Forecasting stock price movements is a complex problem due to the non-stationarity and high level of noise in financial time series. This thesis systematically evaluates the effectiveness of modern foundation models based on the Transformer architecture for predicting short-term price movements. In a two-phase experiment, the Chronos and TimesFM models were compared on intraday data from eleven volatile stocks at 5-minute, 15-minute, and 1-hour intervals. Based on the results, the Chronos model was selected for a more detailed analysis at a 5-minute interval, evaluated in both zero-shot and fine-tuned modes, and compared against the ARIMA and random walk baseline models. The results show that the fine-tuned Chronos model proves to be the most stable, exhibiting the narrowest error distribution, although its median error does not always surpass that of the strong random walk baseline. Nevertheless, qualitative analysis demonstrates that the model is exceptionally successful at capturing clear trends during periods of lower volatility, where it achieves high directional accuracy, while, as expected, failing to predict extreme, unpredictable events. This work confirms the potential of the Transformer architecture, which offers more consistent predictions than classical approaches, and indicates the existence of predictable patterns under certain market conditions.
|
