On DeepAR methods, its extensions, and implementation

Time series forecasting plays a critical role in various sectors such as in finance, but accurate predictions is important for making decisions.

Traditional probabilistic models such as the Vector Autoregression have been foundational, yet their effectiveness is limited by assumptions that do not hold in the unpredictable financial market. These models fail to accurately capture certain features inherent in financial data.

Non-probabilistic models, such as neural networks offer flexibility as having weaker assumptions, but they require extensive data for training, making them less effective in situations with limited historical information.

This doctoral study involves the development, extension, and application of methods that can effectively handle such complexities without compromising predictive accuracy or requiring unrealistic data assumptions.

We will work on DeepAR models, their extensions towards Bayes statistics and Gaussian processes, as well as we will develop appropriate computational methods to facilitate the corresponding challenges. The implementation of the developed methods may involve financial data.