Forecasting Assist Movement via GAN

Project Overview

This paper focuses on using GANs combined with LSTM and CNN models for advanced stock market prediction. I incorporated financial data, technical indicators, and sentiment analysis to build a highly accurate predictive model.

Download paper View on GitHub

Key Achievements

• Implementation of a predictive model using Generative Adversarial Networks (GANs) combined with Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs) to forecast stock market movements
• Incorporated a broad spectrum of 72 assets, including major banks, global economy indices, the VIX, key composite indices, and a diverse range of currency pairs
• Engineered a robust set of features including technical indicators (e.g., moving averages, MACD), Fourier transforms for trend extraction, ARIMA for time-series modeling, and advanced sentiment analysis using Natural Language Processing (NLP) on financial news
• Integrated anomaly detection techniques to identify market irregularities, ensuring the model could handle unexpected market behavior
• Applied Deep Reinforcement Learning strategies (PPO and Rainbow) and Bayesian optimization to dynamically fine-tune the hyperparameters of the GAN model, improving prediction accuracy
• Conducted rigorous validation of the model using real-world data from high-profile stocks such as Apple (AAPL), Microsoft (MSFT), and Amazon (AMZN), achieving Mean Absolute Errors (MAE) as low as 0.0077 and MAPE values under 1.82%
• Successfully conducted ten episodes of reinforcement learning, each with 200 epochs, to refine the GAN's predictive capability

Tools and Technologies Used

• Python, TensorFlow, Keras
• GAN, LSTM, CNN
• Reinforcement Learning (PPO, Rainbow)