The study, led by Tapas Ranjan Patra of Rajendra University with colleagues from Utkal University and Netaji Subhash Chandra Bose College, drew on daily maximum and minimum temperature records from the India Meteorological Department’s gridded archive covering 1951 to 2024. Rather than averaging the state into a single number, the team preserved spatial detail by working at the 1-degree by 1-degree grid-cell scale. Of 30 grid cells overlapping Odisha’s boundary, 28 had complete enough records to enter the modelling stage, a 93.3 percent spatial coverage that allowed coastal, interior and transitional zones to be assessed within one common framework.

The statistical engine of the study was the generalized extreme value distribution, the standard tool for modelling the tails of annual or seasonal maxima and minima. For each grid cell and each extreme series, the researchers fitted five nested candidate models: a stationary model with a fixed location parameter, a model in which the location shifts linearly through time, and three models that added the El Niño-Southern Oscillation index, the Indian Ocean Dipole index, or both as covariates. Bayesian inference, implemented in PyMC with four Hamiltonian Monte Carlo chains and strict convergence diagnostics, propagated uncertainty from parameter estimation all the way through to the return-level estimates, producing posterior intervals rather than single point values.

Diagnostic discipline was central to the workflow. A fit was deemed clean only if it showed finite parameters, a maximum R-hat of 1.01 or less, effective sample sizes of at least 400, zero divergences and no support violations. Among clean candidates, the model with the highest expected log predictive density was preferred, with a parsimony rule favouring simpler structures when candidates competed within two ELPD units. The team completed 767 fits in total, including rescue refits for difficult cases, and every one of the 140 selected grid-series combinations ultimately yielded numerically valid return levels, with five cold-night cases retained under explicit model cautions.

One of the clearest findings concerns the timing of daytime heat. The annual maximum daily maximum temperature and the March-to-June maximum proved almost indistinguishable: they shared the same mean of 41.4 degrees Celsius, the same median, and the same observed range from roughly 35 to 46.4 degrees. In 99.95 percent of valid grid-years, the annual extreme fell within the extended pre-monsoon heat season. This confirms that Odisha’s hottest days are not scattered through the year but concentrated in a well-defined window when high insolation, dry soils and anticyclonic circulation combine to push daytime temperatures toward their limits.

The headline number for heat planners is the recent 50-year return level for pre-monsoon daytime maxima: a spatial median of 43.81 degrees Celsius, with grid-level medians spanning 41.60 to 46.52 degrees. Yet the study is careful about what its zero-change results mean. In 27 of 28 grids, the selected model was stationary, which by construction forces the early-to-recent return-level change to be exactly zero. The authors stress that these zeros are structural properties of the chosen models, not independent evidence that daytime heat has stopped evolving. A separate sensitivity analysis that imposed a time trend uniformly across grids found small positive changes in daytime extremes, though none large enough for the 95 percent credible intervals to exclude zero.

The warm-night story turned out to be strikingly different. Night-time extremes showed much stronger support for time-varying models, and the direction of change was negative: recent return levels for monsoon-season warm-night maxima were lower than early-period values, with the median 50-year return level falling from 30.82 to 30.35 degrees Celsius. In a sensitivity restricted to data through 2007, negative time coefficients persisted in every eligible grid, 12 of 12 for annual warm nights and 19 of 19 for monsoon-season warm nights. The authors caution that this does not mean Odisha is cooling; rather, it reflects seasonally specific behaviour of upper-tail minimum temperatures, consistent with earlier evidence of negative monsoon-season minimum-temperature trends over eastern India, possibly linked to cloud cover, humidity and land-sea influences.

Cold-night extremes in winter behaved differently again, with near-zero and mixed changes and return levels that were more sensitive to how boundary grid cells were handled. Meanwhile, teleconnection covariates earned only a limited role. ENSO appeared in 16 selected models, the Indian Ocean Dipole in 9, and both together in just 6, out of 140 selections, and their coefficient evidence was mixed rather than uniformly directional. The authors interpret this cautiously: large-scale climate modes may modulate Odisha’s extremes only in specific grid cells and seasons, and the data did not justify a state-wide teleconnection narrative.

What emerges is a portrait of thermal risk that refuses to be flattened into a single warming trend. Daytime heat remains anchored in the pre-monsoon season at physically consequential magnitudes, warm-night tails behave according to their own seasonal logic, and cold nights show little coherent change. The authors emphasize that their return levels are conditional summaries of observed tail behaviour, not deterministic forecasts of events arriving once every 50 calendar years. They also note the limits of 1-degree gridded data for urban or district-scale planning, and the fact that their models were fitted independently rather than through a spatial hierarchy.

Even with those caveats, the study offers something regional planners rarely get: a defensible, uncertainty-bounded benchmark of how rare hot days, warm nights and cold nights have actually behaved across Odisha’s landscape. The authors suggest that future work could add spatial hierarchical modelling, compare these observational benchmarks with CMIP6 and CORDEX projections, and test physically grounded covariates such as soil moisture, monsoon onset and humidity-linked heat stress. For a state where heatwaves correlate significantly with recorded mortality, knowing the magnitude and uncertainty of the rarest thermal events, grid cell by grid cell, is a practical foundation for the climate-risk decisions that lie ahead.

Subject of Research: Bayesian extreme-value assessment of observed temperature extremes and return levels across Odisha, India

Article Title: Bayesian assessment of return levels for observed temperature extremes across Odisha in India

Article References: Bayesian assessment of return levels for observed temperature extremes across Odisha in India. (n.d.). https://doi.org/10.1007/s44288-026-00777-y

Image Credits: AI Generated

DOI: 10.1007/s44288-026-00777-y

Keywords: Bayesian inference, extreme value theory, generalized extreme value distribution, temperature extremes, return levels, Odisha, India, heatwaves, ENSO, Indian Ocean Dipole, climate risk, IMD gridded data

The study, published in Theoretical and Applied Climatology, analyzes 124 years of daily rainfall observations, from 1901 to 2024, using the India Meteorological Department’s gridded dataset at a resolution of 0.25 degrees of latitude and longitude. Rather than pooling all extreme rainfall days together, the researchers separated the annual maximum one-day rainfall totals into two climatologically distinct seasons: the southwest monsoon from June to September, known in the region as JJAS, and the northeast monsoon from October to December, abbreviated OND. This separation matters because Tamil Nadu sits in the rain shadow of the Western Ghats during the summer monsoon, while the retreating northeast monsoon, fueled by moisture-laden easterly winds and tropical cyclones in the Bay of Bengal, delivers the state’s most destructive downpours.

