Forecasting HIV Incidence in Somalia Using Single and Hybrid Time Series Models: A Comparative Analysis

Background HIV incidence is an important indicator for monitoring transmission and progress toward HIV control. In Somalia, long-term surveillance is constrained and formal forecasting studies remain limited. This study examined historical trends and compared single and hybrid time-series models for forecasting HIV incidence among people aged 15–49. Methods Annual HIV-incidence estimates for Somalia from 1990–2022 (n = 33) were obtained from the World Bank World Development Indicators (WDI), which republishes UNAIDS modelled estimates. Stationarity was assessed using ADF, PP, and KPSS tests. The methodological specification assigns differencing to integrated models (ARIMA/ARFIMA), while ETS, TBATS, Theta, and NNAR are defined on the original-scale series. Six individual models and ten equal-weight pairwise hybrids were compared using MAPE and sMAPE, with Theil’s U reported where available. The primary evaluation used 1990–2014 for training and 2015–2022 for testing; an additional rolling-origin one-step-ahead sensitivity analysis evaluated four focal models over 2005–2022. Results The deposited rounded WDI series contained 18 unique values; 0.10 occurred in 14 of 33 years (42.4%) and formed a continuous plateau from 2009–2022, making all eight observations in the primary test window identical. Within this fixed holdout, ETS(M,N,N) had the lowest reported individual-model error (MAPE = 0.038%; sMAPE = 0.00038), and ETS–NNAR had the lowest hybrid-model error (MAPE = 0.321%; sMAPE = 0.00321). In the rolling-origin sensitivity analysis, ETS remained competitive (MAPE = 2.787%), but the separation from ETS–NNAR (3.165%) and NNAR (3.543%) was substantially smaller; ARIMA had MAPE = 9.512%. Forecasts for 2023–2029 remained close to 0.10 per 1,000 uninfected population. Model-based prediction intervals widened over time and do not incorporate the separate uncertainty of the underlying UNAIDS estimates. Conclusions The data support a long-term decline followed by a prolonged low reported incidence level, but the flat and rounded tail materially limits discrimination among forecasting models. The fixed-holdout rankings should therefore be interpreted as conditional on the evaluation window rather than as evidence of general model superiority. Forecasts are most appropriately used as planning scenarios alongside strengthened surveillance and improved access to less-rounded, uncertainty-bounded source estimates.