The Gulf's PM2.5 forecast rests on one number a year for each country

Air pollution is associated with higher rates of illness and death from respiratory, cardiovascular and other non-communicable diseases, and accurate prediction is what allows health risks to be anticipated rather than counted afterwards [s1]. In the Gulf, that prediction problem is complicated by an arid climate in which particulate levels are shaped by dust as well as by emissions.

A study published in BMC Public Health on 10 January attempts a forecast anyway, across all six Gulf Cooperation Council states [s1]. The methods are more interesting than the numbers, and the limitation is worth understanding before either is used.

What the study did

The authors describe it as the first multi-country study across the GCC to forecast PM2.5 [s1]. They used annual PM2.5 data from the World Bank database for the six GCC countries, covering 1991 to 2021 [s1].

Against that series they ran traditional time series models — ARIMA, naïve, exponential smoothing — a non-parametric model (NPAR), and machine learning-based hybrid models, evaluating accuracy with RMSE, MAE, MAPE, nRMSE and the Diebold-Mariano test, and validating with rolling cross-validation [s1].

The hybrid models won. NPAR-NNAR produced the lowest RMSE for Saudi Arabia (2.0181), Qatar (2.2852), Bahrain (1.3145) and Kuwait (2.299), while NAIVE-NNAR performed best for the United Arab Emirates (1.0462) and Oman (1.6522) [s1].

On levels, the study reports Qatar consistently highest at 87.90 ± 5.78 µg/m³, followed by Bahrain at 67.42 ± 4.59 and Kuwait at 58.00 ± 5.95 [s1]. The forecast is that GCC countries may see variation in PM2.5 in coming decades, with Qatar and Bahrain facing the highest concentrations among them [s1].

The constraint the paper does not foreground

The input is annual data, at national resolution, for thirty-one years. That is 31 observations per country.

Every methodological sophistication in the paper — the hybrid architectures, the rolling cross-validation, the Diebold-Mariano comparisons — operates on that. Model selection between NPAR-NNAR and NAIVE-NNAR based on RMSE differences of a few tenths, on series of that length, is not a strong basis for concluding that one architecture is genuinely better suited to one country than another. It is more likely to reflect the particular shape of a short series.

More importantly, an annual national mean is a poor proxy for the exposure that damages health. It averages across every day of the year and across every person in a country, so it cannot distinguish a sustained moderate level from a handful of severe episodes, nor an air-conditioned indoor worker from someone working outdoors. The paper's own stated implication — that better forecasting supports early warning capabilities [s1] — sits awkwardly with a model that operates on a one-year timestep. Early warning requires days.

What the burden data says

A companion picture comes from a Global Burden of Disease analysis published in the Pakistan Journal of Medical Sciences in autumn 2025, conducted at King Saud University in Riyadh and covering the same six countries from 1990 to 2021 [s2].

It reports that the highest rate of PM exposure in 2021 was in Qatar, at 56.95 per 100 people, and the lowest in Oman, at 38.86 per 100 people [s2] — a different metric from the concentration figures above, and the two should not be compared directly. Qatar consistently achieved the highest decreases in death and disability-adjusted life year rates between 1990 and 2021 [s2].

The finding that stands out is a divergence. Across the diseases examined — respiratory, cardiovascular, diabetes and total cancers — diabetes was the only one that showed an increase in both deaths and DALYs in the GCC region over the period [s2]. The authors' conclusion is that the rising trend in air pollutant exposure is attributable to death and DALY rates in GCC countries [s2].

That last sentence is stated more strongly than a burden-attribution exercise supports. GBD estimates are themselves modelled, and diabetes trends in the Gulf over three decades are driven by diet, obesity and demographic change to a degree that no PM2.5 series can disentangle. The convergence of two very different national-level datasets on Qatar as the highest-exposure state is worth noting; the causal claim attached to it is not established by either.

What would actually improve the picture

Ground-level monitoring at sub-national resolution, publicly released at daily or hourly timesteps, is the missing input. Several Gulf states operate monitoring networks; the data is not consistently available in forms that support the kind of analysis both of these papers had to substitute modelling for. Until it is, forecasts of Gulf particulate exposure will keep being built on the coarsest possible representation of the problem.

What to watch

Whether GCC environment agencies publish station-level PM2.5 series, and whether future studies separate dust-storm days from background exposure — the distinction that determines whether any of this can support a health warning system.

Sources

1. [s1] Modeling and forecasting air pollution for public health protection based on ML and time series models in Gulf Cooperation Council (GCC) countries. BMC Public Health, 10 January 2026. https://doi.org/10.1186/s12889-025-26164-9
2. [s2] Effect of particulate matter (PM2.5, PM10) on deaths and disability-adjusted life years (DALYs) in Gulf Cooperation Council countries: Global burden of disease time trend analysis 1990-2021. Pakistan Journal of Medical Sciences, September 2025. https://doi.org/10.12669/pjms.41.10.12523