YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
JW3-Trader-1 v2
Multimodal trading model — lower timeframe (1h) + reinforcement learning experiments. Text (OHLCV) + vision (chart patterns), MLX on Apple Silicon.
JW3.ai — the infrastructure layer for autonomous agents. Website · Whitepaper · Glossary
v2 Highlights
- Data: 1h candles (was 4h) — 925,931 windows from 28 Binance USDT pairs, 2017–2026
- Supervised model: 704K train / 222K eval (2025+ strictly out-of-sample)
- RL experiments: REINFORCE fine-tune with LONG/FLAT/SHORT actions, 4h holding window, 6bps round-trip cost
Architecture
- Text branch: LSTM over 48×1h OHLCV candles (7 features: O/H/L/C, volume, returns, spread)
- Vision branch: CNN reading candlestick chart patterns (96×96 rendered images)
- Fusion → policy: direction (UP/FLAT/DOWN) + volatility regime (low/med/high)
Honest Results
Supervised (primary release)
| Metric | Value | Baseline |
|---|---|---|
| OOS accuracy (2025–2026, 222K windows) | 37.8% | 33.3% random |
| OOS loss | 1.0856 | ~1.099 |
Trained on 704,043 windows, 2017–2024. Evaluated on 221,888 windows from 2025–2026 — never seen in training. Above-chance directional signal at 1h cadence.
RL (experimental)
REINFORCE with running-mean baseline, warm-started from the supervised model:
| Epoch | Train reward | OOS reward | OOS acc |
|---|---|---|---|
| 1 | +0.00037 | −0.00039 | 33.5% |
| 2 | +0.00057 | −0.00042 | 34.5% |
| 3 | +0.00064 | −0.00058 | 34.6% |
| 5 | +0.00072 | −0.00072 | 33.8% |
Honest verdict: RL increases training reward but does NOT yet generalize — OOS reward stays negative (the 6bps cost dominates the thin edge). The supervised model remains the better artifact. This is the transparent result we publish; RL with a better reward shape (risk-adjusted, regime-aware) is the next iteration.
Files
| File | Purpose |
|---|---|
jw3-trader-1-1h.safetensors |
Supervised 1h weights (primary) |
jw3-trader-1-rl.safetensors |
RL fine-tuned weights (experimental) |
jw3_train_1h.py |
Supervised 1h training |
jw3_train_rl.py |
REINFORCE RL training |
jw3_data_builder_1h.py |
1h dataset build (embedded prices for RL rewards) |
jw3_model_builder.py |
Shared architecture |
Quick Start
python3 jw3_infer.py --symbol BTC-USDT --interval 1hour # uses best weights
Requirements
pip install mlx mlx-lm numpy pandas
Disclaimer
Research model for educational purposes. Not financial advice.
Built by the JW3.ai team — autonomous AI trading agents, multi-chain DeFi (EVM · Solana · TON · Robinhood Chain), decentralized GPU compute, and the $JW3 token.
- Downloads last month
- -