SXNP-Factor-Models

Feature Engineering for Fundamental Factor Models

This project explores the effectiveness of certain recurrent neural network architectures with long short-term memory units (LSTM RNNs) in the forecasting of stock prices for companies in the STOXX Europe 600 Industrial Goods and Services Index. In particular, the project focused on attempting to identify whether certain attributes of a company’s fundamentals data, such as book value and annual sales, had significant influences on its stock price. The main emphasis of this project was to implement a complete tool that would provide all the functionalities required to take this machine learning problem from the initial raw dataset to a reasonable final forecast result. A significant proportion of the project focused on processing the raw data into a suitable and effective format that could be used as the input of a Keras model. Additionally, by tuning various hyperparameters and experimenting with multiple loss functions and classifications methods, attempts were also made to optimize and improve the performance of the model. Finally, the stock weightings provided by the model were used to construct a portfolio to simulate a real world scenario. Its performance was then compared to the naïve diversification heuristic over a time period consisting of unseen test data.