Predicting the value of financial flows is one of the most challenging machine learning problems, as historical data on value movements contain a lot of noise and are unstable. Current approaches focus on using reinforcement learning and using neural networks to predict the next action. In this thesis, we will explore the use of recurrent neural networks to perform these functions and compare them to each other. We will compare the performance of a neural network with LSTM layers, which have already proven to be very promising both in the field of predicting financial flows and playing games, and a neural network with GRU layers, the use of which has not yet been well explored. By using feed-forward neural networks, we want to better capture temporal dependencies and patterns in the data, and it will also allow us to use a larger amount of data (from past time intervals) to predict the next action to achieve greater accuracy. The results of our experiments show that, in the period between 2019 and 2024, the agent with a feedforward neural network performs best on most financial instruments, while the agent with GRU layers performs best on financial instruments with large and sudden price changes.
|
