ISEPT: Image-Based Selection and Execution Framework for Pair Trading

Pair trading is a statistical arbitrage strategy that seeks market-neutral profits by hedging two assets. Most earlier studies separate pair selection and execution, running each step in its own model. This split prevents trading results from informing selection and often overfits a small set of assets. In addition, many approaches rely on raw or aggregated prices, which struggle to capture the visual patterns and cross-asset interactions that influence returns. Accordingly, we introduce ISEPT, an end-to-end framework that works directly with candlestick chart images and joins pair selection with trading. A convolutional autoencoder (CAE) converts monthly candlestick images into stock-level latent vectors. For each pair, the concatenated vectors pass to a multilayer perceptron (MLP) that predicts the next month's Sharpe ratio. At month-end, realized trading results feed back into the MLP as fresh training data, letting the model adjust to current market conditions and refine pair rankings continuously. Tests on a long historical span of daily Open-High-Low-Close (OHLC) data for S&P 500 constituents show that ISEPT outperforms traditional approaches. Both return on investment and Sharpe ratio improve markedly, indicating that the framework delivers steady long-term gains.