Agent Roster

Classic technical-analysis indicators (RSI, MACD, Bollinger Bands, ATR) combined into a rule-based probability estimate.

Rate-of-change and cross-sectional momentum signals over multiple lookback windows to forecast continuation vs. mean-reversion.

Bayesian update rules applied to a prior probability derived from historical base rates, updated with likelihood ratios from current features.

Calendar-effect and intraday-seasonality patterns (day-of-week, month, options-expiry cycles) encoded as probability adjustments.

Detects low/medium/high volatility regimes using realised vol, VIX term-structure proxies, and regime-switching models.

Runs thousands of GBM simulations with bootstrapped drift and volatility parameters to estimate barrier-touch probabilities.

Gradient-boosted tree ensemble (XGBoost / LightGBM) trained on engineered OHLCV features with time-series cross-validation.

Reinforcement-learning policy trained via PPO on a simulated triple-barrier environment, outputting action probabilities as predictions.

Deep Q-Network (DQN) agent using a pure-numpy two-layer MLP (12→32→16→3). Continuous 12-feature state space, experience replay (capacity 500), target network, and epsilon-greedy TD(0) Q-learning — no heavy RL dependencies.

Sentiment and entity extraction from financial news and earnings transcripts, converted to directional probability scores.

Fine-tuned large language model specialised in financial reasoning; produces probability estimates from narrative context.

Microsoft Qlib alpha factor library integrated as a signal source; alpha factors are ranked and converted to probability outputs.

Monitors statistical drift in feature distributions using CUSUM and Page-Hinkley tests to down-weight stale agent signals.

Correlates commodity-relevant weather data (precipitation, temperature anomalies) with sector price movements.

Tracks open-source repository activity (GitHub stars, commit velocity, dependency adoption) as a leading indicator for tech stocks.

Ensemble meta-learner that blends the outputs of the other 16 agents using a stacking approach with isotonic-regression calibration.

Experimental generalist reasoning agent that synthesises all available signals through chain-of-thought prompting and self-consistency voting.