Forecasting Inflation Using Machine Learning vs. Traditional Time Series Models: Evidence from the Post-Pandemic Period

The post-2021 inflationary episode renewed a long-standing question in applied macroeconomic forecasting: whether flexible machine-learning estimators predict inflation more accurately than classical time-series models, and whether any such advantage is concentrated in periods of structural instability rather than being general. This paper compares classical econometric benchmarks with machine-learning estimators over several forecast horizons, using a recursive pseudo-out-of-sample design in which every model setting is determined exclusively from information available before each forecast origin, so that no specification is chosen with knowledge of the period being predicted. Simple naive benchmarks are included alongside the econometric ones, so that any reported gain is measured against a realistic lower bound rather than against a weak comparator. To distinguish the contribution of the model class from that of the volume of information supplied to it, the machine-learning estimators are evaluated twice: once under the same restricted information set available to the classical multivariate benchmark, and once under a substantially broader set of macroeconomic predictors. Forecast accuracy is assessed by standard loss measures and by tests of equal predictive accuracy appropriate to overlapping multi-step forecast errors, and is compared across a period of low and stable inflation and the post-pandemic cycle, the latter extending through the subsequent disinflation. The objective is to establish the conditions under which the additional flexibility of machine-learning methods, and the additional information they can accommodate, translate into practical forecasting gains.