Ensembles of Convolutional Neural Networks Trained Using Unconventional Data for Stock Predictions

This research paper summarizes a unique method of training neural networks that contrasts conventional methods used for assessing and making predictions with financial data. In general, for limited financial data sets such as those comprising only adjusted closing and volume information, neural networks are trained with raw, or structured, data. Relying on ongoing neural network development strategies and applications, we developed a neural network that could be trained using unconventional data in which prices or indicators could be converted into image data. We applied this method to predictions using daily data of the Apple stock price. Using extracted images, we were able to explore the capabilities of neural networks in greater depth as a result of using convolutional neural networks to capture more information from the images, such as edges and shape patterns, which is more advanced than traditional neural networks trained on raw data. We used a well-known method in data science and machine learning called “bootstrap aggregating” or “bagging.” After multiple experiments, we concluded that the performance metric registered decent to good values; however, we also encountered overfitting, which we plan to address in a future experiment. We found that the primary limitation of our method was overfitting, and in the next experiment, we will explore larger data sets, such as those comprising hourly data.