stock-prediction

this is a collection of experiments training AI to trade.

i reccomend you use my stats dashboard, forecasting platform and API bitbank.nz for cryptocurrency trading.

A production RL trading system for crypto and equities. The core loop is: fast C market environment + PPO policy training + Chronos2 time-series forecasting + live execution on Binance (FDUSD/USDT) and Alpaca (US stocks).

Architecture

Raw OHLCV data (Binance / Alpaca)
        |
        v
  Binary export (export_data_*.py)
        |
        v
  C market environment (pufferlib_market/market_env.c)
        |
        v
  PPO training (pufferlib_market/train.py)
        |
        v
  Evaluate OOS (pufferlib_market/evaluate.py)
        |
        v
  Deploy (trade_execution_listener.py / binan/)

Chronos2 LoRA forecasts feed in as additional features at inference time (see forecast_cache_lookup.py, retrain_chronos2_hourly_loras.py).

Remote GPU training dispatches to RunPod pods and stores checkpoints in Cloudflare R2 (see src/remote_training_pipeline.py, scripts/dispatch_rl_training.py).

Quick start

git clone <repo>
cd stock-prediction
source .venv313/bin/activate
python -m pytest tests/ -x -q

Python 3.13 is the primary version. The .venv313 environment should be pre-populated. To rebuild it:

uv venv .venv313 --python 3.13
source .venv313/bin/activate
uv pip install -e ".[dev]"

GPU (CUDA 12.8) is required for training. CPU-only runs work for unit tests and backtests.

Key scripts

Script	Purpose
export_data_hourly_price.py	Export hourly OHLCV to .bin for RL training
export_data_daily.py	Export daily OHLCV to .bin
pufferlib_market/train.py	Train PPO policy on the C market env
pufferlib_market/evaluate.py	Evaluate a checkpoint OOS
pufferlib_market/autoresearch_rl.py	Automated hyperparameter sweep with OOS eval
retrain_chronos2_hourly_loras.py	Retrain Chronos2 LoRA adapters per symbol
run_crypto_lora_batch.py	Batch LoRA evaluation across symbols
forecast_cache_lookup.py	Disk-cached Chronos2 inference
trade_execution_listener.py	Live trade execution loop
binan/binance_wrapper.py	Binance API wrapper (FDUSD + USDT pairs)
scripts/dispatch_rl_training.py	Dispatch training jobs to RunPod
src/remote_training_pipeline.py	Remote training pipeline (R2 storage)
smart_test_runner.py	Change-aware test runner

C environment

The market simulator is written in C for speed:

cd pufferlib_market
python setup.py build_ext --inplace

This must be compiled with .venv313 active.

Remote training

Training jobs can be dispatched to RunPod GPU pods:

python scripts/dispatch_rl_training.py --config <config>

Checkpoints are synced to Cloudflare R2. See CONTRIBUTING.md for the SSH StrictHostKeyChecking=no rationale.

Testing

# Unit tests (fast, no GPU needed)
python -m pytest tests/ -x -q -k "not slow" --ignore=tests/integration

# Full suite (requires self-hosted GPU runner)
python -m pytest tests/ -x -q

Requirements

• Python 3.13 (.venv313)
• CUDA 12.8 for GPU training
• uv for package management (uv pip install, never plain pip)

License

MIT. See LICENSE.