Forecasting Macrofinancial Time Series with Machine Learning Methods

Unlike the standard supervised learning setting, forecasting time series poses different challenges due to their inherent time dependence, non-stationarity, and susceptibility to structural shifts. Traditional econometric models, while powerful, might struggle capturing the nonlinear patterns and high-dimensional dependencies present in such data. In recent years, machine learning (ML) methods have emerged as promising alternatives, offering flexible approaches to modeling and prediction. In this thesis, we (1) build the foundational knowledge related to timeseries as well as practical applications , (2) introduce the basic building blocks of both classical and ML algorithms and finally (3) compare these models in multiple different settings - low to high amount of training data with low to high forecasting horizons. Finally, all tests are conducted on 3 different real world datasets to allow us to generalize our findings. We find that ensemble methods perform better across majority of settings, with the exception of low signal-to-noise datasets, where simpler methods perform better. Additionally, regularized models tend to perform at least as good as its non-regularized counterparts, though at the cost of computation.