Forecasting time series using convolutional neural network with multiplicative neuron

Abstract

Convolutional neural networks (CNNs) are proven to be efficient in time series forecasting, however architectural selection remains a challenging task. This work aims to propose CNN, utilizing single multiplicative neuron model in forecasting time series, intended to eliminate architectural complexities of classical CNN ensuring its computational efficiency. Applicability of proposed approach is employed on financial time series datasets such as Index, Stocks, Cryptocurrencies and a commodity in evaluating the model’s performance on the basis of RMSE, MAE and $R^2$ values. Further, time-delay effects were also observed in datasets which has been analyzed to improve the accuracy of the proposed model. Based on the lowest RMSE and MAE values, and higher $R^2$ values, the optimal delay value has been analyzed which has been used for forecasting. The result demonstrates that in data sets like NIFTY50, SBI, Bitcoin, and Natural Gas, the forecasting efficiency is improved when compared to classical CNN. The results obtained can be used to draw valuable insights for decision making, which will enable future studies and facilitate easy adaptation in analyzing time series.