The statistical machinery behind the analysis is a Bayesian hierarchical generalized extreme value model, or GEV model, fitted with the INLA-SPDE computational approach. The GEV distribution is the standard tool of extreme value theory: it describes the distribution of block maxima, in this case the largest single-day rainfall total in each season of each year, using three parameters that control the location, scale, and shape of the tail. What makes the Bayesian hierarchical framework powerful is that these parameters are allowed to vary smoothly across space, linked through a latent Gaussian field approximated by the stochastic partial differential equation method, so that neighboring grid cells inform one another and uncertainty is propagated coherently across the entire state. The result is not a single statewide estimate but a map of return levels, each accompanied by a full posterior distribution that quantifies how confident the model is at every location.

The headline finding is a dramatic seasonal asymmetry. At the median of the distribution, one-day rainfall maxima during the northeast monsoon are approximately 1.4 times larger than those of the southwest monsoon. In the upper tail, where the most dangerous events live, the asymmetry widens to a factor of 1.8. The spatial median 50-year return level, the rainfall depth expected on average once in 50 years, is 123.1 millimeters per day for the southwest monsoon but 220.5 millimeters per day for the northeast monsoon. In other words, the season that most engineering studies have historically treated as the primary hazard for peninsular India is, for Tamil Nadu, the milder of the two. The northeast monsoon, shorter and less famous, is unambiguously the state’s dominant extreme-rainfall engine.

One of the study’s most methodologically interesting choices is its treatment of nonstationarity. Many recent studies impose time trends or climate covariates on extreme rainfall models as a matter of course, assuming that a warming world has already shifted the distribution. Here, the authors took a more disciplined approach: nonstationarity in the location parameter, expressed through time, the El Niño-Southern Oscillation, and the Indian Ocean Dipole, was treated as a testable model feature rather than an assumption. Covariates were only retained if the data and validation metrics supported them. This matters because overfitting a trend to 124 years of noisy extremes can produce return levels that look climate-informed but perform poorly out of sample.

The validation strategy is unusually thorough for a regional frequency analysis. Temporal hindcast validation tested whether the model, fitted to earlier portions of the record, could predict extremes in later periods. Spatially blocked cross-validation removed contiguous regions from the fit and asked how well the model transferred to them, a sterner test than leaving out isolated points because it probes whether the spatial structure generalizes. Pre-specified sensitivity analyses, including a late-epoch risk-inflation test restricted to 1961 through 2024, examined whether the full-record estimates might understate present-day risk. These checks revealed genuine limitations. The southwest monsoon tail was found to be under-calibrated: the nominal 50-year level was exceeded at an empirical rate of 0.029, corresponding to an effective return period of only about 34 years. And spatial transferability was regionally heterogeneous, particularly for the northeast monsoon, meaning the maps are more reliable in some districts than others.

The stationarity question produced the study’s most nuanced conclusion. Across the full 124-year record, risk inflation factors, ratios comparing nonstationary or recent-period estimates to the stationary benchmark, came out close to unity, suggesting no dramatic long-term shift. But when the analysis was restricted to the late epoch from 1961 onward, the inflation factors rose, with median 50-year risk inflation values of 1.076 for the southwest monsoon and 1.038 for the northeast monsoon. The authors’ interpretation is careful and worth emphasizing: the full-record stationary result should be read as a benchmark, not as definitive evidence that historical design standards remain adequate today. A seven to eight percent inflation in the recent period may sound modest, but for infrastructure with a design life of 50 to 100 years, even small shifts in the tail compound into materially different flood probabilities.

Why does this matter beyond the statistics? Tamil Nadu has repeatedly experienced catastrophic urban flooding, most visibly in Chennai, where the northeast monsoon of 2015 delivered record-breaking deluges that paralyzed the metropolis. Frequency analyses that blend the two monsoons, or that borrow return levels calibrated for summer-monsoon-dominated regions of India, would systematically underestimate the depth of the design storm that the state’s drainage and reservoir systems must withstand. The finding that the northeast monsoon tail is nearly twice as heavy as the southwest monsoon tail provides a quantitative, uncertainty-quantified basis for seasonally differentiated design criteria, and it aligns with a growing body of evidence that widespread extreme precipitation during the northeast monsoon over south peninsular India has been increasing.

The study is also candid about what its outputs are not. Because the analysis rests on 0.25-degree gridded data, which averages rainfall over cells roughly 25 to 28 kilometers across, the resulting return-level surfaces represent areal-equivalent extremes rather than point rainfall at a specific gauge. Gridded products tend to smooth out the most intense convective cells, so the maps should be understood as spatially coherent hazard-screening surfaces, useful for identifying where risk is concentrated and where station-level analysis is most urgently needed, rather than as ready-made engineering inputs. The authors explicitly recommend local gauge or catchment-scale verification before any design use, a caveat that reflects a broader tension in climate science between the spatial completeness of gridded datasets and the fidelity of point observations.

What the study ultimately delivers is a template. By separating climatologically distinct seasons, treating nonstationarity as a hypothesis to be tested rather than a narrative to be assumed, validating with hindcasts and blocked cross-validation, and publishing uncertainty alongside every estimate, it models the kind of rigor that regional extreme-rainfall assessments will need as climate change reshapes the tails of the distribution worldwide. For Tamil Nadu, the message is concrete: the state’s flood future is written in October through December, and any risk framework that ignores the asymmetry between its two monsoons is measuring the wrong storm.

