Enhancing Financial Time-Series Analysis with TimeGAN: A Novel Approach

Stock price prediction is crucial in the financial sector, impacted by various factors such as economic indicators, news events, and investor sentiment. Traditional time-series prediction models like Long Short-Term Memory (LSTM) and Autoregressive Integrated Moving Average (ARIMA) have been widely used but have limitations. This study explores the application of Time-series Generative Adversarial Networks (TimeGAN) for generating synthetic time-series data that can closely mirror real financial d ata. T imeGAN c ombines t he s trengths o f both supervised and unsupervised learning to maintain temporal dynamics and preserve the relationships between variables over time. Using a Google stock price dataset, we evaluate the performance of TimeGAN-generated synthetic data against real data across key metrics such as R², Mean Absolute Error (MAE), and Mean Squared Log Error (MSLE). The results show that the synthetic data achieved an R2 score of 0.902, an MAE of 0.276, and an MSLE of 0.051, while the real data had an R2 score of 0.956, an MAE of 0.280, and an MSLE of 0.051. Visual analyses using Principal Component Analysis (PCA) and t-distributed Stochastic Neighbor Embedding (t-SNE) further validate that the synthetic data retains the diversity and distribution of the original dataset. This study demonstrates TimeGAN's potential to enhance financial forecasting by providing a reliable and realistic representation of stock price data. Future research can expand TimeGAN's application to other financial d atasets, o ffering robust tools for risk management and investment decision-making.