Answers are grounded in this paper and cite their sources.
Machine learning models are often reported to forecast inflation well, yet nine standard models tested on Ghana's headline inflation through the 2022–23 surge, were all beaten by an ARIMA benchmark and a random walk. Here we ask why that happened and what can be done about it. We show that much of the failure comes from how the target is parameterized. A learner asked to predict the level of a highly persistent series has to rebuild that persistence from a short sample, and tree-based learners cannot forecast above the highest value they saw in training (41.95 percent here, against an actual peak of 54.1 percent). We propose persistence-anchored residual learning (PARL). An autoregressive anchor carries the persistent component, machine learning models are trained only on the anchor's one-step errors, and an adaptive weight, whose memory can be tied to a two-state Markov-switching regime model, decides how much of the correction to use. Prediction intervals come from locally scaled adaptive conformal inference. Using 280 months of Bank of Ghana data (January 2000 to April 2023), strictly one-step-ahead forecasts and a 42-month test window that contains the surge, we find that predicting the monthly change instead of the level cuts the test RMSE of the nonlinear learners by 64 to 80 percent, to values near that of the ARIMA anchor (2.15 percentage points). PARL lowers the RMSE by a further 3 percent (2.08), a gain that is not statistically significant against the anchor (Diebold–Mariano p = 0.41) but is against the random walk ( p = 0.03). Letting the weight's memory depend on the regime gave no additional gain. Adaptive conformal intervals raised the coverage of nominal 90 percent intervals in the test window from 67 to 86 percent and cut the interval score by about a quarter. The results suggest that, for persistent macroeconomic series, how the target is written matters more than which learner is used.
No takes yet. Share an insight, caveat, or question.
Nyame et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: