Abstract: The present research introduces an innovative approach to macroeconomic forecasting based on both existing machine learning techniques and original contributions to the field. We created a forecasting algorithm – dubbed Adaptive-GBT with Predictive Intrapolation (PI) – that is specifically tailored for macroeconomic forecasting, and addresses non-linearity and non-ergodicity of the economy, as well as the problem of high dimensionality in a context of scarce data. We produced forecasting simulations in pseudo-real time for all major economies. Our forecasting algorithm proved more reliable than benchmark forecasts and displayed a much better ability to anticipate turning points. Moreover, Adaptive-GBT with PI allows producing reliable forecasts nine months and twelve months ahead.
Bio: Nicolas Woloszko is a data scientist and an economist. He joined the OECD in September 2016, where he initiated an innovative project that aims at bridging the gap between machine learning and economics. Nicolas has been introducing artificial intelligence in economic analysis in order to make better economic forecasts, and to produce more reliable assessment of the impact of economic policies. His project involves research in both machine leaning and economics. Prior to joining the OECD, Nicolas has worked in research (at Sciences Po), and in a machine-learning start-up (Kynapse). He has a dual background in economics and applied maths from ENS and ENSAE ParisTech.