Every deployed model goes through the same six-stage pipeline — dozens of candidates trained, only the validated survivors ever get combined into what runs on your account.
Independent baseline models, each trained, validated, and ranked on their own before anything gets combined.
Sequence and pattern models trained on the same data, covering recurrent, convolutional, and transformer-family architectures.
Trained after supervised learning, optimizing directly for trading outcomes — not just prediction accuracy.
Reward shaped from profit, risk, drawdown, holding cost, and transaction cost — not raw price prediction.
Every model above is tuned, not left at defaults.
Training accuracy is never trusted on its own — every model has to survive testing designed to break it.
Only models that survive validation get combined. Weak models are rejected outright, not averaged in.