Persistent Hypereutrophy and Limited One-Month-Ahead Forecastability in the Inner Bay of Lake Titicaca: Change Point Analysis and Leakage-Aware Temporal Validation

Abstract

High-altitude lakes are increasingly exposed to nutrient enrichment, organic loading and wastewater-derived contamination, while predictive performance can be overstated when temporally ordered observations are randomly partitioned or target-defining measurements are reused as predictors. We analyzed 177 consecutive monthly water-quality records from the Inner Bay of Lake Titicaca, Peru (January 2011–September 2025), integrating Carlson’s composite trophic state index (CTSI), Hamed–Rao modified Mann–Kendall tests, block-bootstrap Sen slopes, Pettitt change point detection with 12-month block permutation, nutrient stoichiometry, correlation, principal component analysis, leakage-aware contemporaneous classification and one-month-ahead forecasting. Total phosphorus was treated as elemental P and phosphate as PO4. The bay remained chronically hypereutrophic (mean CTSI 74.08 ± 4.07; 83.1% of months). BOD5 increased by 0.547 mg L−1 yr−1, chlorophyll-a by 3.254 mg m−3 yr−1, total suspended solids by 0.752 mg L−1 yr−1 and conductivity by 13.13 µS cm−1 yr−1, whereas total phosphorus declined by 0.085 mg P L−1 yr−1. Block-supported shifts occurred in chlorophyll-a in November 2016 (21.81 to 54.18 mg m−3) and BOD5 in June 2018 (6.54 to 11.77 mg L−1). After removing target-defining variables and preserving temporal order, the best trophic-state classifier had balanced accuracy 0.575 and MCC 0.257. Organic pollution classification had a balanced accuracy of 0.624 but a sensitivity of only 0.271, whereas the fecal-indicator model was unstable (MCC 0.130; ROC-AUC 0.460). Persistence was the best BOD5 forecast (RMSE 4.193 mg L−1; R2 0.390). The best CTSI model explained only 4.8% of future variance, and all thermotolerant coliform forecasts had negative out-of-time R2. Persistent ecological degradation was therefore evident, but monthly observations alone were insufficient for deployment-ready early warning. Higher-frequency sensing, hydrometeorological and wastewater load covariates, spatial replication, direct microbiological measurements and prospective validation are required.