Subject of Research: Seasonal asymmetry in extreme daily rainfall and Bayesian spatial return-level estimation over Tamil Nadu, India

Article Title: Seasonal asymmetry in extreme daily rainfall over Tamil Nadu: Bayesian spatial return levels and a stationarity benchmark

Article References: Patra, T. R., Pathy, A. C., Sk, M. M., Safder, A., & Khatun, R. (2026). Seasonal asymmetry in extreme daily rainfall over Tamil Nadu: Bayesian spatial return levels and a stationarity benchmark. Theoretical and Applied Climatology, 157(11), Article 699. https://doi.org/10.1007/s00704-026-06618-9

Image Credits: AI Generated

DOI: 10.1007/s00704-026-06618-9

Keywords: Tamil Nadu, extreme rainfall, northeast monsoon, southwest monsoon, return levels, generalized extreme value, Bayesian hierarchical model, INLA-SPDE, nonstationarity, ENSO, Indian Ocean Dipole, flood risk

The mathematical foundation of the study lies in extreme value theory, a branch of statistics developed specifically to model the behavior of rare events rather than averages. The Generalized Extreme Value, or GEV, distribution works by taking the single largest daily rainfall value from each year and fitting a flexible probability curve to those annual maxima. That curve has three variants named after their discoverers: the Gumbel family, with a moderately heavy tail; the Fréchet family, whose unbounded tail allows for extraordinarily large extremes; and the Weibull family, whose tail is bounded, meaning there is a physical ceiling on how extreme an event can become. Which variant applies in a given region is governed by the shape parameter, and estimating it correctly is the crux of any return-period calculation, because the tail behavior determines how quickly probabilities decay as you project toward rarer and rarer storms.

The Peak Over Threshold, or POT, approach takes a fundamentally different view of the data. Instead of discarding all but one value per year, it captures every daily rainfall that exceeds a chosen high threshold, thereby using far more information about the upper tail of the distribution. The exceedances are then modeled with the Generalized Pareto distribution, a technique rooted in Picklands’ foundational 1975 work on extreme order statistics. The trade-off is well known among hydrologists: POT extracts more signal from limited records and typically yields narrower uncertainty intervals, but its results are sensitive to how the threshold is selected, whereas GEV discards potentially useful secondary storms within a single year yet is simpler and more standardized in practice.

To capture India’s enormous climatic diversity, the authors divided the country into six homogeneous regions: north, northeast, west, east, central, and peninsular India. This regional partition matters because the drivers of extreme rainfall differ dramatically across the subcontinent, from the orographic uplift along the Western Ghats and the Himalayan foothills to the depression-driven systems that sweep across the Gangetic plains. The rainfall data came from the gridded daily product maintained by the India Meteorological Department, providing a spatially consistent record spanning nine decades, a period during which global warming, rapid urbanization, and land-use change have all reshaped the hydrological cycle.

The shape parameter results reveal a striking geographic split. In Northeast India, one of the wettest places on the planet, the fitted GEV shape parameter was approximately zero, which corresponds to an effectively Gumbel-type tail: extremes grow with rainfall intensity but without the explosive amplification of a heavy-tailed Fréchet distribution. In the other five regions, the shape parameter was negative, indicating bounded Weibull-type tails in which the largest conceivable daily rainfall approaches a finite upper limit. For flood engineers, this distinction is far from academic. A bounded tail implies that 100-year and 1000-year return levels converge toward a ceiling, while a heavier tail keeps pushing design storms upward as the return period lengthens, with direct consequences for the safety margins built into critical infrastructure.

The estimated 100-year return levels quantify just how uneven India’s extreme rainfall landscape is. In western India, the GEV-based 100-year daily rainfall came out at roughly 358 millimeters, with a 95 percent bootstrap confidence interval spanning 308 to 409 millimeters. In northeastern India, the corresponding figure soared to about 1107 millimeters, with an interval of 923 to 1353 millimeters, meaning that in the worst plausible case a single day’s rain in the northeast could exceed the entire annual rainfall of many semi-arid regions. These numbers translate directly into design guidance: a culvert sized for western Indian extremes would be dangerously undersized if transplanted to the Brahmaputra valley, and the wide confidence intervals in the northeast underscore how much statistical uncertainty remains even with ninety years of data.

When the two modeling philosophies were compared directly, the POT method emerged as the more accurate fit in every one of the six regions. Its 100-year return level estimates ranged from about 357 millimeters to 1045 millimeters, closely tracking the GEV values but generally with tighter precision, a benefit attributable to its richer use of threshold exceedances. This finding aligns with a growing body of international literature suggesting that threshold-based models, when their thresholds are chosen carefully, offer the best of both worlds: the theoretical rigor of extreme value theory combined with the statistical efficiency of using multiple extreme events per year rather than just the annual maximum.

Perhaps the most methodologically important contribution of the study is its insistence on out-of-sample validation. The authors split the record into a training period from 1930 to 1999 and a testing period from 2000 to 2019, then evaluated how well models calibrated on the historical data predicted the extremes of the most recent two decades, using the negative log-likelihood as the scoring metric. The results exposed regional differences in predictive calibration that would have remained invisible under conventional in-sample goodness-of-fit testing. A model can fit the data it was trained on beautifully and still misjudge the tail behavior of the future, and the study demonstrates that out-of-sample evaluation is a genuinely better guide to real-world flood risk than fit statistics computed on the same data used for estimation.

