Hail Frequency Europe: ERA5, AR-CHaMo and ESRS E1-9

AR-CHaMo stands for Additive Regression Convective Hazard Model. The framework was developed at the European Severe Storms Laboratory (ESSL) by Rädler and co-authors in 2018 and refined by Francesco Battaglioli and co-authors, whose 2023 paper in Natural Hazards and Earth System Sciences (NHESS, Copernicus) describes the current models. The method is an additive logistic regression: from the atmospheric state in an ERA5 grid cell it estimates the probability of large hail (≥ 2 cm) and of the severity class very large hail (≥ 5 cm), which is why ESSL runs it as a joint lightning and large hail forecasting product.

The model carries that weight because the thing everyone asks for does not exist. "Hail projection for 2050" sounds like a climate scenario and is not one: hail is a sub-grid phenomenon no CMIP5 or CMIP6 model resolves. Site statements rest on observation (ESWD, DWD radar, GDV) and on proxy reanalyses.

AR-CHaMo on ERA5: Battaglioli et al. and the global trend

AR-CHaMo was trained on roughly 24 million lightning observations and 44,000 hail reports, then applied to the hourly ERA5 reanalysis of the ECMWF on a 0.25° × 0.25° grid. The training base explains why the acronym says regression, not simulation: it fits predictors to observed hail and lightning and evaluates that fit in every cell ERA5 covers.

The key publication for practitioners is "Contrasting trends in very large hail events and related economic losses" (Battaglioli et al., Nature Geoscience, 2025). Three main findings for Europe and Germany:

• From 1950 to 2021, the ERA5-AR-CHaMo models show significant increases in hail across large parts of Europe, primarily driven by rising moisture in the lowest atmospheric layers.
• The strongest increase worldwide occurs in northern Italy: very large hail (≥ 5 cm) is now (2012 bis 2021) roughly three times more frequent than in the 1950s.
• For Europe overall, the strongest global rise in very large hailstones is observed, in stark contrast to the southern hemisphere, where decreases are documented.

A methodologically important sub-result: the CAPE energy above the minus 10 °C isotherm (not classical CAPE) proved to be a universally superior predictor for large hail. Anyone making hail statements in a climate context cannot get around this parameter anymore.

ERA5 itself is the fifth generation of global climate reanalyses from the European Centre for Medium-Range Weather Forecasts (ECMWF). Hourly data points since 1940, 31 km grid, over 100 atmospheric levels. Unlike pure observation networks, ERA5 combines physical models with historical measurements and is spatially complete. That is what makes it valuable for hail climatology in regions with sparse observational coverage.

KIT studies, Frontiers 2026 and Kahraman et al.

German hail climatology is essentially carried by the IMK-TRO at KIT Karlsruhe. Three key studies:

• Kunz, Sander & Kottmeier (2009, Int. J. Climatol.): Analysis of hail damage days in Baden-Württemberg 1974 bis 2003. Significant increase in damage days despite constant thunderstorm frequency. Convection indices based on radiosonde data explain over 55 per cent of the annual variance.
• Mohr & Kunz (2013, Atmos. Res.): "Recent trends and variabilities of convective parameters relevant for hail events in Germany and Europe". Confirms that near-surface temperature and humidity indices show positive trends, whereas indices from higher atmospheric layers are neutral or negative.
• Puskeiler, Kunz & Schmidberger (2016, Atmos. Res.): First comprehensive hail statistics for Germany based on C-band radar data 2005 bis 2011. Clear north-south gradient, several hotspots in the lee of low mountain ranges.

Frontiers in Environmental Science (2026, KIT/climXtreme): 15,577 potential hail tracks from 20 summer half-years (2005 bis 2024) on a 3D C-band radar dataset. Spatially a clear north-south gradient. Highest hail frequency south of Stuttgart and in the Bavarian Alpine foothills. Temporally no significant national trend. Spatially differentiated: significant increase along the Baden-Württemberg/Bavaria border and south-east of Munich, significant decrease across large parts of northern and western Germany. Scandinavian blocking patterns offer a teleconnection explanation.

