AWDL-Net: Adaptive wavelet-based deep learning model for multi-scale financial time series forecasting

Abstract

Forecasting financial time series is an extremely difficult task due to the many complexities and characteristics associated with financial time series, especially in emerging markets: nonlinearity, nonstationarity, volatility clustering, and multi-scale dynamics. To help overcome these issues, this study develops a new framework to forecast stock price market returns called the Adaptive Wavelet-Based Deep Learning Network (AWDL-Net). The AWDL-Net framework utilizes both wavelet-based multi-resolution analysis (including the Discrete Wavelet Transform) and adaptive deep learning to improve prediction accuracy. The AWDL-Net framework decomposes a given stock price series into multiple frequency components (via a wavelet-based multi-resolution analysis) so that both short and long-term trends in the data can be extracted at different frequency bands through a series of parallel deep learning learners. Additionally, the predictions made by the various parallel deep learning models will be combined through an adaptive data-driven fusion mechanism (making use of each model's performance) in order to generate a single composite forecast for the stock price series. The framework has been tested on pre-COVID stock prices for major commercial banks operating in Nigeria, and AWDL-NET has outperformed all standalone predictors such as RNN, LSTM, and 1DCNN across each of these banking institutions. For example, AWDL-Net demonstrated reduced error rates than traditional methods for Fidelity Bank (MSE: 0.0353, RMSE: 0.1880, MAE: 0.1562), Stanbic IBTC (MSE: 4.7250, RMSE: 2.1737, MAE: 1.7346) and UBA (MSE: 0.2295, RMSE: 0.4791, MAE: 0.3911). AWDL-Net also outperformed traditional methods for Wema Bank (MSE: 0.0091, RMSE: 0.0954, MAE: 0.0748), Zenith Bank (MSE: 0.8582, RMSE: 0.9264, MAE: 0.7753), GTBank (MSE: 2.4172, RMSE: 1.5547, MAE: 1.3441), and First Bank (MSE: 0.3078, RMSE: 0.5548, MAE: 0.4124). In terms of overall performance across all banks, AWDL-Net had the highest number of banks with the lowest error rates (MSE: 0.0447, RMSE: 0.2115, MAE: 0.1694), thereby substantiating its ability to accurately forecast future market conditions under various economic conditions. This study provides evidence for the effectiveness of using multiscale signal decomposition and adaptive model fusion to model complex nonlinear and temporal dependencies and contributes to the existing literature by presenting a fully adaptive, architecture-agnostic deep learning framework for financial forecasting. The proposed model can serve as a useful tool for both investment analysis and risk management and can also allow investors in emerging market economies plagued by macroeconomic uncertainty to assess the viability of their current investment decisions.