This paper presents a comprehensive framework for statistical arbitrage, decomposing it into three key components: constructing arbitrage portfolios, extracting time-series signals, and determining optimal trading allocations. The authors propose a novel deep learning solution that uses conditional latent factor models (such as Instrumented PCA) to generate arbitrage portfolios as residuals from asset pricing factors, ensuring they are weakly correlated and stationary. For signal extraction, they employ a convolutional transformer, a state-of-the-art machine learning method for time-series pattern recognition, which learns flexible, data-driven filters without parametric assumptions. The trading allocation is optimized via neural networks to maximize risk-adjusted returns under realistic constraints like leverage and turnover limits.

Empirically, the model is tested on daily returns of the 550 largest US stocks from 1998 to 2016, achieving outstanding out-of-sample performance: annual Sharpe ratios exceeding 4 and mean returns of 20% even with short-selling constraints. This performance is four times better than leading parametric benchmarks and twice as good as alternative deep learning models without the convolutional transformer. The study reveals that signal extraction is the most critical element; the choice of factor model has limited impact as long as it includes sufficient factors. The convolutional transformer identifies local asymmetric trend and reversion patterns, with a short-term focus on the last 30 trading days. The profitability is orthogonal to common risk factors and persists after transaction costs, indicating substantial real-world arbitrage opportunities. Overall, the paper demonstrates that flexible, data-driven time-series modeling combined with economic portfolio construction can uncover significant statistical arbitrage in liquid equity markets.