A Novel Model to Minimize Time Series Forecasting Error using Convolutional Attention

Abstract

The main consideration of this paper is time series forecasting using machine learning methods. While hundreds of studies have been conducted on this topic, we propose a novel, yet intuitive model, using technical indicators in combination with a focus on lagged relationships and the machine learning technique called BEDCA (Belief Encoder-Decoder with Convolutional Attention). BEDCA takes into consideration technical indicators to improve predictive accuracy. To evaluate our model, we used historical Coca-Cola stock data, room temperature data, and daily data including the number of COVID-19 case data. We found that our model trained relatively quickly and had a high accuracy (low loss). Overall, BEDCA outperformed the standard Long-Short Term Memory (LSTM) network on all three datasets, achieving MAE reduction of 74.1%, 41.3%, and 69.0%, respectively, although it took 3.05, 3.26, and 3.20 times longer to train on each dataset respectively. Our results highlight that convolutional attention is a promising form of attention and that technical indicators are important considerations for time series data.