Stock price prediction using probabilistic machine learning modelsOcepek, David (Avtor) Demšar, Jure (Mentor) stock price forecastingchance-constrained optimizationrisk-estimationstock tradingneural networksOur thesis had three main goals. The first one was to create a neural network-based stock portfolio trading agent for predicting future stock price distributions. This allows traders not only to estimate future trends for a given stock but also to predict risks associated with the stock. The second goal was to improve this agent's profitability by training it with custom loss functions, specifically designed for the nature of stock forecasting problems. The third goal was to create a realistic brokerage environment for running trading simulations in order to evaluate the success of our models. We developed several neural network-based models and compared them with more traditional time series forecasting approaches. For estimating uncertainty we used Markov chain dropout or the quantile loss. To further improve our results, agents facilitated two kinds of portfolio optimization techniques, the expected return maximization, which did not use uncertainty in its decision-making, and the chance-constrained optimization, which did. Based on our experiments, we conclude that trading commissions strongly affect agent profitability. The most profitable models were the temporal fusion transformer and a simple neural network, resulting in gains of 65% for US residents and 19% for non-US residents. The custom loss function did not yield the desired results and performed worse than mean square error. In terms of optimizers, chance-constrained optimization increases the profitability of agents but is riskier as it is unable to guarantee a certain level of risk to the user.20232023-12-06 09:55:00Magistrsko delo/naloga152756VisID: 35336COBISS_ID: 177588483sl