Stock price prediction with wavelet transform and deep learning methods

Investors consistently exhibit a keen interest in high-yield investment instruments, directing their savings towards such tools. The anticipation of fluctuations and trends in the stock market remains a focus for both individual investors and investment firms. Given the substantial return potential associated with stocks, analysts and researchers also manifest significant interest in the field of stock price forecasting. Nevertheless, stock prices are intricately influenced by multifaceted factors, including the economic state of the country, political and social events, international relations, and sectoral developments. As a result, the ability to predict stock price movements poses a challenging task. Although stock price predictions traditionally rely on statistical and econometric methods grounded in time series analysis, recent years have witnessed an increasing adoption of artificial intelligence and deep learning methodologies. This trend is propelled by advancements in computer memory and data processing capabilities. However, the effectiveness of deep neural networks trained based on historical values is seriously affected by the noise common in financial data. Addressing this issue, the study observes that these networks struggle to accurately predict stock prices. To mitigate the impact of noise in financial variables, the study recommends the application of the wavelet transform method, a widely used technique in signal analysis. The primary objective of this study is to predict the closing values of stock data in different sectors using wavelet transform and deep learning methods, and to compare the performances of the models. To achieve this goal, closing price data of selected companies from various sectors were collected as the dependent variable between June 3, 2013, and May 28, 2021. Among the independent variables are the 20-day lagged values of closing prices for each stock, along with opening prices, lowest prices, highest prices, gram gold prices, Brent crude oil prices, and the exchange rate of the dollar. The analysis aims to evaluate the predictive capabilities of the models within this framework. In the data preprocessing stage, noise in financial data was alleviated through the application of wavelet transform. Subsequently, the denoised financial data were employed to train various deep learning models, including Artificial Neural Networks (ANN), Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU), for predicting stock closing prices. Additionally, these deep learning methods were applied to financial data without any transformation. Upon evaluating the results, it was observed that RNN, LSTM, and GRU models, particularly when trained with Symlets wavelet transform, demonstrated high efficacy in prediction. A total of 8960 models were created through various combinations. Notably, among these models, the denoised RNN model using Symlets wavelet transform for ASELS exhibited superior predictive performance, yielding an RMSE value of 0.0067.