Temperature estimation methods usually involve regression followed by kriging of residuals (residual kriging). Despite the performance of such models, there is invariably a residual which is not necessarily unpredictable because it may still be correlated in time. We set out to analyse such residuals through resort to autoregressive processes. It is shown that the optimal period varies depending on whether it is identified by functions of the form resd = f(resd−1, resd−2, …, resd−p) or by partial correlations. Autoregressive processes significantly improve estimates, which are evaluated by cross-validations. Finally, the two following points are discussed: (1) the assumptions of the autoregressive model on the residuals (the assumptions of linearity, stationarity of space and time are verified empirically) and (2) the identification of the days for which the introduction of this model is really interesting.