Isotonic Regression and Artificial Intelligence in Econometrics: Applications in the Era of Big Data

Abstract

The challenge of big data in econometrics lies in effectively leveraging vast amounts of information collected on economic behavior. Conventional statistical techniques are often inadequate for handling such complex and high-dimensional datasets. Isotonic regression, a nonparametric approach, estimates a regression function while preserving monotonicity constraints. Despite the rapid expansion of machine learning applications in econometrics, the literature still lacks a unified framework that explicitly integrates economically meaningful monotonicity constraints into machine learning pipelines. Existing approaches either preserve interpretability at the cost of predictive power, or maximise accuracy at the cost of theoretical consistency. The present study addresses this gap by proposing a hybrid AI–isotonic framework that reconciles both requirements. The framework deploys isotonic regression in two complementary roles, a smooth predictive estimator that regularises machine learning outputs under monotonicity constraints, and a segmented variant used as a structural diagnostic tool for the identification of regime changes in macroeconomic series. The methodology is applied to three Moroccan datasets, namely the World Bank GDP series (1960–2020), the M3 monetary aggregate published by Bank Al-Maghrib (1985–2025) and the telecommunications revenue series of the Haut-Commissariat au Plan (1997–2024). Empirical results, evaluated through the Root Mean Squared Error, the Mean Absolute Error and the coefficient of determination R², indicate that the hybrid approach reduces the in-sample forecasting error by approximately 90 % on GDP, 98 % on M3 and 76 % on telecommunications revenue, relative to conventional linear regression. These findings confirm that economically constrained machine learning can simultaneously achieve predictive accuracy and theoretical consistency.

JEL Classification: C01; C52; C55

Paper type: Empirical research