B1696
Title: Probabilistic forecasting of binary outcomes in the presence of outliers
Authors: Mikhail Zhelonkin - Erasmus University Rotterdam (Netherlands) [presenting]
Abstract: The problem of forecasting binary outcomes is of prominent importance in various fields including economics, management, finance and medicine, to mention a few. For instance, it can be a default of a company, a click on an online advertisement, or an occurrence of a disease. The traditional approach is to use classification methods, which can be seen as point forecasts. However, from the perspective of a decision-maker, it is valuable to have a probability forecast. The traditional benchmark parametric models, e.g., logistic regression, are unstable in the presence of outliers and data contamination. The alternative machine learning methods are often biased and require recalibration which makes them hardly interpretable. It is shown that logistic regression estimated by robust methods is a viable alternative. Using the influence functions approach, it is shown that the robustly fitted logistic regression delivers well-calibrated forecasts and that the additional variability is negligible.
