Financial time series analysis with 3D convolutional neural networks

Artificial neural networks have become widely used in recent years for financial time series forecasting. Their ability to model complex, nonlinear relationships has allowed them to surpass traditional methods in many applications. Among these, LSTM-based models, which can effectively capture temporal dependencies, are extensively utilized for predicting financial data. Convolutional Neural Networks (CNNs) have also gained traction in the literature and have shown promising results in various studies. In this study, a novel CNN-based approach is proposed, extending beyond classical CNN implementations by performing convolution operations over a three-dimensional data structure. The data preprocessing phase was carefully designed. First, 20 commonly used technical analysis indicators were selected. Each indicator was calculated using 20 different time periods ranging from 5 to 25 days, resulting in a 20×20 two-dimensional matrix for each day. Subsequently, consecutive data from the past 20 days were added to this matrix, forming a three-dimensional input of size 20×20×20 for each day. This data structure enables the CNN model to simultaneously process information related to time, indicator types, and period lengths, thus allowing it to capture more complex patterns. In the labeling phase, each day was assigned one of three labels: "Buy", "Sell", or "Hold", based on the direction of the price movement. The model was trained to learn this classification task and to predict future trading signals. Based on the model's outputs, a trading strategy was implemented and tested over 750 trading days between 2022 and 2024, using both Dow Jones 30 stocks and selected exchange-traded funds (ETFs). As a result of the simulated trades, the strategy achieved an average annual return of 23.87% for stocks and 19.40% for ETFs, outperforming the traditional buy- and-hold strategy. These findings demonstrate that the proposed model can effectively leverage the multidimensional representation of technical indicators in time series data to achieve robust predictive performance.