I. Increase historical data scope to 10+ years ,may require changing the schema entirely for a larger historical API lower cost / normalizations of splits and economic shifts beyond a decade, current implemetnation is only ~2.1 years of historical daily dat a which doesn't encompass the black swan events / recessions that ive lived through. Roc-auc/pr scores likely to be signfiicantly lower but far more representative of the market's movements.
considerations:
pitfall 1: A model trained on say 100 years of data may learn patterns that no longer exist.
mitigations;
- rolling windows
- decay weighting
- regime models
instead of simply dumping every historical observation into training.
pitfall 2: same thing but for the future, paradigm shifts can make current patterns completely redundant.
mitigations;
- Online/continual learning with periodic retraining on recent windows
- Concept drift detection and adaptation mechanisms
- Ensemble models that blend long-memory and short-memory components
- Feature engineering focused on regime-invariant signals (e.g., volatility clustering, correlation structures) rather than raw - - price patterns
- Stress testing against synthetic regime-shift scenarios
- Monitoring distributional stability of input features in production and triggering alerts/retraining when shifts are detected
I. Increase historical data scope to 10+ years ,may require changing the schema entirely for a larger historical API lower cost / normalizations of splits and economic shifts beyond a decade, current implemetnation is only ~2.1 years of historical daily dat a which doesn't encompass the black swan events / recessions that ive lived through. Roc-auc/pr scores likely to be signfiicantly lower but far more representative of the market's movements.
considerations:
pitfall 1: A model trained on say 100 years of data may learn patterns that no longer exist.
mitigations;
instead of simply dumping every historical observation into training.
pitfall 2: same thing but for the future, paradigm shifts can make current patterns completely redundant.
mitigations;