Targeted transformations for macroeconomic forecasting

The crisis periods of the past decades have highlighted the difficulty of forecasting economic indicators due to increased non-linearity and rapidly changing dynamics. To address this challenge, we introduce the Transform-Sparsify-Forecast (TSF) framework. The TSF framework first applies multiple transformations to each predictor to account for non-linear effects. It then dynamically applies a dimension reduction (targeting) to select the most relevant predictors from the high-dimensional dataset. Our approach is straightforward and can be seamlessly incorporated into existing forecasting frameworks. We demonstrate its versatility by applying it to five distinct classes of econometric and machine learning models. Using the FREDMD dataset, we show that the TSF framework substantially improves the forecasting accuracy across all models, with notable gains at short forecast horizons and during periods of high economic uncertainties, attributable to the adaptive realignment of variable interactions and the ability to capture non-linearities. On short horizons, improvements are on average more than twice as high compared to standard methods without transformations. Our findings highlight the importance of the dataset composition and variable selection in effectively capturing evolving relationships and dynamics over time.