Technical companion to /projects/quant. For someone who wants to see the infrastructure under a strategy, not just a backtest curve.
I Use This When...
I want the full loop: ingestion → factor ranking → regime detection → sentiment overlay → paper execution → scheduled cycle → audit.
Why HMM for regime detection, not a trend filter
Trend filters (e.g. price > 200-day MA) work but lag at turning points by design — they need the trend to have already broken before they switch sides. An HMM over return features captures the latent state directly, so the regime flip can be detected earlier than the price has confirmed.
The trade is sensitivity to inputs: an HMM with too few features will chatter; with too many it overfits the calibration window. The current choice is a small feature set anchored to FRED macro inputs (rates, breadth) so the regime is grounded in macro context, not just price.
Why FinBERT as an overlay, not a primary signal
News sentiment over the swing-trading horizon mostly averages out — big-news spikes are short-lived noise compared to the multi-day multifactor signal. FinBERT therefore lives as an overlay that can tip a tie or de-risk a position when sentiment is broadly negative, not as a driver in its own right.
A higher-frequency strategy could push sentiment into the primary signal stack; for swing it would dominate the noise floor.
Why paper, not live
The project is for learning and showing infra quality. Going live would change the work from "build the loop" to "prove the edge" — a different project entirely, and one where being wrong has actual cost. Paper through the IBKR API gets all the plumbing-level lessons (order types, fills, slippage simulation, scheduling) without the risk surface.
Why Streamlit + FastAPI both
Streamlit is excellent for the dashboard (fast iteration on charts, no manual front-end build) and bad as an API (poor request semantics, no clean machine-callable surface).
Splitting the project gives:
- FastAPI for everything programmatic (paper cycle endpoints, historical queries, signal probes).
- Streamlit for the human-facing dashboard.
Both speak to the same backing data, neither pretends to do the other's job.
Why the 15-minute cycle, not real-time
For US-equities swing trading the signal moves at the speed of daily / multi-hour data, not tick. A 15-minute cycle is fast enough to react to intraday regime flips without burning compute on noise. The cycle re-ingests prices, re-ranks the watchlist, re-checks the regime, re-overlays sentiment, and emits any order adjustments.
What broke first
TODO: my recollection is the 15-min cycle's first cron offset missed market-open by enough that regime-flip detection lagged a full bar. Add the timezone / alignment fix.
What I'd rebuild
- Persist cycle results into TimescaleDB so every past decision can be replayed against the inputs it actually saw — turns the system into an audit trail, not just a runtime.
- Surface Sharpe / max-drawdown vs SPY directly in the dashboard so the "is this strategy working?" question has a one-glance answer.
- Walk-forward retraining cadence for the HMM so the regime model doesn't drift on stale calibration windows.