From pixels to profits: trading arbitrage portfolios based on image representations

This paper studies whether convolutional neural networks (CNNs) can extract economically useful signals from image representations of financial time series to form statistical arbitrage portfolios. Following a generalized pairs-trading framework, I encode the joint dynamics of each stock, its factor-based hedge, and the residual spread as RGB images. A CNN ranks stocks by the probability of above-median residual returns in the following month. The strongest baseline specification generates a gross monthly excess return of 0.82% and an annualized Sharpe ratio of 0.97. Hedging improves gross risk-adjusted performance, while comparisons with technical-indicator strategies suggest that conventional predictors capture part of the image-based performance. The image-based portfolio retains a positive alpha after controlling for these benchmarks. Overall, the findings suggest that CNNs can extract economically relevant patterns from image representations, although profitability depends on beta estimation and trading costs.