Models and Reinforcement Learning

Ensemble forecasting, model fusion, and the reinforcement learning loop

Ensemble forecasting

• Base learners: LSTM/Temporal CNNs trained per horizon (short, medium)

• Calibration: Temperature scaling and isotonic calibration on validation splits

• Regularization: Dropout + weight decay; early stopping on information leakage tests

Model fusion

• Meta-learner: Gradient boosting + shallow MLP to stack base predictors

• Cross-validation: K-fold or time-series split; leakage-safe folds

• Objective: Maximize directional utility under turnover and cost constraints

Reinforcement loop

• State: Recent forecasts, realized slippage, risk rule activations, and PnL

• Action: Strategy selection and parameterization

• Reward: PnL with drawdown penalties and risk-adjusted performance (e.g., Sharpe proxy)

• Updates: Online policy updates; periodic batch retraining

Governance and safety

• Roll-forward evaluation before model promotion

• Canary deployments with kill-switches and automatic rollback

• Model registry with versioned artifacts and audit trails