This paper develops a statistical arbitrage strategy based on overnight social media data and applies it to high-frequency data of the S&P 500 constituents from January 2014 to December 2015. The established trading framework predicts future financial markets using Factorization Machines, which represent a state-of-the-art algorithm coping with high-dimensional data in very sparse settings. Essentially, we implement and analyze the effectiveness of support vector machines (SVM), second-order Factorization Machines (SFM), third-order Factorization Machines (TFM), and adaptive-order FactorizationMachines (AFM). In the back-testing study, we prove the efficiency of Factorization Machines in general and show that increasing complexity of Factorization Machines provokes higher profitability - annualized returns after transaction costs vary between 5.96 percent for SVM and 13.52 percent for AFM, compared to 5.63 percent of a naive buy-and-hold strategy of the S&P 500 index. The corresponding Sharpe ratios range between 1.00 for SVM and 2.15 for AFM. Varying profitability during the opening minutes can be explained by the effects of market efficiency and trading turmoils. Additionally, the AFM approach achieves the highest accuracy rate and generates statistically and economically remarkable returns after transaction costs without loading on any systematic risk exposure.
finance social media data Factorization Machine overnight information statistical arbitrage high-frequency trading