As winter turns to spring, flowers begin to bloom and seasonal hurricane forecasts are about to be released. There will be a lot of seasonal hurricane predictions coming out in the next few months and almost all of them will rely on the six-month sea surface temperature (SST) forecasts released by national modelling centres. This raises the question, how accurate are these six-month SST forecasts?
There’s a lot of different ways that we can evaluate how good or bad a forecast is, but one straightforward way is to look at how well it has performed in the past. The plot shows SST forecasts for the Main Development Region of the Atlantic Ocean over the past 18 months, released by ECMWF, a leading modelling centre. By comparing past forecast anomalies (grey lines), released monthly, to observed SST anomalies (yellow line), we can see that the forecasts have correctly predicted the cooling trend in the Main Development Region since June 2024. However, they have often been too aggressive with the cooling (grey forecasts are mostly below the yellow observations). Looking ahead to the start of hurricane season in June, the forecasts predict continued cooling, which would significantly lower hurricane risk compared to 2024. Whether the latest forecasts (blue line) are accurate remains one of the major questions that organizations releasing seasonal hurricane forecasts will need to grapple with. Inigo will be releasing our view in the coming weeks, so stay tuned!
AI technology is now ubiquitous, with specialised hardware and software playing crucial roles in the development and deployment of artificial intelligence models. Popular frameworks like TensorFlow and PyTorch run seamlessly on multiple types of hardware and Graphics Processing Units (GPUs) and Tensor Processing Units (TPUs) are extensively employed to solve computational fluid dynamics problems.
In a recent paper co-authored by Boyang Chen, Amin Nadimy, and colleagues, we demonstrate how convolutional layers commonly used in neural networks can help solve complex partial differential equations. The paper, titled “Solving the Discretised Shallow Water Equations Using Neural Networks“, showcases this innovative approach. The NN4PDEs framework, developed by the Applied Modelling and Computational Group at Imperial College London, enables these use cases. In this method, the convolutional layer provides a discretisation scheme to solve the equations. Though conceptually simple, the paper illustrates the practical implementation for complex equations like the shallow water equations.
Importantly, these tools can effectively model flood risk with the accuracy and spatial resolution needed to represent local water depth during significant floods. We applied this to simulate the flood event impacting Carlisle, UK, in January 2005. The figure below, from the paper, depicts the problem setup and the solution provided by this approach.
While this early work demonstrates that the NN4PDEs approach achieves results and computational performances comparable to traditional computational fluid dynamics tools, it also paves the way for further applications. These applications can seamlessly combine AI models and physics on different scales or allow scaling up hardware to solve complex fluid dynamics problems over large domains.
Wildfire Whiplash – sudden shifts with painful consequences.
Rapid transitions between wet and dry conditions can increase frequency and severity of natural catastrophes like flash flooding or wildfires. The unprecedented wildfires that have devastated Los Angeles County are, in part, a result of this phenomenon. The figure below shows cumulative precipitation data recorded by a rain gauge in downtown LA since 1907. After two exceptionally wet seasons in 2022/23 and 2023/24 (blue lines), the LA hills experienced an explosion of grass and brush growth. However, since then Los Angeles has endured its second driest June to January on record (red line), desiccating the landscape and creating a highly flammable tinderbox environment. The combination of this abundant fuel, alongside strong winds and ignitions has led to the most destructive wildfire season ever recorded.
Inigo has established a partnership with climate scientist Daniel Swain (UC ANR & UCLA) to better understand US wildfire seasonality and associated risks. His latest publication on ‘hydroclimactic whiplash’ is extremely timely and an excellent summary of it can be found on his website – Hydroclimate volatility on a warming Earth – Weather West.
Santa Ana winds intensified the ongoing Franklin fire in Malibu, California, forcing mass evacuations and valiant firefighting efforts. These hot, dry, powerful winds desiccate vegetation, carry embers, and increase flame length, accelerating wildfire spread. From 1990 to 2009, 50% of burned area and 80% of wildfire economic loss in Southern California was linked to Santa Ana winds (Kolden et al., 2018). More recent examples include the 2017 Thomas and 2020 El Dorado fires, which together burned over 300,000 acres.
Climate change is increasing annual average air temperatures and wildfire burned area across the Western US (NOAA, NIFC), but its impact and Santa Ana winds is uncertain. These winds are generated by pressure gradients between the warm Great Plains and the cooler Pacific Ocean that typically form during the autumn months. Reanalysis of weather data from 1979-2020 reveals that strengthening temperature gradients between coastal California and the continental interior may be increasing the frequency of these events (Thompson et al., 2023). However, global climate model outputs suggest that the frequency of Santa Ana winds may decline over the 21st century as the annual window of favourable conditions narrows (Guzman-Morales and Gershunov, 2020).
