Machine Learning for Predicting Race Results: Decoding the Code of Victory and Defeat from Data

Introduction

Stage 18 of the 2025 Tour de France, a queen stage in the Alps. The night before the race, a data analyst for one team ran a machine learning model on their laptop. The model comprehensively analyzed the historical performance of all top 20 riders in the general classification, recent training loads, weather forecasts, route profiles, and physiological trend indicators from the first 17 stages. The prediction: their team leader had a 73% probability of reclaiming the yellow jersey on the final HC climb.

The next day, the prediction came true.

This isn’t magic—it’s the mature application of machine learning in sports prediction. When we have enough data, algorithms can uncover patterns that human analysts struggle to detect.

Core Methods of Machine Learning in Sports Prediction

Feature Engineering: Turning Cycling Knowledge into Data

The quality of a machine learning model depends 80% on the quality of feature engineering. In cycling race prediction, key features include:

Rider Historical Features:

• Performance distribution in the same type of event over the past 3 years
• Historical performance on similar route profiles (climbing, time trial, flat)
• FTP trends and power curves for the current season
• Grand Tour experience index (how many Grand Tours participated in)
• Age and career stage

Real-Time Physiological Features:

• HRV trends over the past 7 days (reflecting recovery status)
• Cumulative fatigue index (chronic training load CTL based on TSS)
• Body weight change trends
• Stability of physiological indicators (the smaller the fluctuation, the more stable the form)

Environmental Features:

• Temperature and humidity forecasts
• Wind direction and speed
• Altitude
• UV index

Route Features:

• Total elevation gain and climb distribution
• Gradient profiles of key climbs
• Technical difficulty of the course (number of corners, descent danger level)
• Road surface type

Tactical Features:

• Team strength (supporting capability of teammates)
• General classification time gap distribution
• Race phase (conservative early vs. attacking late)
• Opponents’ historical tactical preferences

Model Selection

Different prediction tasks suit different models:

Gradient Boosting (XGBoost / LightGBM)

This is currently the most commonly used model family in sports prediction. Reasons:

• Naturally handles mixed feature types (numerical, categorical, sequential)
• Built-in handling for missing feature values
• Provides feature importance rankings, making model decisions interpretable
• Fast training speed, suitable for frequent updates

Use Cases: Predicting race placings, classifying outcomes (whether a rider makes the top 10)

Recurrent Neural Networks (RNN / LSTM)

For time-series data—such as a rider’s performance trends across a multi-day race—recurrent neural networks can capture temporal dependencies.

Use Cases: Predicting daily performance changes in multi-day races, fatigue accumulation effects

Bayesian Models

Bayesian methods are particularly well-suited for handling uncertainty. In situations with incomplete information (for example, a rider lacking recent race data), Bayesian models can provide predictions with confidence intervals.

Use Cases: Estimating a rider’s true ability range, calculating the probability of an upset

Graph Neural Networks (GNN)

Team tactics in cycling involve complex interactions between riders. Graph neural networks can model these relationships:

• Nodes: each rider
• Edges: cooperative (teammate) or competitive relationships between riders
• Edge weights: historical interaction patterns (e.g., Rider A always pulls for Rider B in the final 5 km)

Use Cases: Predicting the impact of team tactics, analyzing peloton dynamics

Ensemble Methods

The most accurate prediction systems typically don’t rely on a single model but instead integrate predictions from multiple models:

Final Prediction = w₁ × XGBoost Prediction + w₂ × LSTM Prediction + w₃ × Bayesian Prediction

The weights w₁, w₂, w₃ are dynamically adjusted based on each model’s historical performance on the validation set.

Real-World Case Studies

Case Study 1: Grand Tour General Classification Prediction

Objective: Predict the final top 10 in the general classification before the start of the three Grand Tours (Tour de France, Giro d’Italia, Vuelta a España).

Dataset:

• Complete results data from the three Grand Tours, 2010-2024
• UCI rankings and race results from the 6 months prior to the race
• Route features (total elevation gain, time trial kilometers, high-altitude sections)
• Team budget and roster strength indicators

Model Performance:

Metric	Accuracy
Top 10 hit rate	7.2/10
Top 3 hit rate	2.1/3
Winner prediction accuracy	58%
Spearman rank correlation	0.82

Notably, the model was more accurate at predicting “consistent” riders, while predictions for “explosive” or “highly variable form” riders were poorer—which aligns with intuition, as instability is inherently difficult to predict.

