ChanFormer is a channel-wise Transformer architecture for multi-asset financial time-series forecasting. It is meticulously designed to study cross-asset interaction and representation learning without temporal patching, serving as a controlled architectural baseline against patch-based models.
- Multi-asset modeling (crypto, equities, or blended)
- One-step-ahead OHLCV forecasting
- Channel-wise Transformer encoder
- Reversible Instance Normalization (RevIN)
- Cross-channel mixer for asset interaction
- Three experimental scenarios (crypto / stock / blended)
ChanFormer
├── README.md
├── LICENSE.txt
├── dataset/
│ ├── crypto/
│ │ └── SUIUSDT_1d_full.csv
│ └── stock/
│ └── AAPL_1d_full.csv
└── src/
├── ablation.py
├── backtesting.py
├── ChanFormer.py
├── emissions.csv
├── overhead.py
├── PatchTST.py
└── Transformer_benchmark.py
All training, evaluation, and forecasting logic is contained in
ChanFormer.py.
Instead of splitting time into patches, ChanFormer:
- Takes a 90-day historical window
- Flattens time per feature channel
- Projects into embedding space
- Applies Transformer encoding
- Mixes channels via MLP
- Predicts next-day OHLCV
Each feature channel becomes one token, removing temporal cross-attention and isolating the impact of channel interaction.
| # | Cryptocurrency | Stock |
|---|---|---|
| 1 | ATOMUSDT | AAPL |
| 2 | BCHUSDT | AMZN |
| 3 | DOTUSDT | AVGO |
| 4 | HBARUSDT | BRK-B |
| 5 | LTCUSDT | GOOGL |
| 6 | MATICUSDT | META |
| 7 | NEARUSDT | MSFT |
| 8 | SHIBUSDT | NVDA |
| 9 | SUIUSDT | TSLA |
| 10 | XLMUSDT | TSM |
- Sequence length: 90 days
- Embedding dimension: 384
- Attention heads: 6
- Encoder layers: 4
- Dropout: 0.10
- Optimizer: AdamW
- Learning rate: 8e-5
- Weight decay: 5e-6
- Batch size: 512
- Max epochs: 6000
- Early stopping patience: 700
- Timestamp alignment across assets
- Forward/backward fill missing values
- Last 365 days reserved for testing
- 20% of pre-test data for validation
- MinMax scaling per feature (fit on training only)
- Gaussian noise injection during training
Computed on scaled space:
- Global MSE
- Global RMSE
- Global MAE
- Per-asset MSE / RMSE / MAE
Predictions are inverse-transformed before saving. Results are as follows:
For each scenario:
{scenario}_forecasts_no_patch_full_ohlcv2.csv{scenario}_no_patch_global_result2.csv{scenario}_no_patch_per_asset_results.csv2{scenario}_model_no_patch_best.pth
Recommended: Python 3.10+
pip install torch pandas numpy scikit-learn
python ChanFormer.py
The script automatically:
- Trains all three scenarios
- Saves best model per scenario
- Exports forecasts and metrics
- Prints final performance