The practical stakes of this work are considerable. Floods rank among the most frequent and destructive natural hazards worldwide, threatening lives, ecosystems, infrastructure, and economic stability, and their severity is being amplified by climate change, urbanization, and hydrological variability. Previous research, including landmark analyses of increasing extreme rain events over India in a warming environment, has documented that the character of Indian rainfall extremes has been shifting, which makes the choice of statistical model even more consequential: an outdated or poorly validated tail model can silently understate the risk that a dam, levee, or urban drainage network will be overwhelmed. By providing a regionally consistent assessment that explicitly separates model fit from predictive skill, the study gives Indian water managers a clearer map of where each tool can be trusted.

The research also opens doors for future refinement. The analysis assumes stationarity, treating the 1930 to 2019 record as drawn from an unchanging climate, whereas a growing literature on non-stationary extreme rainfall over Indian cities suggests that allowing distribution parameters to evolve with time or with climate covariates could sharpen return-level estimates further. Extending the out-of-sample framework to non-stationary models, to finer spatial resolutions, and to sub-seasonal extremes would be natural next steps. For now, the message of this study is clear and actionable: for diagnosing extreme precipitation across India, the Peak Over Threshold method offers the most accurate regional fits, the northeast demands special caution because of its extraordinary rainfall magnitudes and wide uncertainty bands, and no flood risk assessment should rely on in-sample fit alone when the true test is how well a model anticipates the storms of the next twenty years.

Subject of Research: Extreme value analysis of extreme precipitation and return levels across India using GEV and POT methods

Article Title: Assessing the suitability of GEV and POT for extreme precipitation diagnostics over India

Article References: Jain, J., Muñoz-Arriola, F., & Khare, D. (2026). Assessing the suitability of GEV and POT for extreme precipitation diagnostics over India. Theoretical and Applied Climatology, 157(11), Article 693. https://doi.org/10.1007/s00704-026-06643-8

Image Credits: AI Generated

DOI: 10.1007/s00704-026-06643-8

Keywords: extreme precipitation, extreme value theory, Generalized Extreme Value distribution, Peak Over Threshold, return levels, flood risk, India, monsoon, Generalized Pareto distribution, out-of-sample validation, hydroclimatology, Theoretical and Applied Climatology

The research, led by Alexander Lee Rischmuller of the Research Unit Sustainability and Climate Risk at Universität Hamburg, together with Benjamin Poschlod and Jana Sillmann, focuses on southern Germany, a region whose complex topography makes it a demanding proving ground for rainfall statistics. The team drew on the Canadian Regional Climate Model Large Ensemble, known as CRCM5-LE, a fifty-member set of regional climate simulations dynamically downscaled to a resolution of roughly 12.5 kilometres. Because each of the fifty members differs only by tiny perturbations in its initial conditions, the spread among them represents the chaotic internal variability of the climate system itself. Stacking all fifty realizations for the period 1980 to 2019 yields a homogeneous, gap-free sample of two thousand simulated years, an archive of extreme rainfall that no real weather station could ever hope to match.

That archive served as what statisticians call a perfect model experiment. Rather than asking whether the climate model perfectly reproduces reality, the researchers treated the full two-thousand-year ensemble as the ground truth. From it they computed what they term effective return levels, for example the hundred-year rainfall intensity represented by the twentieth-highest annual maximum in the entire sample. They then carved the archive into artificial sub-samples of thirty to one hundred years, mimicking the typical length of observational records for sub-daily rainfall, and asked each candidate model to reproduce the true hundred-year return level from these fragments. Each sub-sample size was drawn one hundred times, generating a rigorous test of accuracy and robustness across twenty-five grid cells and six rainfall durations ranging from one to forty-eight hours.

The statistical backbone of the exercise is the Generalized Extreme Value distribution, the canonical tool of extreme value theory. When rainfall maxima are sampled in blocks, typically annual maxima, their distribution converges to the GEV, which is described by three parameters: location, scale, and shape. The shape parameter is the critical one, because it governs the behaviour of the distribution’s tail and therefore the estimated magnitude of rare events. Small errors in the shape parameter can dramatically distort estimates of the hundred-year return level, and the problem worsens as sample sizes shrink. The study’s central innovation lies in how the shape parameter is handled within a duration-dependent GEV model, which pools data across all durations into a single coherent framework rather than fitting each duration separately and risking physically inconsistent curves.

The six models tested spanned the full spectrum of current practice. Two frequentist approaches, a standard GEV fitted by L-moments and a duration-dependent GEV fitted by maximum likelihood, represented the state of the art. Three Bayesian hierarchical variants of the duration-dependent GEV followed, differing in whether the shape parameter was allowed to vary across durations, across space, or across both. A fourth, non-hierarchical Bayesian model completed the set. Hierarchical models share information across dimensions through hyperpriors, a mechanism known as partial pooling that shrinks individual estimates toward a common mean and stabilizes inference when data are sparse. All Bayesian models were fitted with the No-U-Turn sampler, an efficient form of Hamiltonian Monte Carlo implemented in the Stan framework.

The verdict was unambiguous. The hierarchical model with a shape parameter that varies with rainfall duration but remains fixed across space, abbreviated dGEV-BHM-ξ(d), delivered the highest accuracy and the narrowest confidence intervals across nearly every test. When fitted to a thirty-year sample, the typical situation for sub-daily rainfall records in Germany, it cut the relative error of the hundred-year return level to 8.8 percent, compared with 18.1 percent for the standard L-moments GEV. Its return level concordance, the fraction of modelled hundred-year estimates falling within the confidence interval of the true ensemble value, reached 27 to 35 percent, roughly ten percentage points above every competitor. The most flexible model, which let the shape parameter vary over both space and duration, performed worst, because its extra complexity could not be supported by samples of one hundred years or fewer.

The physical insight behind the winning configuration is telling. The fitted shape parameters revealed heavier tails for short convective bursts, with values around 0.15 to 0.2 for one- to three-hour storms, declining to roughly 0.06 for forty-eight-hour events. This duration dependence matches earlier global surveys of rainfall tails, while the spatial constancy of the shape parameter within a small region echoes previous findings that neighbouring sites share similar tail behaviour. By fixing the shape parameter across space, the model reduces the number of free parameters, suppresses overfitting, and shields the tail estimate from unlucky sub-samples dominated by internal climate variability. The resulting intensity-duration-frequency curves, demonstrated for a locality near Würzburg, reproduced the true hundred-year return levels across all durations, whereas the maximum-likelihood benchmark overestimated rainfall for durations of six hours and longer.

