Self-Attentive Moving Average for Time Series Prediction

Time series prediction has been studied for decades due to its potential in a wide range of applications. As one the most popular technical indicators, moving average summarizes overall changing patterns over past period and is frequently used predict future trend time series. However, traditional indicators are calculated by averaging data with equal or predefined weights, ignore subtle difference importance different steps. Moreover, unchanged weights will be applied across series, regardless differences their inherent characteristics. In addition, interaction between dimensions ignored when using averages scales trends. this paper, we propose learning-based indicator, called self-attentive (SAMA). After encoding input signals based on recurrent neural networks, introduce self-attention mechanism adaptively determine at steps calculating average. Furthermore, use multiple heads model SAMA scales, finally combine them through bilinear fusion network prediction. Extensive experiments two real-world datasets demonstrate effectiveness our approach. The codes work have released.