Subteams

Team Members Focus Spec Doc
🗄️ Data & Features (~7) (red) quentin, aditya, ryan T, April, nikhil, erika, will Kalshi + NOAA + Binance ingestion, feature engineering → live_features.parquet see Reference section
🧠 Modeling & Intelligence (~9) (green) oilver, austin, jason, alex, justin, shreyes, evan, alina, vicky, Jollen, anzhe FinBERT/VADER NLP sentiment (internal) + XGBoost fusion model + probability calibration → predictions.json see Reference section
⚡ Execution (~4) (blue) Kelly sizing, risk checks, order management, dry-run → live trading see Reference section

references

Key Metrics

Metric What it measures Target
Brier Score Model calibration. Random = 0.25, perfect = 0.0 < 0.20
Sharpe Ratio Risk-adjusted return on backtest > 1.0
Win Rate % of trades that close profitably > 52%
Edge per Trade Avg (p_model − p_market) on winning trades > 0.05
Dry-Run Trades Proof the system is running autonomously > 50

Data Contracts

Team 1 → Team 2 · data/features/live_features.parquet (refreshed every 15 min)

Field Type Notes
contract_id str Kalshi market ticker e.g. KXBTC-25APR14-T100000
timestamp datetime UTC
market_price float [0–1] Normalized from Kalshi 0–100 cents
volume_24h float
days_to_resolution float
price_change_1h float Delta vs. 1h ago
price_change_6h float
market_category str "weather" / "crypto" / "sports"

💡 Note on NLP signals: nlp/ and models/ are both owned by Team 2 (Modeling & Intelligence). Sentiment scores are an internal Team 2 artifact — they flow from nlp/sentiment.py into models/predict.py at runtime and are cached in nlp/sentiment.json. The only output Team 2 exposes externally is signals/predictions.json.

Team 2 → Team 3 · signals/predictions.json (refreshed every 15 min)

Field Type Notes
contract_id str
timestamp datetime UTC
p_model float [0–1] Calibrated probability of YES outcome
confidence float [0–1] Model uncertainty — used for position sizing