Perhaps the most consequential finding, however, concerns a diagnostic tool that practitioners rely on routinely: the Anderson-Darling goodness-of-fit test. Across 720,000 hypothesis tests, the researchers found that the test’s verdicts aligned with model flexibility rather than predictive skill. The most flexible models passed the test most often, while the best-performing hierarchical model was rejected most frequently. The reason is a classic bias-variance trade-off. A flexible model can overfit the quirks of a small sample, earning a passing grade on the fit test while generalizing poorly to the rare quantiles that actually matter for risk assessment. A model that passes an Anderson-Darling test on a thirty-year record, the authors conclude, is not necessarily capable of predicting very rare probabilities of occurrence. For climate-risk applications, the Akaike information criterion proved far better aligned with true predictive skill.

The implications reach well beyond southern Germany. Because the recommended model requires only sub-daily precipitation data from several stations with similar characteristics, it is directly applicable to the short, sparse records that dominate observational hydrology, and to high-resolution climate simulations that typically cover only a few decades. The framework could sharpen the rainfall return levels used to design stormwater systems, inform catastrophe models, and support adaptation planning in a warming climate where extreme rainfall is intensifying. The authors acknowledge limitations: the models assume stationarity, which is reasonable for the 1980 to 2019 window they examined but would need non-stationary extensions for climate projections, and the spatially constant shape parameter restricts the approach to relatively homogeneous regions, though clustered applications or covariate-based extensions could widen its reach. Bayesian computation is also more demanding than L-moments. Yet as floods grow more severe and design standards come under scrutiny, the study makes a compelling case that borrowing strength across space and duration, rather than squeezing every drop of flexibility from scarce data, is the more reliable path to knowing how hard the rain of the next century will fall.

Subject of Research: Statistical estimation of extreme precipitation intensity-duration-frequency curves using Bayesian hierarchical models and a large climate model ensemble

Article Title: Bayesian hierarchical modelling of intensity-duration-frequency curves using a climate model large ensemble

Article References: Rischmuller, A. L., Poschlod, B., & Sillmann, J. (2026). Bayesian hierarchical modelling of intensity-duration-frequency curves using a climate model large ensemble. Advances in Statistical Climatology, Meteorology and Oceanography, 12(1), 1-19. https://doi.org/10.5194/ascmo-12-1-2026

Image Credits: AI Generated

DOI: 10.5194/ascmo-12-1-2026

Keywords: extreme precipitation, intensity-duration-frequency curves, Generalized Extreme Value distribution, Bayesian hierarchical model, climate model large ensemble, return levels, flood risk, shape parameter, Anderson-Darling test, southern Germany, internal climate variability, extreme value theory

The stakes are far from abstract. France’s offshore wind sector is projected to reach a cumulative capacity of 3.6 gigawatts by the end of 2027, driven by the commissioning of seven major projects. These turbines and their foundations are engineered to survive more than two decades in one of the most hostile environments on Earth, and their designs rest on a deceptively simple statistical assumption: that the statistics of the ocean, including the mean, the variance, and above all the extremes of wave height, remain stationary over time. Climate change breaks that assumption. Rising sea levels, migrating storm tracks, and shifting seasonal patterns mean that the wave climate a turbine faces in 2080 may be fundamentally different from the one measured during the historical record used to certify it.

At the heart of the study lies a class of statistical tools known as generalized extreme value, or GEV, models. These models describe the distribution of extreme values, such as the maximum wave height recorded in a given block of time, using three parameters: a location parameter that sets the central tendency of the extremes, a scale parameter that controls their spread, and a shape parameter that governs the behavior of the distribution’s tail, which determines how catastrophic the rarest events can become. Traditionally, engineers fit a GEV distribution to annual maxima, extracting the so-called 100-year return level, the wave height expected to be exceeded on average once per century. But annual maxima throw away most of the data, leaving only one point per year and producing wide, often impractically large, uncertainty bands.

Raillard and his team took a different route. Instead of annual maxima, they modeled monthly maxima, which multiplies the available data by a factor of twelve. The catch is that monthly maxima are not identically distributed, because wave climate is strongly seasonal. To handle this, the researchers turned to Generalized Additive Models, allowing the location and scale parameters of the GEV distribution to vary smoothly across the months of the year using spline functions, flexible curves fitted to the data. The shape parameter, notoriously difficult to estimate, was held constant. From the fitted monthly distributions, the team reconstructed the distribution of annual maxima by multiplying together the twelve monthly cumulative distribution functions, an operation justified by statistical tests showing no remaining time dependence in the model residuals once seasonality was accounted for.

The framework then went a step further into genuinely non-stationary territory. The researchers extended their model so that the seasonal cycles themselves evolve over time, using spline surfaces defined over the joint space of month and year. Cyclic P-splines captured the annual seasonality, while thin plate regression splines captured its long-term drift. Fitted independently to each of eight global climate models from the CMIP6 ensemble, under both a low-emission scenario (SSP1-2.6) and a high-emission scenario (SSP5-8.5), the model could shrink gracefully to a stationary form wherever the data showed no evidence of change, while flexibly tracking complex, non-linear evolutions elsewhere. Model selection was performed using the Akaike Information Criterion, and uncertainty was quantified through a Monte Carlo procedure simulating model parameters one thousand times.

