Spike Detector
Pre-market binary classifier for significant intraday stock moves
The Problem
This is not a price predictor. It is a binary spike classifier: given everything knowable before the open (~9:15 AM ET), will a ticker make a significant intraday move relative to its own recent volatility? Thresholds adapt per ticker so a volatile name is not scored the same way as a quiet one.
The practical goal each morning is a short watchlist of the 2-3 names most likely to move ~3%+, with a direction call and risk-limited bracket orders.
Pipeline
Features span FinBERT news sentiment, technicals, macro, earnings, calendar effects, and a statistical-mechanics layer (entropy, sector magnetization, Boltzmann calibration, Ising overlay). A two-stage XGBoost stack first estimates P(spike), then P(up | spike).
Cloudflare Worker crons kick GitHub Actions for predict/trade at open, a delayed A/B account at 10:00 AM, end-of-day validation, and bi-weekly retrain. Telegram broadcasts the watchlist and trade confirms with no human in the loop.
Status
Fully automated paper trading is live: morning predict, dual-account A/B timing experiment, EOD validation charts, and scheduled retrain. Focus now is calibration quality and whether delayed fills beat open fills under real slippage.