Case Study 2: Real-Time One-Day Race Prediction

Objective: Provide real-time updates of each rider’s win probability during a one-day road race.

Data Streams:

• Real-time GPS position and speed
• Power meter data (if publicly available)
• Features of the remaining route
• Current weather data
• Peloton split status

Model Update Frequency: Every 60 seconds

This kind of real-time prediction system is similar to a poker odds calculator—as the race progresses, more information is revealed, and the predictions gradually converge toward the final outcome.

Spectator Experience Application: Displaying each rider’s real-time win probability percentage on the live broadcast, increasing the tension and engagement of watching.

Case Study 3: Personal Training Goal Prediction

For amateur cyclists, machine learning is equally valuable. Models can predict based on your training data:

• Estimated finish time for a target event
• Probability of achieving a specific performance goal (e.g., finishing Wuling in under 3 hours)
• Optimal pacing strategy
• Required training volume and intensity adjustments

Limitations and Ethics of Prediction

Unpredictable Factors

No matter how powerful machine learning is, it cannot predict:

• Mechanical failures: flat tires, dropped chains, derailleur damage
• Crashes: road hazards, sudden weather changes, peloton chaos
• Sudden health issues: gastrointestinal problems, cramps, heat exhaustion
• Mental breakdowns: performance collapse due to excessive pressure
• Tactical surprises: completely unexpected attack timing

These “black swan events” are part of what makes sports compelling, and they are factors that prediction models can never fully capture.

The Risk of Overfitting

Sports data has relatively small sample sizes (only a few major races per year), making overfitting a serious risk. If a model performs perfectly on historical data but fails on new events, that’s a classic symptom of overfitting.

Countermeasures:

• Rigorous cross-validation (holding out the most recent year’s data as a test set)
• Regularization techniques to limit model complexity
• Continuous monitoring of model performance on new data with timely updates

Ethical Considerations

Impact on Betting Markets

Highly accurate race prediction models, if used for sports betting, could raise fairness concerns. Organizations with advanced models could gain an unfair advantage in betting markets.

Rider Privacy

The physiological data required by models (HRV, power, heart rate) involves riders’ health privacy. The collection, use, and sharing of data requires explicit consent from riders.

Race Suspense

If predictions become too accurate and are widely disseminated, could they reduce spectator interest? A moderate degree of uncertainty is central to the appeal of sports.

Application Potential for Taiwanese Cycling Events

Existing Data Foundation

Cycling in Taiwan has already accumulated a considerable data foundation:

• Strava Data: riding records from a large number of Taiwanese cyclists
• Race Results Databases: historical results from the Tour de Taiwan, KOM, and county-level cycling races
• Route Data: complete GPS and gradient information for Taiwan’s major cycling routes
• Weather Data: high-resolution meteorological data provided by the Central Weather Administration

Feasible Application Directions

1. KOM King of the Mountain Prediction: Predict the top 10 finishers and the winning time based on the participant list and historical results
2. Personal Goal Assistant: Help riders predict completion times for specific routes (e.g., Wuling, Fengguizui, Beiyi)
3. Strava Leaderboard Analysis: Analyze leaderboard trends and predict the next challenger to take the top spot
4. Live Race Broadcast Enhancement: Display prediction information in real time on the broadcast screen

Conclusion

Machine learning will not replace racing—it allows us to understand racing more deeply. When algorithms reveal hidden patterns in the data, our appreciation of the sport does not diminish; it grows. Because we finally see the physiological logic behind every attack and the statistical improbability in every comeback victory.

Prediction models tell us what “should” happen, and the beauty of sport lies in its frequent reply—the data is wrong. It is precisely this defiance of prediction that keeps sports forever captivating.

Related Reading

• Post-Race Analysis and Goal Setting: Turning Every Race into a Stepping Stone for Progress
• Research Review: Shifting Efficiency and Precision Test Report of Electronic vs. Mechanical Shifting in Muddy Off-Road Conditions: Physiological Characteristics of Elite Athletes (Article 1013)
• Research Review: Shifting Efficiency and Precision Test Report of Electronic vs. Mechanical Shifting in Muddy Off-Road Conditions: Physiological Characteristics of Elite Athletes (Article 1316)
• Research Review: Shifting Efficiency and Precision Test Report of Electronic vs. Mechanical Shifting in Muddy Off-Road Conditions: Physiological Characteristics of Elite Athletes (Article 1247)