Though the evolution of Santa Ana winds as the climate changes is unclear, their devastating effect on wildfire activity continues to increase. California’s rainy season arrives ~27 days later than it did in 1960 (Lukovic et al., 2021). A extended wait for wildfire-relieving rains prolongs the period where vegetation dried over the summer months is exposed to Santa Ana winds. As a result, we may see an increasing number of late-season Santa-Ana-fuelled fires like the Franklin fire in the future.
Solid and dashed blue lines show the mean monthly precipitation for California between 1943-1963 and 2010-2023, respectively (National Oceanographic and Atmospheric Association). Yellow bars show the monthly average number of Santa Ana events recorded between 1948 and 2012 (Guzman-Morales et al., 2016).
Earlier this year on September 27th Asheville, NC was the center of the devastation brought by hurricane Helene, with record-breaking rainfall which resulted in extreme flooding. The flooding inundated homes, businesses, and roads, displacing hundreds of residents and leading to billions of dollars in damages and hundreds of houses destroyed in Buncombe County. With 227 fatalities Helene became the deadliest hurricane to hit the US since Katrina in 2005.
With one of the longest daily rainfall records in the U.S., dating back to the 1870s, Asheville’s historical data highlights the peculiar nature of this event. The figure shows, for each year in the historical record, the maximum 24-hour (blue) and 72-hours (yellow) rainfall totals. While Helene’s 1-day rainfall (and peak rainfall rate) were significant but not unprecedented and were exceeded at least twice in the historical record, the 3-day total far exceeded anything in Asheville’s history. This record-breaking 3-day rainfall was fuelled by two days of heavy rain before Helene’s arrival, caused by a unique meteorological setup. The precursor rain saturated the ground and raised river levels, leaving the region vulnerable to catastrophic flooding once Helene made landfall.
This event highlights the growing challenge of preparing for extreme weather. The sequence and intensity of rainfall, rather than daily peaks alone, proved decisive in causing the devastation. At Inigo, we’re prioritizing the evaluation of flood risk models and maximizing our learning from these events to improve our preparedness and resilience in the face of increasingly severe weather patterns. Asheville’s experience underscores the urgency of better understanding flood risk and adapting to a changing climate to mitigate future risks.
AI weather forecasting is in the news again this week, with Google DeepMind’s release of GenCast. AI weather models have made rapid progress for forecasts up to 15 days in advance, but less progress has been made on the notoriously difficult problem of seasonal predictions. Accurate seasonal predictions would provide an opportunity for society to prepare well in advance of hazardous conditions. Here at Inigo we’ve spent the last few months refining our strategy for seasonal hurricane prediction. A major part of that work has been creating I-SPARK (Inigo Seasonal Prediction and Analysis of RisK), a new AI tool to help make accurate seasonal hurricane predictions at insurance relevant lead times of three to nine months.
I-SPARK uses freely available sea surface temperature forecasts to predict the level of risk for the upcoming hurricane season. The figure shows some of the sea surface temperature regions that the model will be using to make its December initiated forecast in a few weeks time. Interestingly, our model disregards most ENSO information until after spring due to low predictability. As the figure inset shows, the model skill rapidly increases by February, but then falls off in spring. This gives us confidence that useful hurricane risk information can be derived as early as February, but we need to do more work to understand why the model struggles in spring.
This work is being presented at the American Geophysical Union conference in early December, 2024. You can check out the full conference poster and additional details on GitHub. Please reach out if you want to learn more.
The 2024 hurricane season began with the earliest Category 5 hurricane ever recorded. In an unprecedented twist there was an eerily long and unexpected lull during the peak season, before ending with a flurry of activity in the Gulf. Next month, Charles Powell, from the University of Cambridge, and I will be sharing a retrospective of 2024 as part of the work that Inigo is funding under its InSPIRe programme.
The chart compares the timing of the first and last major hurricanes for the climatological periods 1980-2020, 1980–2000 and 2000–2020 and compares these to 2024. The vertical lines denote the temperature range between the first and last major hurricane and the horizontal line the mean temperature. From the climatological data we can see that there is a fairly well-defined start to the first major hurricane in the season, but that the end of the season appears to be extending, with the last major hurricane occurring later in the 2000–2020 period compared with the earlier 1980–2000 period. The first major hurricane of 2024, Beryl, is a remarkable outlier becoming a major hurricane on the 30th of June. The last major hurricane of 2024, Rafael, reached major hurricane status on the 6th of November, which, while not unprecedented, is approximately two standard deviations from the long-term mean.
Our analysis shows that the average temperatures in the Atlantic main development region have increased between the early and later periods, 1980–2000 and 2000–2020, with 2024 being one of the hottest years on record. As global temperatures continue to rise due to climate change, warmer oceans are able to sustain major hurricanes earlier and later in the season potentially leading to extended hurricane seasons This could pose a challenge for the insurance sector, as amplified hurricane risk over a longer season increases the potential for catastrophic loss events.