The seemingly contradictory counter-finding comes from Kahraman et al. (Nature Communications, September 2025), Met Office and British universities. Using km-scale climate simulations under a high-emission scenario (RCP 8.5, +5 °C), the picture for central Europe shifts:

• Severe hail (≥ 2 cm) may become rarer, because a higher freezing level and weaker large-scale weather patterns inhibit hail formation.
• Very large hail (≥ 5 cm) may still increase, through more tropical storm types with very strong updrafts, particularly in southern Europe.
• For central Europe and Germany this means, under an extreme scenario, regionally fewer but more intense events.

These findings do not contradict the historical ERA5 trends, they describe a possible future under strong warming. For site-level risk assessment both views belong in the analysis, because they address different time horizons.

Why CMIP6 cannot project hail directly and which proxies help

CMIP6 global models typically have horizontal resolutions of 50 to 150 km and use parameterised convection. Hail cells form on horizontal scales of 1 to 20 km within minutes, far below grid cell size. Three consequences:

1. Convection parameterisation cannot realistically represent either hail growth or melting processes during fall.
2. Spatial resolution is insufficient for orographically triggered convergence (Black Forest, Swabian Alb, Alpine foothills).
3. Hail statistics are not directly output by CMIP6.

Those who still need climate projections for hail relevance work with proxies:

The literature converges: a combination of CAPE-based instability and wind shear reproduces historical hail climatologies better than single parameters. AR-CHaMo outperforms most composites at medium and long forecast horizons. For an ESRS-E1-9 assessment the choice matters, because it carries the justification for the uncertainty range. Anyone diving deeper into the difference between direct projections and proxy-based statements will find the methodological frame in the CORDEX vs. CMIP6 data choice guide and the appropriate methodology framework in ISO 14091 in practice for banks and insurers.

Datasets for site-level assessment

Four data sources are practically relevant for German industrial sites:

• ESSL European Severe Weather Database (ESWD): Over 310,000 quality-controlled reports for large hail, gusts, tornadoes and heavy precipitation. Operational since 2006, with historical data before. DWD is a cooperation partner. Limitation: inhomogeneous report density, population-dependent, lower coverage in northern Germany and rural regions. Access via ESSL membership.
• DWD radar network: 17 C-band Doppler radars, complete national coverage. A homogenised archive dataset since 2005, basis of the KIT studies. Available as RADOLAN/RADKLIM for precipitation, HAIL-Composite for hail probability, and convection indices from radiosonde stations in Essen, Meiningen, Munich and Lindenberg.
• GDV Naturgefahrenreport: Insured loss data since 1973, price-adjusted, with regional breakdown. Around 70 tables per annual report. The GDV actively supports research access.
• Insurer-side sources: Insurers such as Verti publish annual hail atlases with state and city rankings. Usable for first orientation, too unsystematic for ESRS reporting.

Radar observation and modelled probability answer different questions, and conflating them is the standard error in site dossiers. DWD radar and the ESWD archive record what happened, with a date and a place, but only inside network coverage and back to the homogenised archive. AR-CHaMo records nothing: it estimates how probable large hail was per grid cell and hour back to 1950, including cells no observer stood in, and never proves a stone fell on site. The gap between them is the honest uncertainty range.

Which combination a site needs is a question of scale. For a single plant, ESWD density plus the radar climatology of the enclosing cell, read against the AR-CHaMo trend sign, stays a desk exercise. Across a portfolio it breaks: ESWD density is population-dependent, so a Ruhr-area plant and one in Upper Swabia are not on equal terms. I rank portfolios on the radar climatology alone and use ESWD only on the extremes.

Hail hotspots in Germany

The regional evidence base from radar, insurance and model data:

The strongest evidence rests on the 20-year DWD radar dataset, analysed in Frontiers (2026). The north-south gradient is consistent across all studies. The significant increase along the BW/Bavaria border is a newer finding with high statistical robustness. For site managers this means: an industrial site choice in the Allgäu or south of Munich should account for the hail trend with a 30 to 40 per cent higher probability of very large hailstones by the end of the century (under a moderate warming scenario).

GDV loss statistics 2023 to 2025

Three consecutive years with marked variation:

• 2023: Total natural hazard losses 5.6 billion euros. Of which storm/hail in property insurance 2.7 billion, in motor 1.3 billion (fourth highest value ever). Hotspots in June (740 million) and August (1.5 billion). In Kassel more than every third comprehensive-insured vehicle was damaged, the highest claim frequency since 1984.
• 2024: 5.7 billion overall, storm/hail property 1.8 billion. Highest state losses in Bavaria and Baden-Württemberg, each around 1.6 billion (including the June flooding).
• 2025: Marked decline. Property natural catastrophe losses 1.4 billion nationally, 3 billion less than 2024. Of which storm/hail 1.0 billion. In Bavaria only 118 million, less than a tenth of the previous year.

