IntraFormer – Channel-Sequential Transformer for Next-Day Equity Close Prediction

IntraFormer is a lightweight, channel-first transformer architecture designed specifically for next-day closing price prediction on daily equity data (OHLCV).

Instead of treating time steps as the sequence (common in most time-series transformers), IntraFormer treats the OHLCV channels as the sequence and projects the time dimension — a design choice that has shown surprisingly strong empirical performance on individual stock prediction tasks.

Key Features

• Channel-sequential Transformer encoder (no patching)
• Reversible Instance Normalization (RevIN) tailored for financial time series
• Optional positional embeddings on channels
• Ensemble of 3 independently trained models (different seeds)
• Ablation suite comparing RevIN, positional embeddings, layer count, dropout
• Realistic walk-forward backtesting with transaction costs (5 bps)
• Comparison scripts vs traditional deep models (LSTM/GRU/CNN-LSTM/CLAM)
• Carbon footprint estimation example (training + inference energy)

Project Structure

thequantscientist-intraformer/
├── README.md
├── LICENSE                  # MIT
└── src/
    ├── IntraFormer.py       # Main single-stock training & evaluation script
    ├── ablation.py          # Ablation study: RevIN / pos emb / layers / dropout
    ├── backtesting.py       # Portfolio-level walk-forward backtest
    ├── computation.py       # Energy & CO₂ footprint measurement example
    ├── traditional_benchmark.py   # LSTM / GRU / CNN-LSTM / CLAM baselines
    └── transformer_benchmark.py   # Autoformer / FEDformer / Informer / iTransformer
    └── ultimate_results.csv   # (example output – not versioned)

Quick Start

1. Requirements

pip install torch pandas numpy scikit-learn
# optional: for transformer benchmarks
pip install einops

2. Data Preparation

Place daily OHLCV CSV files in data/ with format:

Date,open,high,low,close,volume
2020-01-02,74.06,75.15,73.80,75.09,135480400
...

Example filename: AAPL_1d_full.csv

Update paths in the scripts if needed.

3. Run main experiment (30 stocks, ensemble of 3)

cd src
python IntraFormer.py

→ Produces predictions and metrics in ~/qa/IntraFormer_all_equity_*

4. Run ablation study

python ablation.py

→ Compares 5 configurations across all stocks

5. Run backtest

After generating predictions:

python backtesting.py

Shows long-only, top-N, and long-short portfolio performance with realistic costs.

Model Architecture (simplified)

Input:  (B, 60 days, 5 channels: OHLCV)

→ Reversible Instance Normalization (per sample)
→ Linear projection: 60 → d_model=256   (per channel)
→ Add learnable channel positional embedding   (optional)
→ TransformerEncoder (3 layers, 8 heads, GELU FFN)
→ Take close-channel representation (index 3)
→ MLP head → scalar next-day close prediction
→ Denormalize (if RevIN used)

Results Highlights (your data)

• Best ablation usually includes RevIN + positional embeddings + 3 layers
• IntraFormer frequently outperforms LSTM/GRU baselines by 10–30% in RMSE/MAE on many names
• Portfolio strategies show positive Sharpe even after transaction costs in several configurations

Citation

If you find this work useful in your research, feel free to cite:

@misc{intraformer2026,
  author = {Nguyen Quoc Anh},
  title  = {IntraFormer: Channel-Sequential Transformer for Daily Equity Close Prediction},
  year   = {2026},
  url    = {https://github.com/thequantscientist/intraformer}
}