Skip to content

[Download Layer] increase historical data scope #19

Description

@ieatyoursushi

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

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions