HFT Signal Architecture Backlog

Inspired by a Polymarket BTC scalper that turned $0.90 → $408k by exploiting CLOB lag vs spot price. The latency arbitrage itself doesn't transfer to stocks, but several architectural patterns do.

Prioritized

Idea	Status	Notes
Prediction confidence gate	Backlog	If max(class_probabilities) < 0.65, emit "no prediction" instead of forcing a weak call — improves signal quality without new data
Multi-signal convergence requirement	Backlog	Only fire a prediction when 2+ independent signals agree (price momentum + sentiment + vol regime); mirrors the bot's skip-on-conflict logic
Real-time WebSocket price feed	Backlog	Replace hourly yfinance polling with Binance/Alpaca WebSocket for US tickers — reduces data latency from ~60min to seconds
Hard skip rules	Backlog	Structured abstain conditions: thin sentiment data, low price history, conflicting momentum signals — prevents noisy predictions from entering the dashboard

On Radar

Idea	Notes
Graph-based signal topology	Map tickers/signals as nodes, correlations as edges; detect convergence clusters before predicting — high complexity, high payoff for sector analysis
Per-trade / daily prediction caps	Limit how many predictions fire per cycle; prioritize highest-confidence calls only — mirrors the bot's daily cap + hard stop discipline
Signal lag detection	Measure and log how often sentiment lags price moves; if consistent lag exists, exploit it as a leading indicator