Agents And Forecasting

Configuration

The repo keeps behavior in config files:

configs/agents/
├── single_asset_ppo.json
└── multiple_asset_ppo.json

configs/predictions/
├── ar1_baseline.json
├── arimax_price_only.json
├── arimax_sentiment.json
├── gaussian_process_price_only.json
├── gaussian_process_sentiment.json
├── lstm_price_only.json
├── lstm_sentiment.json
├── ridge_arx_price_only.json
├── ridge_arx_sentiment.json
├── lightgbm_direct_price_only.json
├── lightgbm_direct_sentiment.json
├── lightgbm_price_only.json
└── lightgbm_sentiment.json

Training modules and forecast generators read configs rather than embedding direct file paths in the browser layer.

PPO commands

npm run preprocess
npm run train:single
npm run train:multi

Forecast commands

npm run predict:baseline
npm run predict:arimax
npm run predict:ridge
npm run predict:lightgbm
npm run predict:lightgbm:direct
npm run predict:gp
npm run predict:lstm

Single-asset PPO

The single-asset PPO module trains one model per commodity CSV under data/training/commodity_outputs/ and emits:

  • action

  • greedy action

  • hold/buy/sell probabilities

  • entropy

  • position

  • reward

  • net worth

Multi-asset PPO

The multi-asset PPO module trains a shared portfolio policy across the configured commodity set. Output rows contain one action-probability block per commodity plus shared portfolio state.

Forecast outputs

The forecast layer currently exposes:

  • a baseline trend forecast

  • Ridge ARX price-only

  • Ridge ARX price-plus-sentiment

  • ARIMAX price-only

  • ARIMAX price-plus-sentiment

  • LightGBM price-only

  • LightGBM price-plus-sentiment

  • Direct multi-horizon LightGBM price-only

  • Direct multi-horizon LightGBM price-plus-sentiment

  • Gaussian process price-only

  • Gaussian process price-plus-sentiment

  • LSTM price-only

  • LSTM price-plus-sentiment

Ridge ARX, ARIMAX, one-step LightGBM, Gaussian process, and LSTM emit two evaluation variants:

  • observed_history

  • recursive_path

Direct multi-horizon LightGBM is exposed as:

  • direct_multi_horizon

Dashboard loading

The browser demo does not retrain models. It discovers generated or committed CSV outputs under:

data/agent_outputs/
data/prediction_outputs/

The app-side loaders normalize those files into a stable UI shape for the charts.