This lecture has tried to reflect the dynamic nature of the subject material. Downscaling of seasonal to interannual predictions is a rapidly developing field and while substantial progress has been made, a number of unanswered questions remain regarding the best methodologies. To assist the reader to digest this subject matter, the following is a summary of some of the main points in this lecture:

• To justify making a contrasting anomaly forecast over a small spatial scale, evidence of physical mechanisms are needed (such as in Fig. 4.2).
• Sum High resolution regional climate models (RCMs) can be used to contribute to that evidence, along with diagnostic analysis of observations and statistical downscaling. The RCMs can potentially provide such diagnostic insight when they are driven with observed data (in practice, the approximations for observed fields as given by reanalysis fields), as well as with output from GCM seasonal prediction experiments.
• RCMs are being evaluated as a dynamical downscaling tool, and clearly provide improved spatial and temporal data above GCMs.
• Statistical downscaling of numerical model output: The large-scale output from GCMs or RCMs can be used as predictors for the needed downscaled information, such as station scale (or sub-regional rainfall index scale) predictands. This corrects the climate model for systematic spatial biases as well biases in mean and variance. It reduces the problem of predictor selection that is so difficult for purely statistical approaches.
• Daily weather sequences can be generated from these and other seasonal mean forecasts. Approaches include weather generators and analogue methods.
• Techniques exist to create daily weather sequences from the daily fields predicted by GCMs. These will involve some element of perfect-prognosis - that is assuming GCM fields can be used in relationships that are established between observed fields and the observed target downscaled information (such as daily station precipitation). Methods include regression based approaches (regressing large-scale predictors to the downscaled information) or approaches based on daily weather states. Related to the latter is the Hidden Markov Model for daily precipitation. Approaches that use daily GCM fields and have some reliance on perfect prognosis need to be compared with the simpler methods that generate daily weather sequences from monthly or seasonal timescale predictions. The latter may currently have the advantage that they can better extract prediction information from the GCM that is spatially displaced from observed patterns.
• When daily weather sequences are generated, the prediction is not interpreted as the exact sequence of daily weather, but rather, the statistics of the daily weather, like the number of storms in the season and the number of dry spells.
• The techniques can be applied to generate downscaled weather parameters, or information that more closely relates to decision support, such as stream flow. Or, the daily weather sequences can be used to drive impact models, such as crop models or stream flow models. These can be linked into utility evaluation models (including economic models) to provide estimates of long-term benefits (and costs) of response strategies to the downscaled seasonal forecast information. Ultimately, the information should integrate into decision support frameworks, which may involve provision of scenario consequences for different management strategies.

Acknowledgments: This material has evolved through discussions with many colleagues. Its development has also been assisted by the Columbia Center for New Media Teaching and Learning. (See Credits)