The wave data themselves came from a numerical wave model, WaveWatch III, forced by the eight CMIP6 general circulation models, covering a historical period from 1985 to 2014 and a future window from 2071 to 2100 at three-hour temporal and half-degree spatial resolution. Because global climate models are too coarse to represent local conditions accurately, the team applied a statistical bias correction known as the CDF-t method, which maps the cumulative distribution of the global model onto that of a high-resolution local reanalysis, HYWAT for the English Channel and Atlantic and MED-WAV for the Mediterranean. The correction was applied month by month at six representative offshore sites chosen to span the diversity of the French seafront, and it markedly tightened the scatter of the climate models around the reanalysis reference.

The results paint a vivid picture of a changing wave climate. At the representative site in the Eastern English Channel, historical significant wave heights, the average height of the highest third of waves, ranged roughly between 3 and 7 meters. Under the high-emission scenario, that range widens to about 2 to 8 meters, with winter extremes pushing toward 8 meters and summer minima dipping near 2 meters. Perhaps most striking is the temporal redistribution: in most models the calm summer period grows longer and peaks later in the year, with the lowest wave heights shifting from June toward July, and January extremes exceeding December values by 0.2 to 0.5 meters. The seasonal cycle itself appears to be migrating toward the end of the year, hinting that conventional definitions of the seasons may need revision in a warming ocean.

When the researchers translated these shifts into design quantities, the ensemble average of the 100-year return level rose from approximately 6.15 meters in the historical period to 6.75 meters under SSP1-2.6 and 7 meters under SSP5-8.5 at the Eastern Channel site. Across all seafronts, the equivalent lifetime design level, the wave height that should not be exceeded over the operational life of a wind farm, increased almost everywhere: from about 7.4 meters historically to 8.2 meters under the high-emission scenario in the English Channel, and from 11 to 13 meters to 13 to 15 meters along the Atlantic coast. The North Atlantic site showed a clear increase of roughly 2.5 meters. The Mediterranean told a messier story, with weaker seasonal signals, much larger inter-model spread, and some physically implausible outliers, reaching simulated wave heights of up to 20 meters with uncertainties of 30 meters, likely reflecting coarse model resolution and poorly resolved coastal dynamics in that enclosed basin.

Methodologically, the payoff of using monthly maxima was substantial. Because the model is fitted to twelve times more data, the confidence intervals around the return levels narrowed dramatically compared with the classical annual-maxima approach, which the authors note can yield spurious results when data are scarce. Compared with an earlier study reporting an uncertainty of 13.15 meters for the 100-year return level under a comparable scenario, the new method achieved a maximum uncertainty of about 10 meters, a reduction the authors attribute to the larger effective sample size. The team also introduced a new definition of the design condition itself: rather than a fixed 100-year return level, they computed the quantile of the maximum wave height over the entire lifetime of the structure, calibrated so that the cumulative probability of failure over, say, thirty years matches the conventional stationary target.

The implications ripple outward from engineering offices to energy markets. Under-designed structures face reduced operational lifespans, escalating maintenance costs, and, in the worst case, catastrophic failure, while over-conservative designs inflate capital expenditure unnecessarily. By providing a statistically rigorous, uncertainty-aware way to translate climate projections into design loads, the framework offers engineers, designers, and stakeholders a tool for building offshore wind farms that remain both safe and economically viable across decades of environmental change. The authors see their work as a starting point rather than an endpoint, pointing toward refined wave modeling along the French coastline, parametric treatments of evolving seasonality, and joint analyses of wave height and peak period under future climates. As the ocean’s moods continue to shift, the mathematics used to tame them is shifting with it.

Subject of Research: Non-stationary extreme value analysis of projected changes in significant wave height for offshore wind farm design along the French coasts

Article Title: Non-stationary GEV models for estimating design sea-states in a changing climate – applications to offshore wind farms along the French coasts

Article References: Non-stationary GEV models for estimating design sea-states in a changing climate – applications to offshore wind farms along the French coasts. (n.d.). https://doi.org/10.5194/ascmo-12-195-2026

Image Credits: AI Generated

DOI: 10.5194/ascmo-12-195-2026

Keywords: offshore wind, significant wave height, GEV distribution, non-stationary extremes, CMIP6, climate change, return levels, French coasts, English Channel, Mediterranean Sea, extreme value analysis, wave climate

The research, conducted by Muhammad Aslam Mohd Safari and Tosiyuki Nakaegawa, analyzed daily rainfall records from 51 synoptic weather stations spanning the Japanese archipelago from 1901 to 2020. Rather than simply pulling out the single wettest day of each year, as many traditional analyses do, the team employed a peaks-over-threshold framework, a cornerstone of modern extreme value theory. In this approach, every day on which rainfall exceeds a carefully chosen high threshold counts as an extreme event, which means the method extracts far more information from the historical record than annual maximum sampling alone. The trade-off is complexity: extreme events often arrive in clusters, as when a stalled rain band lingers over the same region for days, and treating those clustered days as independent observations can badly distort the statistics.

That clustering problem is precisely where the new study makes its most distinctive contribution. The researchers applied a dependence-adjusted framework in which the degree of temporal dependence in the extremes, summarized by a quantity known as the extremal index, guided how events were declustered at each station. Depending on what the data demanded, they either left the record unblocked or grouped consecutive days into blocks of five, seven, or nine days, retaining only the peak within each block. This station-specific strategy reduces the artificial inflation of event counts that arises when a single multi-day storm contributes multiple threshold exceedances, while avoiding the over-smoothing that aggressive declustering can introduce. The extremal index estimates, derived from established methods for inference on clusters of extreme values, effectively let each station tell the analysts how much independence its extremes actually possess.

Choosing the threshold itself is one of the most delicate steps in any peaks-over-threshold analysis. Set it too low, and the theoretical model that justifies the analysis breaks down because moderate, non-extreme rainfall contaminates the sample. Set it too high, and so few events remain that uncertainty explodes. The study combined three complementary diagnostics to navigate this dilemma: mean residual life plots, which look for the range where average excess rainfall becomes approximately linear in the threshold; shape-parameter stability plots, which identify thresholds where the estimated tail behavior stops drifting; and return-level stability checks, which confirm that the headline quantities of interest hold steady across plausible threshold choices. This triangulated approach, drawing on a substantial methodological literature on threshold selection for extreme values, was applied independently at every one of the 51 stations.