Regional consistency: Baden-Württemberg, Bavaria and Thuringia lead the hail loss ranking in almost every year. Northern Germany (Mecklenburg-Vorpommern, Schleswig-Holstein, Bremen) sits regularly at the lower end. Average loss per event in BW around 3,300 euros, Bavaria around 3,000, both clearly above the national average.

Peak years such as 2013 or 2021 (Europe-wide) caused hail losses in Germany in the double-digit billion range. Northern Italy reached a record single-event loss of 6 billion US dollars in 2023, an impressive confirmation of Battaglioli's trend findings.

ESRS-E1-9-compliant assessment path for hail

ESRS E1-9 requires the quantification of financial effects of material physical climate risks on assets, separated into acute and chronic risks, with affected asset values disclosed across short-, medium- and long-term time horizons. ESRS E1-2 (AR 10 bis 13) requires a climate scenario analysis as the basis. Where direct projections are missing, the framework explicitly allows analogies, proxy indicators and expert judgement, provided the approach is methodologically justified and transparently documented.

For hail at European industrial sites a five-step path:

1. Historical hail exposure class from geocoding against the ESWD density map, DWD radar hail climatology (Puskeiler/Kunz 2016, Frontiers 2026) and GDV regional data. Output: one of the five exposure classes below.
2. Trend adjustment by overlay with the AR-CHaMo ERA5 trend for the region (1950 bis 2021). Southern Germany: trend amplifier. Northern Germany: trend dampener or neutral.
3. Proxy scenario analysis using CAPE and Lifted Index trends from CMIP6 for the chosen SSP scenario, as a proxy for future hail exposure change (mid-term to 2050, long-term to 2100).
4. Financial exposure across building assets, outdoor infrastructure (PV modules, cooling plants, outdoor storage), business interruption risks. Linked to insurance cover and deductibles.
5. Documentation for E1-9 with justification of the datasets (ERA5/AR-CHaMo, ESWD, GDV), the uncertainty range and references to peer-reviewed literature. Explicitly state that direct CMIP6 hail projection is methodologically unavailable and why proxies are adequate.

Step 1 is where the literature stops: no AR-CHaMo paper and no radar climatology says from which frequency a site counts as exposed. That is a reporting question. The five classes I work with:

The numeric bands behind the classes, hail days and events per year, come from my own threshold research across 13 climate parameters and 8 hazard categories, calibrated against DWD Klimaatlas, UBA KWRA 2021, IPCC AR6, GERICS and EURO-CORDEX. No source document defines a hail threshold, so the bands are mine, and they belong in the disclosure with their parameter and reference period.

The path transitions cleanly into the ISO 14091 in practice methodology and into insurance and coverage gaps. Hail stays a methodological special case for as long as the direct projection is missing. Nobody building an audit-proof E1-9 assessment can shortcut that with "hail projection 2050"; the limits have to be explicit, and auditors reward that. Walking the five steps for one named site is what a climate risk analysis does.

Frequently asked questions

Further resources

• Climate risk analysis for companies: Methodology to ISO 14091
• ISO 14091 in practice: What banks and insurers expect
• CORDEX EUR-11 vs. CMIP6: Climate data choice for risk analysis
• RCP and SSP climate data: A full guide
• Climate risks and insurance gaps
• Physical vs. transition climate risks

Primary sources:

• Rädler, A. T., et al. (2018) and Battaglioli, F., et al. (2023): AR-CHaMo (Additive Regression Convective Hazard Model), European Severe Storms Laboratory. Model description in Natural Hazards and Earth System Sciences 23, 3651 (2023), Copernicus.
• Battaglioli, F., et al.: Contrasting trends in very large hail events and related economic losses. Nature Geoscience, 2025. Kahraman et al., Nature Communications, September 2025. KIT / climXtreme, Frontiers in Environmental Science, 2026.
• Data: ESWD (ESSL), DWD radar climatology (RADKLIM), GDV Naturgefahrenreport.