Federated learning in stock predictions

Abstract

Stock market prediction is about learning the future value of a particular stock and it can give a better yield in terms of profit if predicted accurately. Investors utilise various analysis tools to create stock market forecasts to profit from it. However, factors such as news, political concerns, and natural disasters make stocks difficult to anticipate, thus applying analysis tools will only enhance the possibilities of profiting in the market. Generally, there are three types of stock prediction methodologies: technical analysis, sentimental analysis, and fundamental analysis. There are numerous different machine learning architectures that can evaluate the accuracy of predicting stocks, but can a decentralised stock prediction with a Federated Learning Mechanism be comparable to a conventional centralised setup in Forecasting performance? In this investigation, we proposed a federated learning mechanism, resolving the challenges by allowing users to store the data locally, and learning a shared model by aggregating locally computed updates. We used Long Short-Term Memory (LSTM) which is a type of recurrent neural network to compare the accuracy of the model between federated machine learning and normal machine learning methods. The experimental results show that regardless of model parameter adjustments, the traditional LSTM method outperforms the federated LSTM method in accuracy on stock prediction.