Deconstructing India’s Headline CPI: A Comparative Analysis of Climate, Macroeconomic, and Commodity Shocks

Abstract

Headline Consumer Price Index (CPI) inflation in India responds to concurrent domestic agricultural supply
disruptions, global commodity price movements, and monetary conditions. This paper compares two
forecasting-and-attribution paradigms on the same monthly data and the same chronological holdout. The
machine learning (ML) paradigm comprises ElasticNet, XGBoost, Support Vector Regression (SVR), and
a Stacking Regressor; the time-series (TS) econometric paradigm comprises a regularized ARDL(p, q), a
block-structured Dynamic Factor Model, a VARX, and a Bates-Granger optimal forecast combination.
Each model is estimated under four feature sets (Full, Climate-only, Macroeconomic-only, Commodity-
only) with two CPI persistence controls. The Stacking Regressor under the Full configuration delivers the
lowest holdout error (RMSE = 0.634, R2 = 0.858), 21% below the best Full-configuration TS model, a gap
that survives Benjamini-Hochberg-adjusted Diebold-Mariano testing. Grouped SHAP, Group Permutation
Feature Importance (GPFI), and expanding-window stability diagnostics identify Commodity, in particular
primary food prices, as the Core Driver of CPI in the ML paradigm and as the category with the largest
permutation loss in the TS paradigm. Macroeconomic variables are a Redundant Driver once Commodity
information is present, and Climate variables are Mixed/Inconclusive. The performance advantage of the
ensemble over dynamic time-series specifications reflects sparse selection among collinear wholesale food
indicators and the parameter variance costs associated with distributed lag expansion, though these
mechanisms remain unisolated by controlled ablation.