Market-driven Forecast Combination Trees
-
Series
-
Speaker(s)Ekaterina Kazak (University of Birmingham, United Kingdom)
-
FieldEconometrics, Data Science and Econometrics
-
LocationUniversity of Amsterdam, Roeterseilandcampus, room E5.07
Amsterdam -
Date and time
October 03, 2025
12:30 - 13:30
In this paper, we suggest a novel approach to combining forecasts, whereby the weights are modeled as a nonparametric function of market conditions. It is assumed that the optimal weights are identical across observations within a given regime and differ across regimes. The number of regimes and their duration are identified in a fully data-driven way using tailor-made machine learning algorithm. The new forecast combination method is applied to forecast downside market risk measures, demonstrating that localising the weights using bagged modified decision trees significantly enhances forecasting performance. The superior statistical and economic performance of the introduced forecast combination technique is illustrated through an application to daily returns of 30 large cap stocks.