Time series forecasting has been a standard task of time series analysis. Until recently, purely neural methods for time series forecasting were considered sub-optimal in comparison to well known statistical methods (like ARIMA or Vector AutoRegression) or hybrid ones. However, recently an avalanche of deep learning methods for time series forecasting has appeared, claiming to outperform previous state-of-the-art methods.
For this thesis, the candidate will develop a benchmark suite for time series forecasting with various methods, such as rolling mean, ARIMA, LSTM, Temporal Fusion Transformer, DeepAR, NBeats, Graph Neural Networks, Moirai/TimeGPT. The methods will be compared for their accuracy and performance in the task of uni- and multi-variate forecasting for multiple horizons.
Bibliography:
Deliverable (besides the thesis itself): A Python library (possibly using other libraries) running and comparing multiple time series forecasting methods. The process should be highly automated.