A hybrid model of Multivariate Bayesian Time Series Model with Multi-TemporalConvolution Network for Forecasting Stock Market during Covid-19 Pandemic

This paper presents an alternative method for fore- casting multiple correlated time series data. The proposed hybrid algorithm has two steps comprising a base step of fitting a multivariate Bayesian structural time series (MBSTS) and a post model fitting diagnostic step of running the residuals through a multi-input/output temporal convolutional network (M-TCN) with multiple time scale feature learning. The first step allows for the inclusion of a possibly high dimensional regression component and spike and slab priors are used to find a parsimonious model. The second step is an alternative to the subjective residual based diagnostic procedures in the time series literature and enhances forecasting accuracy. The main advantage of M-TCN is that it can capture sequential information by expanding the field of convolution kernel without increasing the number of parameters. Two applications are presented in this paper- the first uses pre-lockdown data on eleven Nifty stock sectoral indices to forecast the trend of stock price movement and also the first post lockdown upturn. The second application uses stock market data of pharmaceutical companies manufacturing COVID-19 vaccines. In both cases, newspaper and social media sentiment serve as the regression component. The proposed hybrid model is shown to outperform bench- mark models for the two motivating datasets. The hybrid model produces greater accuracy than other models such as LSTM (Long short-term memory), Bidirectional Encoder Representa- tions from Transformers (BERT) based LSTM and GRU (Gated recurrent units), etc.