Once thresholds were fixed and events declustered, the team fitted generalized Pareto distributions, the mathematical workhorse for modeling exceedances over high thresholds, whose tail shape parameter determines whether extremes are bounded, exponentially decaying, or heavy-tailed. Crucially, the researchers did not commit to a single fitting method. They compared maximum likelihood estimation, the standard approach that performs well with ample data, against estimation via L-moments, a robust technique based on linear combinations of order statistics that is often more stable when samples are small or outliers are influential. They also tested a simpler exponential alternative, which assumes a lighter tail. Model performance was then judged honestly: the record was split so that 80 percent of the data trained the models and the held-out 20 percent tested them, and the winning model varied from station to station. The lesson, the authors argue, is that no one-size-fits-all model exists for Japanese rainfall extremes, and station-by-station validation should become standard practice.

Before trusting any of these fits, the team also interrogated the assumption of stationarity, the idea that the statistical behavior of extremes has not changed over the 120-year record. They examined trends in both the annual frequency of threshold exceedances and the magnitude of the largest events. This matters because a warming climate is expected to intensify heavy precipitation, and several prior studies using large-ensemble regional climate simulations have projected robust increases in extreme rainfall over Japan under future warming scenarios. If the historical record itself shows drift, the classical extreme value models would need modification. The study found that while the stationarity assumption was broadly workable, it warrants caution at some locations, a caveat the authors flag explicitly alongside residual dependence detected at eight stations where the declustering did not fully eliminate serial correlation.

The headline outputs of the analysis are return levels: the rainfall magnitudes expected to be exceeded once every 5, 10, 50, 100, or 200 years on average, each accompanied by a 95 percent confidence interval. These numbers are the currency of infrastructure design and hazard zoning, and the study’s maps of them reveal a striking geographic pattern. The highest return levels concentrate in Kyushu and the Okinawa-Amami region in the country’s southwest, where tropical cyclones and the seasonal rain front deliver some of the heaviest daily totals on record. Yet the pattern is far from smooth. In one of the study’s most vivid illustrations of local contrast, the 100-year return level at Kochi, on the island of Shikoku, came out at roughly twice the corresponding value at Matsuyama, a station not far to the northwest. Such sharp gradients underscore how topography, proximity to moisture sources, and storm tracks conspire to make rainfall hazard an intensely local phenomenon.

How do these estimates stack up against official figures? The researchers compared their results with probable precipitation estimates published by the Japan Meteorological Agency and with values reported in earlier studies. The comparison showed broad spatial agreement, which is reassuring for the coherence of the underlying data and methods, but also revealed station-specific discrepancies. Some of those differences are expected, since the agency’s procedures and the study’s dependence-adjusted, station-specific modeling differ in sampling and fitting choices. The authors present the comparison not as a verdict on either approach but as evidence that return-level estimation carries genuine methodological sensitivity, and that consumers of these numbers should always look at the accompanying uncertainty rather than treating a single figure as gospel.

Indeed, uncertainty is a central character in this story. The confidence intervals around return levels widen substantially as the return period lengthens, which is an inherent feature of extreme value statistics: the further one extrapolates beyond the observed record, the less the data can constrain the answer. A 200-year return level estimated from 120 years of data rests on the tail behavior of the fitted distribution in a region where observations are sparse by definition. By reporting intervals alongside point estimates at every station, the study gives planners a defensible range rather than a false promise of precision. That honesty is particularly valuable in a country where the 2018 and 2020 disasters demonstrated, in human terms, exactly what the upper tail of the rainfall distribution can do.

The practical payoff of this work lies in its consistency and its caution. By applying a single, transparent, dependence-aware protocol across 51 stations and 120 years, the study delivers a common yardstick for rainfall hazard that regional authorities can use to compare risk across the archipelago, while the station-specific modeling respects the reality that Kochi and Matsuyama, or Kyushu and Hokkaido, are not interchangeable. The authors position the results as a historical baseline for region-specific rainfall-hazard planning, a foundation against which future projections from climate models can be anchored. As warming continues to load the dice toward heavier downpours, knowing precisely where the historical extremes are worst, and how uncertain those estimates remain, is a quiet but consequential advance in the science of staying dry when the sky does its worst.

Subject of Research: Statistical modeling of extreme daily precipitation extremes in Japan using dependence-adjusted peaks-over-threshold methods

Article Title: Dependence-Adjusted Peaks over Threshold Modeling of Extreme Daily Precipitation in Japan

Article References: Safari, M. A. M., & Nakaegawa, T. (2026). Dependence-Adjusted Peaks over Threshold Modeling of Extreme Daily Precipitation in Japan. Water Resources Management, 40(13), Article 532. https://doi.org/10.1007/s11269-026-04897-5

Image Credits: AI Generated

DOI: 10.1007/s11269-026-04897-5

Keywords: extreme precipitation, peaks over threshold, generalized Pareto distribution, extremal dependence, return levels, Japan, flood risk, threshold selection, L-moments, climate extremes, Water Resources Management, rainfall hazard

The research, led by Aaron Jerome Rampersad of the University of Canterbury with Christianne Marie-Claire Faith Zakour of the Loss and Damage Youth Coalition, addresses a problem that has haunted Caribbean hydrology for decades. Rainfall networks across the region are sparse, records are riddled with gaps, and conventional methods for building design rainfall curves quietly assume data that simply do not exist. The stakes are not abstract. Hurricane Ivan in 2004 damaged or destroyed approximately 89 percent of Grenada’s housing stock, inflicting losses near 900 million US dollars, roughly twice the national GDP. Hurricane Beryl in 2024 caused an estimated 218 million dollars in damage and triggered parametric insurance payouts of 55.6 million dollars. Designing infrastructure against the wrong rainfall statistics has direct, measurable consequences.

