Quantifying Detection Latency in River Water Quality Regulatory Compliance Monitoring via Multivariate Exponentially Weighted Moving Average Control Charts: A 36-Year Analysis of São Paulo State Watersheds Benchmarked Against Eight European Union Countries

The statistical efficiency with which regulatory monitoring networks detect compliance-to-non-compliance transitions in river pollution has never been systematically quantified. This study develops a multivariate exponentially weighted moving average (MEWMA) statistical process control framework, extended for strongly non-normal river pollutant data, to quantify pollution detection efficiency across six São Paulo State watersheds (UGRHIs) over 36 years (1990–2025), drawing on 646,356 records from 141 CETESB monitoring stations benchmarked against eight European Union countries (GRQA v1.4; 998,410 records). Using dissolved oxygen, BOD 5 , total phosphorus, and nitrate under CONAMA 357/2005 Class 2 thresholds, Phase I analysis reveals four of the six UGRHIs already exceeding pollutant limits before 2005, two of them chronically so, establishing degraded baselines. Phase II MEWMA charts yield out-of-control ARL₁ ARL 1 values of 3.5–5.4 annual observations (mean 4.4), of the same order of magnitude as the EU reference (mean 3.5) — yet Alto Tietê dissolved oxygen (1.1 mg/L) is 8.8-fold below the EU median (9.7 mg/L). This broad comparability in detection sensitivity does not indicate comparable water quality: ARL-based metrics measure relative process change and must be interpreted alongside absolute pollutant concentrations to avoid misleading regulatory conclusions. A chronic degradation detection paradox is documented in UGRHI 10 (Sorocaba/Médio Tietê) and UGRHI 06 (Alto Tietê), where zero alarms were generated despite BOD 5 and TP violation rates above 80%, because the chronically polluted baseline renders the chart blind to relative deterioration. Annual monitoring implies detection lags of 42–65 months; quarterly monitoring reduces this to 11–16 months. These findings provide the first ARL-based quantification of pollution detection efficiency for Brazilian monitoring networks and offer actionable criteria for optimising CETESB sampling frequency under CONAMA 357/2005.