Phases 1–5: Data, features, and base models
Summary of the foundation layer (see roadmap.md for full task lists).
Phase 1 — Data pipeline
Multi-source ingestion: yfinance/Massive prices, FRED macro, news, optional Appwrite persistence, DuckDB local matrix, Scrapy spiders. DataPipeline orchestrates fetch with cache TTLs from config.toml.
Phase 2 — Feature engineering
Moving-average matrix (EMA, SMA, KAMA, VWAP, HMA), realized volatility, Hurst exponent, five-regime classifier. Sentiment preprocessing hooks (FinBERT later removed).
Phase 3 — Technical + macro models
Technical scoring: MA crossovers, RSI, MACD, Bollinger squeeze, volume confirmation. Macro: VIX, yield curve, DXY, per-ticker impact; regimes RISK_ON / TRANSITION / CRISIS.
Phase 4 — Sentiment model
Removed — torch/FinBERT dropped; weights redistributed to technical/macro in config.
Phase 5 — Fundamental model
Sector-relative valuation (P/E, P/S, EV/EBITDA) and analyst revision hooks from stored fundamentals.