Most market models are trained on the last 20–30 years of data. That window contains exactly one rate cycle, one inflation regime, and one global liquidity environment. A real macro view needs centuries — World Wars, gold standards, the 70s, the 30s — and most systems quietly throw that data away.
Compiled a unified daily/monthly dataset from 1871 forward, with consistent definitions across regimes (gold standard, Bretton Woods, fiat).
Built regime-detection models that learn from cycles instead of fitting on the latest 10 years.
Layered a News Brain that aligns market moves with the actual narrative drivers of each era — not just the price action.
Ship a v3 (data-model-2) rebuild with cleaner abstractions for adding new asset classes without retraining the whole stack.
Built historical data and modelling components for studying regimes across different market eras.
Cross-period evaluation remains necessary to establish whether a model generalizes beyond particular regimes.
The research direction emphasizes versioned inputs and reproducibility rather than presenting historical fit as a forecasting guarantee.