A week ago I posted CatalystAlert here and someone asked about the AI version.
It's done.
*What's new in V2:*
1. *ML Impact Predictor* (XGBoost)
- Trained on 448 historical catalyst patterns (2015-2025)
- Direction accuracy: 76.67%
- Magnitude MAE: 3.09%
- Top features: catalyst_type (0.649), price_momentum_30d (0.149), volatility_30d (0.110)
2. *Likelihood of Approval (LOA) scores*
- Random Forest model
- Based on phase, indication, company track record
3. *Optimal Entry Predictor*
- Suggests when to enter before a catalyst
- Analyzes pre-catalyst price patterns
4. *Smart Money Score* (-100 to +100)
- Combines SEC Form 4 insider trades + 13F institutional holdings
- Flags unusual activity near catalyst dates
5. *Auto-retraining*
- Weekly cron job retrains models
- Drift detection triggers retraining if accuracy drops
*Technical details:*
- ML Service: Python FastAPI + XGBoost/Random Forest
- Training data: 1,884 catalysts, 490K stock price rows, 727 tickers
- Deployed on VPS with Docker Compose
- Models stored in Docker volume, hot-reloaded
*What stayed free:*
- 992 companies, 2,192 drugs, full catalyst calendar
- SEC insider filings, EMA Europe coverage
- Twitter bot posts movers 3x daily (@catalystalert)
The class distribution in training data is skewed (73.7% positive outcomes)
which inflates accuracy. Working on balancing techniques.
Questions I'm wrestling with:
- Is 76% accuracy actually useful for trading decisions?
- Should I open-source the ML pipeline?
- What would make the predictions more trustworthy?
Original post: https://news.ycombinator.com/item?id=46189127
catalystalert.io
