Finding physical connections in a statistical model: Using global uncertainty and sensitivity analysis on a multiple predictor analog methodology for downscaling and bias correcting precipitation forecasts

With the recent availability of computationally expensive Numerical Weather Prediction Model archives, the use of analog or pattern matching methodologies has become a viable option for downscaling and bias correction of forecasts. Numerous studies have been performed that show skill using these methods for local scale temperature, reference evapotranspiration and precipitation forecasting. However, as this is a statistical, pattern matching approach, little study has gone into ascertaining a relationship between local climatic physics and the choice of predictors to use in a given condition. This study uses global sensitivity and uncertainty analysis with subsequent Monte-Carlo filtering to determine if there are such linkages with regard to seasonality, lead-day of the forecast, and precipitation event magnitude.;While this study showed that non-atmospheric input factors (search window size and number of analogs used) had the greatest effect in model output variability, it was determined that variations in the weighting of atmospheric predictors does offer the potential for improved forecast skill under different conditions. It was found that for the case of the Tampa Bay region of Florida, the influence of precipitation, precipitable water, relative humidity, the meridional wind vector and the zonal wind vectors changed with respect to the winter dry season and the summer wet season as well as event threshold. Analysis showed these changes to be statistically significant. Relationships are posited whereby these changes are potentially attributed to climatic physics. An example is the increased influence of the east-west wind vector in the summer season as opposed to winter. Due to the common occurrence of convective storm formation from the sea breeze, it is plausible that this signal is showing up in the analog method. While this study is an initial foray into connecting physics to a statistical model and statistically significant variations in predictor influence was observed, further research is required to validate any causal relationships.