.
On October 26, 2026, the journey into developing a horse racing prediction app using AI encountered unexpected challenges. Despite achieving trifecta wins in two out of three races, the overall performance over two days revealed troubling discrepancies.
Recent Results and Performance Review
In the latest betting simulation, the app recorded:
- Total races targeted: 6
- Trifecta wins: 2
- Total virtual purchase: ¥6,000
- Total virtual payout: ¥4,110
- Overall balance: -¥1,890
- Recovery rate: 68.5%
While a profit on the single day was encouraging, the cumulative loss over two days raised questions about the app's reliability.
Data Collection vs. Utilization
The endeavor to create a robust prediction app originated from the casual suggestion that ChatGPT could successfully predict horse racing outcomes. Initiating the project sparked a new challenge: to determine whether the AI effectively utilized the historical data collected.
Over time, a database was built, comprising:
- Target period: 1986–2021
- Total race entries: 137,273
- Raw race IDs: 10,047
- Time-series verification races: 2,303
This significant volume of data is meant to enhance predictive capabilities. However, verifying the integrity and usability of this information remained a challenge, particularly concerning rights and acquisition timing.
Discovering Key Issues in Prediction Mechanisms
A deep dive into the app's coding revealed the alarming fact that the settings for AI predictions were empty. Specifically, the model’s input allowlist was devoid of information across key stages:
- 24 hours before the race: 0 items
- 60 minutes before the race: 0 items
- 10 minutes before the race: 0 items
The lack of registered data inputs significantly undermined the credibility of the app's predictions. Although predictions were conducted using some past race references, the integration of newly gathered data to inform these predictions was fundamentally flawed.
The Importance of Unique Horse IDs
Further issues became evident during the evaluation of three races, notably the Kikuka Sho. Out of 51 horses analyzed, the critical identifying information—unique horse IDs—was missing. This absence presents a major complication: without unique identifiers, the AI cannot accurately track the performance of individual horses across various races. Identifying horses solely by name cannot ensure the data integrity necessary for consistent performance evaluations.
Refining Data Processing and Storage Concerns
The app also showed potential flaws in handling race data and verifying records. For instance, entry registrations were mistakenly counted as actual race results. This resulted in skewed historical data being fed into the AI system. Additionally, working with Google Sheets prompted storage capacity issues, necessitating enhancements to ensure accurate predictions without retroactive approvals of invalidated ones.
Understanding Predictions Under Dynamic Conditions
On October 25, even with shifting track conditions, the predicted performance for 51 horses remained unchanged. This highlighted the importance of distinguishing between two scenarios: when new information actually demands a revision in predictions and when the information has failed to trigger any adjustments.
Developing New Testing Mechanisms for AI Predictions
Addressing these challenges has resulted in the implementation of a new testing mechanism designed to responsibly utilize the amassed data. This process involves creating predictions without prior knowledge of race outcomes, thereby ensuring an unbiased testing environment. Two distinct stages of testing have been introduced:
- Stage 1: Hide results during prediction creation.
- Stage 2: Compare saved predictions with the actual outcomes after they have been established.
Through this rigorous validation method, known as walk-forward validation, the model's accuracy and effectiveness can progressively be assessed.
The Path Ahead for Horse Racing AI
As of now, although significant advancements have been made in improving the horse racing AI, challenges remain. The app can evaluate horses from multiple races, generate reference values, and save predictions effectively. Yet, fully grasping the nuances of horse racing through AI remains a work in progress.
Continuous testing and refinement are needed to align the predictive capabilities of the app with the complexities of horse racing. While the code updates have passed initial tests, it's crucial to remember that flawless testing does not equate to perfect real-world predictions. The journey of enhancing the prediction capabilities of the AI is ongoing, and with each hurdle overcome, the potential for more accurate forecasts draws closer.
