AI models
August 21, 2026
The analyses offered by SignalTurf rely on several artificial intelligence models working together.
Rather than relying on a single analysis method, SignalTurf combines different approaches to get a more complete view of each race. This architecture helps reduce the limitations inherent to any single model and improves the overall consistency of the predictions.
This chapter presents the main principles behind this approach, without going into the technical details that make up SignalTurf’s know-how.
Why use several models?
In artificial intelligence, no model is perfect.
Each approach has its own strengths and limitations. Some are better at identifying general trends, while others excel in more specific situations.
Instead of looking for a universal model, SignalTurf chooses to combine several specialized models that analyze the same race from complementary angles.
This approach, commonly used in machine learning, generally produces more robust analyses than relying on a single model.
Specialized analyses
Each model has its own way of interpreting the information available before a race.
Without revealing their internal workings, each can be thought of as bringing a different perspective to the same event.
The results obtained are then compared and combined to produce the final analysis shown in SignalTurf.
The goal isn’t for one model to “be right” over the others, but for their analyses to complement each other.
An identical method for every race
Regardless of the meeting or discipline, the models always apply the same analysis method.
They aren’t manually adjusted for a particular race, and no human intervention alters their results before publication.
This approach ensures a consistent evaluation of every race analyzed by SignalTurf.
Learning from history
Before being used in production, the models are trained on an extensive history of French horse races.
This training allows them to identify relationships between many characteristics observed before a race starts and the results seen afterward.
They don’t memorize past races. They learn to recognize statistical patterns that they then reuse to analyze new races.
Each new race is therefore evaluated independently, based on the information available at the time of calculation.
Combining the results
Once the individual analyses are complete, SignalTurf brings their conclusions together to produce a single analysis.
This combination is an essential step in how the app works.
It notably helps to:
- limit the influence of any single model;
- take advantage of each approach’s strengths;
- produce more stable analyses over time.
The methods used to perform this combination are specific to SignalTurf and evolve regularly through ongoing research and experimentation.
Constant evolution
Artificial intelligence models aren’t set in stone.
They undergo continuous improvement to benefit from advances in machine learning and the ever-growing history of races available.
Some improvements concern the models themselves directly, while others relate to data preparation, how analyses are combined, or the evaluation methods used.
The goal always remains the same: progressively improving the quality and stability of the analyses offered.
Performance measured before every update
Every new version of a model is evaluated before being integrated into SignalTurf.
Its performance is compared to that of previous versions on independent datasets to verify that an improvement is genuinely observed.
An update is therefore not adopted because it’s newer, but because it demonstrates a measurable benefit according to the criteria SignalTurf uses.
This approach favors real improvements over purely theoretical changes.
Early access to new models
Premium subscribers with early access automatically use the most recent versions of the models as soon as they go into production.
When a new model is deemed mature enough, it becomes the reference model for these users.
After an observation period, it’s gradually rolled out to all users and replaces the previous version.
This setup allows improvements to be introduced gradually while maintaining the service’s stability.
Early access to new models
Early access lets Premium subscribers permanently benefit from the most recent generation of artificial intelligence models.
This works through a rotation system.
When a new generation of models demonstrates better performance than the one currently in use, it’s first reserved for Premium subscribers. At the same time, the model that was previously reserved for them becomes the public model available to all users.
In other words, Premium subscribers always have access to the most recent generation, while free version users benefit from the previous generation, whose performance has already been validated.
This setup allows Premium subscribers to permanently benefit from the most recent and highest-performing models validated by our internal evaluations.
Free version users then benefit from these improvements once the next generation goes into service and the previous Premium model becomes the new public model.
Artificial intelligence… with its limits
Even the highest-performing models can’t predict the outcome of a horse race with certainty.
The analyses produced by SignalTurf rely exclusively on the information available before the race starts and remain subject to the uncertainty inherent to any sporting competition.
The models aim to identify the most plausible scenarios based on the data analyzed, but they never constitute a guarantee of results.
That’s why SignalTurf always pairs its predictions with other indicators, such as the confidence level and historical statistics, to provide more context for their interpretation.
Go further
You now know the general principles behind the artificial intelligence models used by SignalTurf.
The Confidence level chapter explains how the analyses produced by these models are complemented by an indicator designed to make them easier to interpret.
You can also check out Performance to learn how the models’ results are measured and presented in the app.