The team assembled an archive of 28 rain gauges drawing on records from Grenada’s National Water and Sewerage Authority and the Grenada Airports Authority. Before any analysis, the raw material was daunting: 7,671 observed station-days from the water authority network alongside 14,885 daily values from the airport, with most non-airport stations suffering gaps ranging from isolated days to entire missing years. Only Point Salines, with an approximately continuous 40-year record, approached the completeness that standard frequency analysis assumes. For many stations, the eventual curated dataset guaranteed a minimum of 12 years of usable records, a thin foundation for estimating rainfall quantities associated with 50- or 100-year return periods.

The workflow begins with a diagnostic innovation. Rather than relying on the classical semivariogram, the geostatistical workhorse that bins station pairs by separation distance, the authors computed site-specific Pearson correlations directly between every pair of stations. This approach, adapted from techniques used to study non-stationary spatial correlation in earthquake ground motions, sidesteps a known weakness: with so few stations, lag-bin averaging obscures the very structure the analysis is meant to reveal. The verdict was clear. Daily rainfall dependence in Grenada is governed primarily by how far apart two stations are, with elevation acting as a secondary influence whose effect shifts with the seasons. March, one of the driest months, showed strikingly coherent spatial rainfall, while June, as the Intertropical Convergence Zone migrates northward, produced far more scattered behavior.

Those diagnostics fed directly into the gap-filling stage. The team fitted spatial correlation models to the daily rainfall field using Bayesian inference, implemented in Python with the NumPyro library and the No-U-Turn Sampler, treating model parameters as probability distributions rather than fixed numbers. Ordinary kriging driven by these Bayesian-inferred models then reconstructed missing daily values, and it outperformed a full bench of competitors including inverse distance weighting, radial basis functions, and Gaussian process regression. Adding elevation dissimilarity as a covariate delivered marginal but consistent gains. Strict quality control followed: imputations were only retained when at least six donor stations contributed and the kriging prediction variance stayed below 25 percent of the marginal daily variance, lifting every station’s completeness above 70 percent.

With the reconstructed dataset in hand, the researchers turned to extreme value theory. Instead of the Gumbel distribution used in earlier Grenadian studies, which effectively fixes the shape parameter at zero and can underestimate rare rainfall quantiles, they fitted the full generalized extreme value distribution to annual maxima and the generalized Pareto distribution to peaks over threshold, the latter using declustering to ensure that exceedances separated by less than 24 hours counted only once. Because most stations contributed just 12 years of annual maxima, weakly informative priors were specified through an empirical Bayes-style strategy informed by preliminary analytical fits, stabilizing inference on the tail-shape parameter that controls how heavy the rainfall distribution’s upper end truly is. Markov chain Monte Carlo sampling, with four chains and extended warm-up, produced full posterior distributions for every parameter.

The physical signals that emerged are plausible for a mountainous Caribbean island. The shape parameter correlated moderately with elevation, at 0.25 for the generalized extreme value model and 0.34 for the generalized Pareto model, consistent with orographic enhancement steepening the upper tail of the rainfall distribution where moisture-laden trade winds are forced over the central highlands. Interpolating the posterior return levels across the island required choosing among kriging variants, and intrinsic collocated cokriging with a residual correlogram, which exploits elevation as a secondary variable, won on leave-one-out cross-validation. Five poorly constrained stations with only about four years of reliable data were excluded after sensitivity testing showed they destabilized the fitted spatial structures. The final maps show the highest predicted extremes in the island’s northeast, broadly matching Grenada’s known climatic zoning, though accompanied by appropriately large uncertainty estimates.

The study is unusually candid about its own limitations. A sensitivity analysis traced every annual maximum and threshold exceedance back to its source, classifying each as observed or imputed, and then refitted the models using observed extremes only. Where imputed values dominated a station’s extreme sample, return levels shifted dramatically, with differences approaching 80 percent for some generalized Pareto estimates and exceeding 50 percent for some extreme value estimates; at the 25-year return period, most stations stayed within roughly 25 percent. The authors attribute this partly to the smoothing inherent in kriging, which produces conditional-mean predictions rather than stochastic realizations and can dampen localized extremes. They suggest that future work propagate imputation uncertainty directly, through empirical Bayesian kriging, multiple conditional realizations, or hierarchical models treating missing rainfall as latent quantities.

What elevates the work beyond a single-country case study is its transferability. The complete workflow, from site-specific correlation diagnostics through Bayesian model fitting to island-wide interpolation, is publicly available through GitHub and Zenodo repositories, and the underlying Grenada Daily Rainfall Database has been released on Zenodo. The authors also outline an engineering validation path, proposing two-dimensional flood simulations in flood-prone catchments such as St. John’s and Great River to test whether the estimated rainfall fields produce physically reasonable inundation. For small island developing states facing intensifying hurricanes and rising adaptation costs, the message is straightforward: with Bayesian methods, even a fragmented, decades-old network of rain gauges can yield defensible, spatially explicit design rainfall, provided the uncertainties are confronted rather than hidden.

Subject of Research: Bayesian estimation of multi-station rainfall intensity-duration-frequency curves and extreme daily rainfall mapping in data-limited island settings, applied to Grenada

Article Title: A Bayesian workflow for multi-station IDF curve development in data-limited island settings: application to Grenada

Article References: Rampersad, A. J., & Zakour, C. M.-C. F. (2026). A Bayesian workflow for multi-station IDF curve development in data-limited island settings: application to Grenada. Theoretical and Applied Climatology, 157(10), Article 633. https://doi.org/10.1007/s00704-026-06558-4

Image Credits: AI Generated

DOI: 10.1007/s00704-026-06558-4

Keywords: IDF curves, extreme rainfall, Bayesian inference, Grenada, kriging imputation, generalized extreme value distribution, generalized Pareto distribution, spatial correlation, small island states, geostatistics, return levels, climate risk