Seize the (intra)day: Features selection and rules extraction for tradings on high-frequency data

Abstract

We describe a method to develop trading rules based on the responses of self-organizing maps (SOMs), trained under various distance metrics. The effectiveness of the procedure is examined on 5 min data of S&P MIB financial index, and its performances compared with those of classical buy and hold trading technique. The noticeable innovations of the paper include the methodology itself, which brings SOMs into a decision making tool to operate into the market; the focus on intraday tradings bars; the systematic study of how changes in the distance metrics employed to train the maps may affect the overall performances, and, finally, the discussion of system performances both in the absence and in the presence of commission fees. At the current stage the results, evaluated with both financial and statistical indicators, bring us to the following conclusions: (a) self-organizing maps can be helpful to localize profitable intraday patterns, achieving more stable performances than common trading rules; (b) working with proper metrics may enhance the overall performances; (c) trading strategies based on unsupervised neural networks make exploitable with profits almost continuous trades, until commission fees maintain below suitable thresholds.