AI Learning
The AI continuously learns — and proves itself first
AI continuously learns from completed predictions and evaluates candidate models before deployment. A candidate model is only promoted when it objectively beats the production model on data neither has ever seen — otherwise it is rejected and recorded.
How the loop works
- Predictions are stored and evaluated only after their horizon fully matures
- Online learning is gated: improvements must hold on out-of-sample windows
- Full retraining runs on a fixed schedule with walk-forward validation
- Deployment requires a strict margin over the current production model
- Every decision — deploy, reject, rollback — is logged and auditable
Why so strict? Many “AI trading” products improve their reported numbers by accidentally (or deliberately) letting future information leak into training, or by cherry-picking favorable windows. The deployment gate here is deliberately hard to pass, and it runs on out-of-sample data the candidate has never influenced — so a promotion means something.
The learning loop
Honest measurement, by design
Four evidence classes that are never mixed.
Live observed
Real predictions issued by production models, evaluated only after maturity. This is the only class that reflects what subscribers would actually have seen.
Out-of-sample
Walk-forward validation folds and the protected test set used for deployment decisions. Data the model never trained on.
Backtest
Historical simulations with realistic fees and slippage. Useful context, never presented as live performance.
Paper trading
Forward simulated trades with no real money. Clearly labelled everywhere it appears in the product.