We propose a fully data-driven, interpretable, and computationally scalable framework to characterize sudden stratospheric warmings (SSWs), extract statistically significant precursors, and produce machine learning (ML) forecasts. By successfully leveraging the long-lasting impact of SSWs, the ML predictions outperform sub-seasonal numerical forecasts for lead times beyond 25 d. Post-processing numerical predictions using their ML counterparts yields a performance increase of up to 20 %.

Sudden stratospheric warmings are extreme weather events where the winter polar stratosphere warms by about 25 K. An improved representation of small-scale gravity waves in sub-seasonal prediction models can reduce forecast errors since their impact on the large-scale circulation is predictable multiple weeks ahead. After a sudden stratospheric warming, vertically propagating gravity waves break at a lower altitude than usual, which strengthens the long-lasting positive temperature anomalies.

This study quantifies the impact of poorly constrained parameters used to represent aerosol–cloud–precipitation interactions on precipitation and cloud forecasts associated with uncertainties in input atmospheric states. Uncertainties in these parameters have a non-negligible impact on daily precipitation amount and largely change the amount of cloud. The comparison between different weather situations reveals that the impact becomes more important when convection is triggered by local effects.

We use a statistical dynamical model to generate precipitation forecasts for lead times of 2–6 weeks over southwest Asia, which are needed to support humanitarian food distribution. The model signal-to-noise ratio is used to identify a smaller subset of forecasts with particularly high skill, so-called subseasonal forecasts of opportunity (SFOs). Precipitation SFOs are often related to slowly evolving tropical phenomena, namely the El Niño–Southern Oscillation and Madden–Julian Oscillation.

The El Niño–Southern Oscillation (ENSO) and Arctic Oscillation (AO) are phenomena that affect the weather and climate of North America. Although ENSO hails from from the tropical Pacific and the AO high above the North Pole, the spatial patterns of their influence on a North American tornado environment index are remarkably similar in computer models. We find that when ENSO and the AO act in concert, their impact is large, and when they oppose each other, their impact is small.

Accurate predictions of the stratospheric polar vortex can enhance surface weather predictability. Stratospheric events themselves are less predictable, with strong inter-event differences. We assess the predictability of stratospheric acceleration and deceleration events in a sub-seasonal prediction system, finding that the predictability of events is largely dependent on event magnitude, while extreme drivers of deceleration events are not fully represented in the model.

Northern Hemisphere winter weather is influenced by the strength of westerly winds 30 km above the surface, the so-called polar vortex. Eurasian autumn snow cover is thought to modulate the polar vortex. So far, however, the modeled influence of snow on the polar vortex did not fit the observed influence. By analyzing a model experiment for the time span of 110 years, we could show that the causality of this impact is indeed sound and snow cover can weaken the polar vortex.

While the assumption of Gaussianity leads to many simplifications, ensemble forecasts often exhibit non-Gaussian distributions. This work has systematically identified the presence of a specific case of non-Gaussianity, bimodality. It has been found that bimodality occurs in a large portion of global 2 m temperature forecasts. This has drastic implications on forecast skill as the minimum probability in a bimodal distribution often lies at the maximum probability of a Gaussian distribution.

Forecasting temperature patterns beyond 2 weeks is very challenging, although occasionally, forecasts show more skill over Europe. Our study indicates that the level of skill varies concurrently for two distinct forecast systems. It also shows that higher skill occurs when forecasts are issued during specific patterns of atmospheric circulation that tend to be particularly persistent. These results could help forecasters estimate a priori how trustworthy extended-range forecasts will be.

This work assesses the forecast of the temperature over the Sahara, a key driver of the West African Monsoon, at a seasonal timescale. The seasonal models are able to reproduce the climatological state and some characteristics of the temperature during the rainy season in the Sahel. But, because of errors in the timing, the forecast skill scores are significant only for the first 4 weeks.

We use an advanced statistical approach to investigate the dynamics of the development of sudden stratospheric warming (SSW) events in the winter Northern Hemisphere. We identify distinct signals that are representative of these events and their event type at lead times beyond currently predictable lead times. The results can be viewed as a promising step towards improving the predictability of SSWs in the future by using more advanced statistical methods in operational forecasting systems.

The prediction capabilities of GPS, operational (low-resolution) and targeted (high-resolution) radiosondes for data assimilation in a Mediterranean heavy precipitation event at different model resolutions are investigated. The results show that even if GPS provides accurate observations, their lack of vertical information hampers the improvement, demonstrating the need for assimilating radiosondes, where the location and timing of release was more determinant than the vertical resolution.

Why do some monsoon seasons start early, while others start late? For the Australian monsoon, some previous research suggested the El Niño–Southern Oscillation in the months before the onset influenced the monsoon timing. This research tests if this is still correct and if other large-scale climate patterns also influenced onset timing. We found that a strong La Niña pattern usually coincided with an early onset but weak La Niña and El Niño patterns did not show a consistent pattern.

Weather variability controls the transport of ozone from the stratosphere to the Earth’s surface and water vapor from oceanic source regions to continental land masses. Forecasting these types of transport has high societal value because of the negative impacts of ozone on human health and the role of water vapor in governing precipitation variability. We use upper-level wind forecasts to assess the potential for predicting ozone and water vapor transport 3–6 weeks ahead of time.

We analysed a tornadic supercell that tracked through the northern Highveld region of South Africa for 7 h. We found that atmospheric conditions were conducive for tornado-associated severe storms over the region. A 4.4 km resolution model run by the South African Weather Service was able to predict this supercell, including its timing. However, it underestimated its severity due to underestimations of other important factors necessary for real-world development of these kinds of storms.

In September 2018 an intense Mediterranean cyclone with structural similarities to a hurricane, a so-called medicane, caused severe damage in Greece. Its development was uncertain, even just a few days in advance. The reason for this was uncertainties in the jet stream over the North Atlantic 3 d prior to cyclogenesis that propagated into the Mediterranean. They led to an uncertain position of the upper-level disturbance and, as a result, of the position and thermal structure of the cyclone.

Seasonal outlooks of Atlantic hurricane activity are required to enable risk reduction measures and disaster preparedness. Many seasonal forecasts are based on a selection of climate signals from which a statistical model is constructed. The crucial step in this approach is to select the most relevant predictors without overfitting. Here we show that causal effect networks can be used to identify the most robust predictors. Based on these predictors we construct a competitive forecast model.

Tropical storms are key for harnessing midlatitude weather prediction skill 2–8 weeks into the future. Recently, stratospheric activity was shown to impact tropical storminess and thus may also be important for midlatitude prediction skill on these timescales. This work analyzes two forecast systems to assess whether they capture this additional skill. We find there is enhanced prediction out through week 4 when both the tropical and stratospheric phenomena are active.

The mesoscale convective system (MCS) that affected Germany at Pentecost 2014 was one of the most severe for decades. However, the predictability of this system was very low. By moving the model domain by just one grid point changed whether the MCS was successfully simulated or not. The decisive factor seems to be small differences in the initial track of the convection: cooler air near the coast inhibited development there, but tracks slightly more inland found more favorable conditions.