arXiv subjects

Martin Leutbecher

Publications and source records attributed to Martin Leutbecher.

8 recordsLinked to original sources

arXiv2026

On the sensitivity of machine-learned probabilistic weather forecast models to scale-aware scoring rules

Probabilistic forecast models can be machine-learned from data using loss functions based on scoring rules such as the Continuous Ranked Probability Score (CRPS). This note summarises a preliminary study comparing versions of AIFS-CRPS, a global weather forecast model, trained with different univariate and multivariate scoring rules that aim to explicitly represent scale-awareness in the loss function. In the first part, we compare the (almost) fair CRPS, a fair global energy score, and a graph energy score based on node neighbourhoods. Across standard verification metrics, forecast skill is broadly similar. In the extratropics we find only small differences, while in the tropics the graph energy score setup performs somewhat better and the global energy score shows some degradation. These results suggest that multivariate scores are a viable alternative to CRPS-based training for global machine-learned weather forecasting. In the second part of the study, we analyse how different scoring rules and scale-aware loss constraints shape the spectra of forecast fields. It is apparent that any form of explicit scale-awareness improves realism. Here, the largest differences are likely associated with different effective weights per scale.

physics.ao-ph↗

arXiv2026

Downscaling weather forecasts from Low- to High-Resolution with Diffusion Models

We introduce a probabilistic diffusion-based method for global atmospheric downscaling implemented within the Anemoi framework. The approach transforms low-resolution ensemble forecasts into high-resolution ensembles by learning the conditional distribution of finer-scale residuals, defined as the difference between the high-resolution fields and the interpolated low-resolution inputs. The system is trained on reforecast pairs from ECMWF IFS, using coarse fields at 100 km to reconstruct fine-scale variability at 30 km resolution. The bulk of the training focuses on recovering small-scale structures, while fine-tuning in high-noise regimes enables the generation of extremes. Evaluation against the medium-range IFS ensemble target shows that the model increases probabilistic skill (FCRPS) for surface variables, reproduces target power spectra at small scales, captures physically consistent multivariate relationships such as wind-pressure coupling, and generates extreme values consistent with those of the target ensemble in tropical cyclones.

physics.ao-ph↗

arXiv2025

Fair Box ordinate transform for forecasts following a multivariate Gaussian law

Monte Carlo techniques are the method of choice for making probabilistic predictions of an outcome in several disciplines. Usually, the aim is to generate calibrated predictions which are statistically indistinguishable from the outcome. Developers and users of such Monte Carlo predictions are interested in evaluating the degree of calibration of the forecasts. Here, we consider predictions of $p$-dimensional outcomes sampling a multivariate Gaussian distribution and apply the Box ordinate transform (BOT) to assess calibration. However, this approach is known to fail to reliably indicate calibration when the sample size n is moderate. For some applications, the cost of obtaining Monte-Carlo estimates is significant, which can limit the sample size, for instance, in model development when the model is improved iteratively. Thus, it would be beneficial to be able to reliably assess calibration even if the sample size n is moderate. To address this need, we introduce a fair, sample size- and dimension-dependent version of the Gaussian sample BOT. In a simulation study, the fair Gaussian sample BOT is compared with alternative BOT versions for different miscalibrations and for different sample sizes. Results confirm that the fair Gaussian sample BOT is capable of correctly identifying miscalibration when the sample size is moderate in contrast to the alternative BOT versions. Subsequently, the fair Gaussian sample BOT is applied to two to 12-dimensional predictions of temperature and vector wind using operational ensemble forecasts of the European Centre for Medium-Range Weather Forecasts (ECMWF). Firstly, perfectly reliable situations are considered where the outcome is replaced by a forecast that samples the same distribution as the members in the ensemble. Secondly, the BOT is computed using estimates of the actual temperature and vector wind from ECMWF analyses.

stat.ME↗

arXiv2025

A multi-scale loss formulation for learning a probabilistic model with proper score optimisation

We assess the impact of a multi-scale loss formulation for training probabilistic machine-learned weather forecasting models. The multi-scale loss is tested in AIFS-CRPS, a machine-learned weather forecasting model developed at the European Centre for Medium-Range Weather Forecasts (ECMWF). AIFS-CRPS is trained by directly optimising the almost fair continuous ranked probability score (afCRPS). The multi-scale loss better constrains small scale variability without negatively impacting forecast skill. This opens up promising directions for future work in scale-aware model training.

