Baseline-aware monthly PM10 prediction in Istanbul and Ankara: a city-level panel study comparing persistence, AR(1), and random forest models

Abstract

Monthly PM 10 prediction was evaluated for Istanbul and Ankara with a 2020 − 2023 city-month panel combining air-quality, meteorological, and traffic-intensity variables. The calendar index comprised 96 city-months, of which 92 had observed PM 10 . Missing PM 10 targets were excluded from scoring. The primary analysis required an observed one-month PM 10 lag; a sensitivity analysis permitted linearly interpolated values only for lag construction. Persistence, city-specific AR(1), and Random Forest models were assessed in five expanding-window rolling-origin folds and a ten-month chronological holdout. Predictive R 2 was calculated as 1 − SSE/SST. In pooled cross-validation, Random Forest achieved RMSE = 6.89 µg/m 3 , MAE = 5.53 µg/m 3 , and R 2 = 0.50; corresponding values were 7.40, 5.68, and 0.43 for AR(1), and 8.65, 6.65, and 0.22 for persistence. In the final holdout, persistence achieved the lowest error (RMSE = 4.57 µg/m 3 ; MAE = 3.71 µg/m 3 ; R 2 = 0.55), followed by Random Forest (5.87, 4.71, and 0.26) and AR(1) (6.20, 5.14, and 0.17). Block-bootstrap intervals supported an AR(1) RMSE reduction in pooled cross-validation; holdout differences from persistence included zero. Model ranking therefore varied between development and final holdout periods, with persistence providing the strongest holdout performance.