Long-term urban water demand forecasting remains a central yet unresolved challenge for water-utility planning. Despite its importance for capital investment, drought preparedness, and financial strategy, forecasting practice has advanced little in recent decades and continues to rely on short, local historical records, deterministic projections, and simplified assumptions about how demand evolves over time. These approaches obscure structural uncertainty, fail to represent the dynamic interactions between demand, pricing, and hydrologic conditions, and rarely evaluate predictive performance under true out-of-sample conditions. This dissertation examines these limitations and develops three complementary methodological contributions to strengthen long-term demand analysis under uncertainty. The first paper develops a stress-testing framework that explicitly characterises multiple sources of uncertainty in long-term demand models, including uncertainty in model structure, exogenous drivers, and behavioural elasticity parameters. Rather than seeking a single authoritative forecast, the framework propagates these uncertainties through a representative econometric model to generate an ensemble of plausible demand trajectories. Global sensitivity analysis and scenario-discovery methods are then used to quantify the contribution of individual uncertainties and to identify combinations of conditions that give rise to high or otherwise operationally challenging demand futures. This provides a quantitative basis for understanding the breadth and structure of uncertainty in long-term demand forecasts. The second paper develops a dynamic modelling framework that represents demand as an endogenous component of the socio-hydrologic-financial system within which utilities operate. Using state-transition models that incorporate drought restrictions, pricing responses, and behavioural persistence, the study shows how long-term consumption evolves through interactions between demand, policy, and hydrologic conditions. The results demonstrate that static forecasting models systematically mischaracterise these dynamics, leading to biased long-range projections. The third paper evaluates long-term forecasting performance using pooled, multi-utility datasets and non-parametric machine-learning methods. By testing models on utilities not included in their training data, the study provides one of the first systematic assessments of out-of-sample predictive skill for long-term urban water demand. The results show that models calibrated on single-utility datasets generalise poorly, whereas pooled-data machine-learning approaches substantially improve predictive accuracy. Together, these contributions establish a stress-testing framework that fully characterises uncertainty in long-term demand models and supports the informed construction of scenarios for use in broader analyses of water-resource system performance.