This paper presents a comparison between parallel linear and parallel neural network models. Parallel models consist of 24 separate models, one for each hour of the day. Each parallel model decomposes the load into a linear Auto- Regressive (AR) part and a residual. Exogenous linear and neural network model performance is compared in predicting this residual. Three days or 72 hours of current and delayed weather variables are available as exogenous inputs for the residual models. Input selection comprises of testing the bootstrapped performance of a linear model. The inputs are ordered using 4 methods derived from a mix of the T-ratio of the linear coefficients and Principal Component Analysis (PCA). The neural network models are found to give superior results due to the non-linear AR nature of the residual.