The Inclusion of the Volume-Price-Product Factor for the Trend Forecasting of Futures Time Series Data

Abstract

Predicting time series data involves extracting features and forecasting trends from observed phenomena. Although deep learning algorithms are widely used in this field, their emphasis on prediction accuracy may not be optimal for futures time series data. For a futures time series, achieving high prediction accuracy alone is not sufficient. This is because, in some cases, ten accurate predictions may not compensate for a single loss. Therefore, a high accuracy rate does not necessarily translate into good returns. Existing methods have yet to provide practical and reliable approaches for predicting futures time series data. The primary contributions of this study are as follows: First, we employ the Vapnik-Chervonenkis (VC) dimension and error function from the perspective of binary classification for futures time series data to elucidate the generalization ability of the simple moving average model. Furthermore, we offer theoretical guidance to enhance predictive performance by introducing effective factors (i.e., features) that positively impact prediction results. By incorporating influential features, the discrimination of the loss function can be enhanced, making it easier to adjust the parameters and minimize the overall loss function value. Consequently, this improves the overall return rate, which is achieved by introducing additional factors to minimize the error values in the loss function. This explains why the proposed moving average model, enhanced by the introduction of the volume-price-product factor, achieves good prediction performance.

Keywords

• Time series analysis
• deep learning
• average models
• prediction methods
• quantization signal trading