Trading by estimating the quantized forward distribution

Abstract

ABSTRACTIn this article, a novel algorithm is developed for electronic trading on financial time series. The new method uses quantization and volatility information together with feedforward neural networks for achieving high-frequency trading (HFT). The proposed procedures are based on estimating the Forward Conditional Probability Distribution (FCPD) of the quantized return values. From past samples, the conditional expected value can be learned, from which FCPD can be obtained by using a special encoding scheme. Based on this estimation, a trading signal is triggered if the probability of price change becomes significant as measured by a quadratic criterion. Due to the encoding scheme and quantization, the complexity of learning and estimation has been reduced for HFT. Extensive numerical analysis has been performed on financial time series and the new method has proven to be profitable on mid-prices. In order to beat the secondary effects, we focus on the most liquid assets, on which we managed to achieve positive profits.