[ICAIF 2025] ISEPT: Image-Based Selection and Execution Framework for Pair Trading

Official implementation of “ISEPT: Image-Based Selection and Execution Framework for Pair Trading (ICAIF ’25)”
Paper: https://doi.org/10.1145/3768292.3770346

Pair trading aims to earn market-neutral profits by hedging two assets and exploiting temporary mispricings. However, many prior approaches separate pair selection and execution, which prevents trading outcomes from improving the next round of selection and can lead to overfitting. Moreover, using only raw/aggregated price series often fails to capture the visual patterns traders commonly rely on.

We propose ISEPT, an end-to-end framework that directly uses candlestick chart images and links pair selection ↔ trading through a Sharpe-ratio feedback loop. A Convolutional Autoencoder (CAE) encodes monthly candlestick images into stock-level latent vectors, and an MLP predicts the next-month Sharpe ratio for each candidate pair. At month-end, realized trading results are fed back as new training data so the model continuously adapts to market regime changes.

---

Key Highlights

• Image-based representation: converts OHLC data into candlestick images to capture within-asset + cross-asset patterns.
• Unified selection & execution loop: uses realized Sharpe ratios to retrain and refine pair rankings each month.
• Long-horizon evaluation: tested on S&P 500 constituents with out-of-sample evaluation spanning ~20 years (2004–2024).

---

Method Overview (ISEPT)

(1) Candlestick Image Generation

• Slide a 21-trading-day window (≈ 1 month) across the prior 12 months and render each candlestick images of size 64×64.
• Apply log scaling to reduce price-scale distortions across tickers.

(2) Pair Selection via CAE → MLP

• CAE encodes candlestick images into latent vectors.
• For each stock, average embeddings over T = 12 months, then concatenate the two stocks’ vectors to form a pair representation.
• MLP predicts the pair’s next-month Sharpe ratio, ranking all candidate pairs and selecting the Top-100 for trading.

(3) Sharpe-Ratio Feedback Loop (Adaptive Learning)

• At each month-end, compute realized performance and feed top/bottom pairs’ Sharpe ratios back into the MLP training set.
• This closes the loop: selection → trading → realized Sharpe → retraining → improved selection.

---

Experiments

Dataset

• Daily OHLC for S&P 500 constituents, preprocessed from Jan 1990–Dec 2024.
• Out-of-sample evaluation period: Jan 2004–Jun 2024.

Baselines

• GATEV (distance/SSD) and VIDYAMURTHY (correlation + Engle–Granger cointegration) as classical pair-selection/trading baselines.

Main Results (2004–2024)

ISEPT-based strategies substantially improve ROI and risk-adjusted performance vs. classical methods (see Table 1 in the paper).

---

Citation

If you use this code in your research, please cite:

@inproceedings{kim2025isept,
  title={ISEPT: Image-Based Selection and Execution Framework for Pair Trading},
  author={Kim, Nayoung and Lee, Jangwook and Kang, Yuncheol},
  booktitle={Proceedings of the 6th ACM International Conference on AI in Finance},
  pages={413--421},
  year={2025}
}