Forecasting Internet adoption in Southeast Asia with machine learning: Do mobile and fixed broadband improve on temporal persistence?

Abstract

Internet use has expanded rapidly across Southeast Asia, but national trajectories remain heterogeneous and short-horizon forecasting evidence is limited. This study evaluated whether routinely reported connectivity indicators improved one-year-ahead forecasts of the World Bank indicator “individuals using the Internet (\% of population)” beyond a persistence benchmark for the 11 current ASEAN member states. A longitudinal archive contained 1,120 observed country-indicator-year records, with a 2000–2024 analytical panel. Rolling-origin evaluation covered 2016–2024, with each test year predicted only from earlier observations. The revised primary ridge-regression specification selected its penalty parameter within each outer training window by temporally ordered nested validation. Dependence-aware uncertainty used country-cluster bootstrap intervals, test-year resampling, a crossed country-by-year bootstrap sensitivity analysis, and an exact test on nine annual mean loss differentials. The nested ridge model achieved mean absolute error (MAE) 3.70 percentage points versus 5.04 for persistence, a 26.6\% reduction, with RMSE 5.83, pooled R² 0.941, and persistence-relative squared-error skill 0.343. The MAE difference was -1.34 percentage points (country-cluster 95\% CI -2.13 to -0.57), although the crossed country-by-year interval included zero. On the identical 89-observation complete-case sample, adding lagged fixed broadband changed MAE from 3.852 to 3.887 (difference 0.035; country-cluster 95\% CI -0.104 to 0.178), showing no clear incremental short-horizon gain. Out-of-sample permutation importance identified previous-year Internet use as the dominant predictor. These results support leakage-resistant forecasting benchmarks while cautioning against causal or AI-readiness interpretations.