physics.ao-ph↗

arXiv2025

Unbiased calculation, evaluation, and calibration of ensemble forecast anomalies

Long-range ensemble forecasts are typically verified as anomalies with respect to a lead-time dependent climatological mean to remove the influence of systematic biases. However, common methods for calculating anomalies result in statistical inconsistencies between forecast and verification anomalies, even for a perfectly reliable ensemble. It is important to account for these systematic effects when evaluating ensemble forecast systems, particularly when tuning a model to improve the reliability of forecast anomalies or when comparing spread-error diagnostics between systems with different reforecast periods. Here, we show that unbiased variances and spread-error ratios can be recovered by deriving estimators that are consistent with the values that would be achieved when calculating anomalies relative to the true, but unknown, climatological mean. An elegant alternative is to construct forecast climatologies separately for each member, which ensures that forecast and verification anomalies are defined relative to reference climatological means with the same sampling uncertainty. This alternative approach has no impact on forecast ensemble means but systematically modifies the total variance and ensemble spread of forecast anomalies in such a way that anomaly-based spread-error ratios are unbiased without any explicit correction for climatology sample size. Furthermore, the improved statistical consistency of forecast and verification anomalies means that probabilistic forecast skill is optimised when the underlying forecast is also perfectly reliable. Alternative methods for anomaly calculation can thus impact probabilistic forecast skill, especially when anomalies are defined relative to climatologies with a small sample size. Finally, we demonstrate the equivalence of anomalies calculated using different methods after applying an unbiased statistical calibration.

physics.ao-ph↗

arXiv2024

AIFS-CRPS: Ensemble forecasting using a model trained with a loss function based on the Continuous Ranked Probability Score

Over the last three decades, ensemble forecasts have become an integral part of forecasting the weather. They provide users with more complete information than single forecasts as they permit to estimate the probability of weather events by representing the sources of uncertainties and accounting for the day-to-day variability of error growth in the atmosphere. This paper presents a novel approach to obtain a weather forecast model for ensemble forecasting with machine-learning. AIFS-CRPS is a variant of the Artificial Intelligence Forecasting System (AIFS) developed at ECMWF. Its loss function is based on a proper score, the Continuous Ranked Probability Score (CRPS). For the loss, the almost fair CRPS is introduced because it approximately removes the bias in the score due to finite ensemble size yet avoids a degeneracy of the fair CRPS. The trained model is stochastic and can generate as many exchangeable members as desired and computationally feasible in inference. For medium-range forecasts AIFS-CRPS outperforms the physics-based Integrated Forecasting System (IFS) ensemble for the majority of variables and lead times. For subseasonal forecasts, AIFS-CRPS outperforms the IFS ensemble before calibration and is competitive with the IFS ensemble when forecasts are evaluated as anomalies to remove the influence of model biases.

physics.ao-ph↗

arXiv2024

Ensemble size dependence of the logarithmic score for forecasts issued as multivariate normal distributions

Multivariate probabilistic verification is concerned with the evaluation of joint probability distributions of vector quantities such as a weather variable at multiple locations or a wind vector for instance. The logarithmic score is a proper score that is useful in this context. In order to apply this score to ensemble forecasts, a choice for the density is required. Here, we are interested in the specific case when the density is multivariate normal with mean and covariance given by the ensemble mean and ensemble covariance, respectively. Under the assumptions of multivariate normality and exchangeability of the ensemble members, a relationship is derived which describes how the logarithmic score depends on ensemble size. It permits to estimate the score in the limit of infinite ensemble size from a small ensemble and thus produces a fair logarithmic score for multivariate ensemble forecasts under the assumption of normality. This generalises a study from 2018 which derived the ensemble size adjustment of the logarithmic score in the univariate case. An application to medium-range forecasts examines the usefulness of the ensemble size adjustments when multivariate normality is only an approximation. Predictions of vectors consisting of several different combinations of upper air variables are considered. Logarithmic scores are calculated for these vectors using ECMWF's daily extended-range forecasts which consist of a 100-member ensemble. The probabilistic forecasts of these vectors are verified against operational ECMWF analyses in the Northern mid-latitudes in autumn 2023. Scores are computed for ensemble sizes from 8 to 100. The fair logarithmic scores of ensembles with different cardinalities are very close, in contrast to the unadjusted scores which decrease considerably with ensemble size. This provides evidence for the practical usefulness of the derived relationships.

stat.ME↗

arXiv2018

Statistical post-processing of dual-resolution ensemble forecasts

The computational cost as well as the probabilistic skill of ensemble forecasts depends on the spatial resolution of the numerical weather prediction model and the ensemble size. Periodically, e.g. when more computational resources become available, it is appropriate to reassess the balance between resolution and ensemble size. Recently, it has been proposed to investigate this balance in the context of dual-resolution ensembles, which use members with two different resolutions to make probabilistic forecasts. This study investigates whether statistical post-processing of such dual-resolution ensemble forecasts changes the conclusions regarding the optimal dual-resolution configuration. Medium-range dual-resolution ensemble forecasts of 2-metre temperature have been calibrated using ensemble model output statistics. The forecasts are produced with ECMWF's Integrated Forecast System and have horizontal resolutions between 18 km and 45 km. The ensemble sizes range from 8 to 254 members. The forecasts are verified with SYNOP station data. Results show that score differences between various single and dual-resolution configurations are strongly reduced by statistical post-processing. Therefore, the benefit of some dual-resolution configurations over single resolution configurations appears to be less pronounced than for raw forecasts. Moreover, the ranking of the ensemble configurations can be affected by the statistical post-processing.

stat.AP↗