Extended logistic regression is a recent ensemble calibration method that extends logistic regression to provide full continuous probability distribution forecasts. It assumes conditional logistic distributions for the (transformed) predictand and fits these using selected predictand category probabilities. In this study extended logistic regression is compared to the closely related ordered and censored logistic regression models. Ordered logistic regression avoids the logistic distribution assumption but does not yield full probability distribution forecasts, whereas censored regression directly fits the full conditional predictive distributions. The performance of these and other ensemble postprocessing methods is tested on wind speed and precipitation data from several European locations and ensemble forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF). Ordered logistic regression performed similarly to extended logistic regression for probability forecasts of discrete categories whereas full predictive distributions were better predicted by censored regression.

The crch package provides functions for maximum likelihood estimation of censored or truncated regression models with conditional heteroscedasticity along with suitable standard methods to summarize the fitted models and compute predictions, residuals, etc. The supported distributions include left-or right-censored or truncated Gaussian, logistic, or student-t distributions with potentially different sets of regressors for modeling the conditional location and scale. The models and their R implementation are introduced and illustrated by numerical weather prediction tasks using precipitation data for Innsbruck (Austria).

Standard-Nutzungsbedingungen:Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden.Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen.Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in Achim ZeileisUniversität Innsbruck Abstract Non-homogeneous regression is often used to statistically post-process ensemble forecasts. Usually only ensemble forecasts of the predictand variable are used as input but other potentially useful information sources are ignored. Although it is straightforward to add further input variables, overfitting can easily deteriorate the forecast performance for increasing numbers of input variables. This paper proposes a boosting algorithm to estimate the regression coe cients while automatically selecting the most relevant input variables by restricting the coe cients of less important variables to zero. A case study with ensemble forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF) shows that this approach e↵ectively selects important input variables to clearly improve minimum and maximum temperature predictions at 5 central European stations.

No abstract

Wind power forecast evaluation is of key importance for forecast provider selection, forecast quality control, and model development. While forecasts are most often evaluated based on squared or absolute errors, these error measures do not always adequately reflect the loss functions and true expectations of the forecast user, neither do they provide enough information for the desired evaluation task. Over the last decade, research in forecast verification has intensified, and a number of verification frameworks and diagnostic tools have been proposed.However, the corresponding literature is generally very technical and most often dedicated to forecast model developers. This can make forecast users struggle to select the most appropriate verification tools for their application while not fully appraising subtleties related to their application and interpretation. This paper revisits the most common verification tools from a forecast user perspective and discusses their suitability for different application examples as well as evaluation setup design and significance of evaluation results. KEYWORDS evaluation metrics, forecast evaluation, testing forecast performance, wind power INTRODUCTIONWind power has become an important power source in many power systems. In Europe, it already covers approximately 12% of the total electricity demand. 1 However, variability and limited predictability of its production challenge power systems and markets, making forecasts required for optimal operation (eg, load balancing and maintenance) and trading. A lot of research has been carried out in the development of wind power forecasting models, and a variety of models have been proposed for different applications and types of forecasts. These include deterministic point predictions, probabilistic forecasts of various forms, and multivariate predictions or predictions for specific events such as ramps or gusts, 2 see, eg, Giebel et al 3 for a general state-of-the-art report on wind power forecasting or Kariniotakis 4 for a recent coverage of challenges related to wind power forecasting (and extension to other renewable energy sources).One of the current challenges, which is rarely covered and discussed, is forecast verification, maybe since many believe that verification frameworks are well-established and forecast users are content with their use. Forecast evaluation is crucial for model development, selection of the best forecast provider, or for quality control. Some of its main goals include estimation of future error statistics, comparison of the forecast accuracy of different forecasts, or finding flaws in a certain forecast model. Unfortunately, it is not the case that current knowledge in forecast verification and existing verification frameworks can give us the whole information about objective quality of forecasts and their value to forecast users. The original view on forecast quality and value (inspired by meteorological applications) was laid out in the 1980s by Murphy. 5,6 More recently, this aspect was discussed by L...