Forecasting under uncertainty : combining and evaluating predictive models in economics and finance
(2023)
Files
ThesisRoccazzella.pdf
Open Access - Adobe PDF
- 4.56 MB
Details
- Authors
- Supervisors
- Vrins, Frédéric
- Abstract
- In recent years, the field of forecasting has experienced significant growth, thanks to advancements in statistical techniques, increased data availability, and the use of open-source predictive models. While there are methods to quantify the risk or statistical uncertainty of a predictive framework, decision-makers also face Knightian uncertainty due to incomplete knowledge of the underlying data-generating process or the likelihood of unknown external shocks. When the decision maker is asked to form an expectation of a variable of interest and multiple forecasts are available, we show that optimal and robust combinations of forecasts offer superior forecast performance. These techniques enforce a non-negativity constraint on the weights and use penalties (or shrinkage) to penalize the divergence from a reference combination scheme. Their benefits include performing forecast selection and combination in a single step while also mitigating the risk of over-fitting by shrinking the optimal but potentially unstable solution to a sub-optimal but satisfactory one. Combinations provide useful insights also when evaluating competing forecasting strategies. We also use optimal combinations to introduce a novel forecast encompassing test under constrained parameter space. By explicitly considering that weights are constrained in the unit simplex, the proposed test has adequate size properties and is more powerful than a competitive test that ignores such constraints. We apply these newly developed combination and evaluation techniques to tackle operational challenges in the credit industry and when evaluating monetary policy in the euro area. Model risk presents a significant challenge for both practitioners and regulators in the credit industry, as there is a risk of a mismatch between the expected and actual results of a model. Decisions concerning model risk include selecting the most appropriate modeling framework, and predictors to employ. We find that combining models that are built on security-specific characteristics and a limited number of established recovery rate determinants permits a favorable trade-off between accurate predictions and interpretability. We also stress the relevance of incorporating economic uncertainty measures and economic predictors in designing recovery rate forecast methods for corporate bonds and consumer credit. In some circumstances, however, the decision maker is confronted with the likelihood of a new event that, despite impacting the target of interest, is overlooked by the available models. In this case, uncertainty arises from the unavailability of supported forecasts. We show how drawing comparisons between the new event and past experiences can aid decision-making. For instance, we design a model that describes the cross-border shock transmission mechanism in the European sovereign bond market during the Eurozone debt crisis. This helps us to evaluate the risk of a potential second crisis during the COVID pandemic. While the proposed methods are tailored to financial and economic forecasting, this thesis offers guidance to the broader forecasting community when dealing with uncertainty.
- Affiliations
Citations
Roccazzella, F. (2023). Forecasting under uncertainty : combining and evaluating predictive models in economics and finance. https://hdl.handle.net/2078.5/233562