The purpose of this study was to evaluate the accuracy and operational suitability of machine learning models for demand-side water forecasting at Nakuru Water and Sanitation Services Company, where 21.7 per cent of the water entering the London Ward network never reached a paying consumer. Four models (ARIMA, FB Prophet, LSTM and Bidirectional LSTM) were re-evaluated on a 964-day operational series under a chronological protocol with a 192-day held-out window, each compared against the naive baseline admissible under its own information set. Models were separated by forecast horizon: lag-based models consume yesterday's realised consumption whereas decomposable models consume only the calendar date, and a rationing schedule must be published in advance. As configured in the parent study, Prophet achieved a mean absolute error of 128.3 m³/day, effectively identical to a straight line fitted to the training data (129.2 m³/day): it had recovered the trend and nothing further. The training partition's autocorrelation showed a dominant 30-day cycle (r = +0.867) and negligible weekly structure (r = +0.203), the reverse of the configured seasonality. Re-specifying Prophet with a cycle length selected on an inner validation split reduced held-out error to 47.7 m³/day, a 62.8 per cent reduction, outperforming the best lag-based model by 16.9 per cent without observed consumption. The study concluded that diagnostic-driven seasonality specification contributed more to forecast accuracy than algorithm choice, and recommends deriving seasonal structure from the training autocorrelation rather than from software defaults.