A1197
Title: Constrained machine learning and stacking for real-time survey forecast combination
Authors: Adam Csapai - The Institute of Economic Research of Slovak Academy of Sciences and University of Economics in Bratislava (Slovakia) [presenting]
Abstract: Survey-based forecast combinations are often evaluated using simple averages, as more flexible methods tend to overfit in small real-time samples. The aim is to examine whether machine learning and stacking can improve the combination of professional survey forecasts when estimation is tightly constrained. Monthly forecasts from Slovak commercial banks for current-year GDP growth and inflation over 2007-2024 are combined using a strictly real-time rolling design that accounts for data publication lags. Linear shrinkage methods, non-linear machine learning models, and non-negative stacking based on out-of-sample base predictions are compared to the equal-weight benchmark. For inflation, regularised linear combinations deliver stable gains, reducing mean squared forecast errors by 10-20 percent, while non-linear methods systematically underperform. For GDP growth, several non-linear models and restricted stacking schemes yield large improvements, with error reductions of up to 40 percent relative to the benchmark. The results demonstrate that the usefulness of machine learning in survey forecast combination is target-specific and depends critically on real-time constraints, providing guidance on when complexity adds value in small panels of professional forecasters.
