The present paper develops Adaptive Trees, a new machine learning approach specifically designed for economic forecasting. Economic forecasting is made difficult by economic complexity, which implies non-linearities (multiple interactions and discontinuities) and unknown structural changes (the continuous change in the distribution of economic variables). The forecast methodology aims at addressing these challenges. The algorithm is said to be “adaptive” insofar as it adapts to the quantity of structural change it detects in the economy by giving more weight to more recent observations. The performance of the algorithm in forecasting GDP growth 3- to 12-months ahead is assessed through simulations in pseudo-real-time for six major economies (USA, UK, Germany, France, Japan, Italy). The performance of Adaptive Trees is on average broadly similar to forecasts obtained from the OECD’s Indicator Model and generally performs better than a simple AR(1) benchmark model as well as Random Forests and Gradient Boosted Trees.
