Time series forecasting plays a vital role across various domains, particularly in finance,
where accurate predictions of returns, volatility, interest rates, and economic indicators
are essential for asset management. However, financial time series are inherently com-
plex and exhibit characteristics such as regime switching, volatility clustering, and a high noise-to-signal ratio, making them difficult to forecast using traditional methods. In recent years, the exponential rise of Large Language Models (LLMs), such as Open AIs
ChatGPT, has reshaped the landscape of artificial intelligence. These models, powered by the Transformer architecture, have achieved remarkable results in tasks such as machine translation and text generation. This success has triggered significant interest across a variety of research fields to explore how the architectural backbone of Transformers can be applied to other domains, such as time series forecasting. This thesis, conducted in collaboration with Assenagon Asset Management S.A., aims to investigate the effectiveness of Transformer-based models for financial time series forecasting. In the first part of this thesis, we provide a comprehensive theoretical analysis of the Transformer architecture and its components, with a focus on the attention mechanism. We then introduce several Transformer variants that have been specifically modified for time series forecasting, such as the PatchTST and Crossformer models, as well as a brief discussion on the emerging field of LLM-based forecasters. The second part of the thesis focuses on the empirical evaluation of these models. We assess the forecasting accuracy of the Transformer-based models for prices, returns, and volatility, both as a whole and at the component level, and compare their performance with traditional statistical methods. Overall, the findings of this thesis, both theoretical and empirical, offer valuable insights into the effectiveness of Transformer-based models in financial time-series forecasting.
«
Time series forecasting plays a vital role across various domains, particularly in finance,
where accurate predictions of returns, volatility, interest rates, and economic indicators
are essential for asset management. However, financial time series are inherently com-
plex and exhibit characteristics such as regime switching, volatility clustering, and a high noise-to-signal ratio, making them difficult to forecast using traditional methods. In recent years, the exponential rise of Large Lan...
»