<p>Accurate demand forecasting is central to supply chain optimization, yet modern retail time series exhibit strong nonlinearity, intermittency, and regime-dependent variability that challenge classical statistical and deep learning models. This study investigates the applicability of quantum reservoir computing (QRC) for retail demand prediction under noisy intermediate-scale quantum (NISQ) compatible constraints. Using Walmart M5 dataset, spanning smooth, erratic, intermittent, and lumpy demand regimes, we systematically compare classical echo state networks (ESNs) and hybrid quantum-classical reservoirs under a unified global hyperparameter optimization protocol. The proposed variational quantum reservoir (VQR) employs angle encoding, nearest-neighbour entanglement, and fixed quantum dynamics with ridge regression readout, thereby avoiding barren plateau issues associated with trainable quantum circuits. Empirical results over a 28-day forecasting horizon demonstrate that multivariate QRC achieves the best global performance (WRMSSE <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\approx \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>≈</mo> </math></EquationSource> </InlineEquation> 0.808), outperforming both univariate quantum and classical reservoir baselines. Regime-specific analysis reveals that quantum reservoirs exhibit superior robustness in erratic and intermittent demand patterns and benefit significantly from multivariate feature integration, indicating enhanced expressive absorption through entanglement-driven feature mappings. Capacity scaling experiments further show nonlinear performance gains with increasing qubit count, contrasting with the saturation behaviour of classical reservoirs. Although current simulation-based quantum inference incurs substantial computational overhead, the findings highlight the potential of compact quantum reservoirs as high-capacity forecasting engines for future hardware implementations.</p>