Modelling of stock market volatility using hybrid wavelet transform data preprocessing and artificial neural network in the Kenyan securities market

Abstract

Financial time-series modelling is a complex task, and methods such as artificial neural networks have been used for over the past two decades. Deep neural networks have shown to be more efficient in many domains, including physics, engineering, biomedicine, signal processing, mathematics, and statistics. This study proposes a model that uses deep learning technique combined with discrete wavelet data preprocessing to improve stock volatility forecasting in the Kenyan frontier market. The study discusses the role of wavelet transforms in time series analysis and demonstrates their advantages in general and in time series denoising, using a sample of 4 stocks of time series data from the Nairobi Securities Exchange (NSE). The forecasting model proposed for financial time series compares the performance of a Discrete Wavelet Transform Long Short Term Memory and standard Long Short Term Memory. The financial time series data is decomposed using the Discrete Wavelet Transform, and the resulting approximation and detail coefficients are used as input variables in a Long Short-Term Memory neural network to predict future stock returns. This study is built on the Python scripting environment and use TensorFlow as a system for deep learning for training, prediction, and comparison. In order to establish the efficacy and superiority of the recommended hybrid model, the forecasting performance of the analyzed models is evaluated using four metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Bayesian Information Criterion (BIC), and Akaike Information Criterion (AIC). According to the empirical findings the DWT-LSTM models for all data sets used were better at predicting stock volatility than the standard LSTM models, with lower performance metrics values.

Keywords: Wavelet Transform, Long Short-Term Memory, Optimizers, Stock volatility.