Typhoon Shanshan made landfall in the Kagoshima Prefecture of southern Japan on August 29th bringing considerable flooding and rainfall with it. Some areas broke rainfall records with over 300mm of rain in 24 hours. Early forecasts showed the potential for even more catastrophic outcomes with the storm forecast to make landfall as a powerful typhoon between Osaka and Tokyo, a highly populated area. However, the storm ended up making a sharp turn westward before turning north and making landfall (yellow track on map). The map below shows the forecast tracks for the ECMWF traditional and AI models initialized on August 23rd with the forecast strength of the storm indicated by text in each track point.
Weather models generally did a poor job forecasting this storm. One of the first models to pick up on the westward turn was the AI model run by ECMWF (blue track), one of the world leaders in weather prediction. In many ways this was a win for the AI models, they accurately forecast the track of a dangerous storm hours and days before traditional weather models (the ECMWF traditional model is the orange track). However, this success came with a massive caveat. The ECMWF AI model forecast the westward turn, but also forecast the storm would never rise above tropical storm strength, a known and recurring issue for AI weather models. In reality, the storm briefly reached category 4 strength, with sustained winds of 215km/h, before weakening to a category 1 storm prior to landfall. All together, Typhoon Shanshan presented an interesting case study in the promise and shortcomings of AI weather models.
Inigo Welcomes Patrick Ball to the Climate Science Team.
“As a Risk Scientist within Inigo’s Catastrophe Research team, my role is to help our underwriters and exposure managers leverage the latest scientific advances and cutting-edge technology when assessing the likelihood and impact of natural disasters to our insurance portfolio. This work includes evaluating and optimising catastrophe models, collaborating with academic institutions and insurtech companies, and performing bespoke research projects. This scientific data-driven approach ensures Inigo stands out from the crowd! While our team covers all natural disasters, my specialty is earthquakes. I hold a masters and PhD in Geophysics from the Universities of Oxford and Cambridge, respectively, and prior to joining Inigo I worked as a researcher at several universities and a catastrophe model vendor”.
Idalia made landfall as a major hurricane this week in a season where we’ve been anticipating what might become of a Battle Royale between the anomalously warm seas in the Atlantic and an ongoing El Niño in the Pacific. A reminder: we’re not even climatologically half way through the season yet…
Taking a look at the sea surface temperatures in the Gulf of Mexico, it’s little surprise we saw a strong landfall: in the week leading up to Idalia, the sea in its path was the warmest it’s been in since 1982 (when this particular sea temperature dataset started). The chart here shows the average sea temperature for the period of Aug 23-29th for every year for the past 40 years in the red box shown that broadly straddles the region through which Idalia passed and developed. Naturally other factors will always influence how a storm develops, but you don’t typically get the heavy-hitters of hurricanes without the warmer seas.
And now for the rest of the season…
Lahaina, located on the northwest coast of the Hawaiian island of Maui, experienced a devasting wildfire in August 2023. Several factors likely contributed to the severity of this wildfire including enhancement of strong easterly winds by Hurricane Dora, Katabatic winds flowing from the West Maui Mountains, located to the east of Lahaina, and drought conditions.
The precipitation rate for all of Hawaii has been extracted from the NCEP Reanalysis data between 1980 and 2023 and is shown in the graph. The precipitation rate for February 2023 was the highest recorded since 1980 at nearly three times the average. Precipitation rates throughout the spring were above average but by June it is below average, with July being extremely dry in the bottom 5% of years.
The February and spring rainfall likely contributed to a growth of vegetation. During dry seasons an increase in fuel availability contributes to the wildfire risk. Precipitation rates over Hawaii have been well below normal for June and July leading to drought conditions across Hawaii. According to the US Drought Monitor the most severe droughts over Hawaii in South and West Maui (where Lahaina is located). It is likely that the drought conditions combined with an abundance of fuel from the wet February and spring were contributory factors in the severity of the Lahaina wildfire.
The Copernicus Climate Change Service produces monthly seasonal forecasts from many different weather forecasting centres to help us understand possible future conditions around the globe.
The chart below shows the expected rainfall through the months of September to November from five different seasonal forecast models. Greens show rainier conditions, brown shows drier conditions. The rainier conditions can be indicative of increased tropical cyclone activity. This year we have an El Nino – that usually weakens hurricane activity – but a warm Atlantic, that can increase activity. So, any indications of what is come in such a difficult year to forecast are always useful.
It’s interesting to note the green, wetter region across the Tropical Atlantic. However,this region exists more towards the central / eastern Atlantic, which may be indicative of a busier hurricane season here, but this anomalous wetness is reduced towards the eastern seaboard of the US – although rainfall is still expected to be slightly above average here. All to play for with the three key months ahead – but will we be spared hurricane landfalls with a busy tropical season staying over the sea?
The Atlantic Ocean is starting to wake up and with it eyes are turning towards the upcoming hurricane season. In spring of each year various public and private entities forecast the level of expected hurricane activity during the upcoming Atlantic hurricane season that officially starts in June. 2024 brought forecasts warning of a potentially catastrophic hurricane season. Those forecasts largely verified with 5 US landfalls and NOAA declaring an extremely active season, but there were surprises with a lull during the typical peak of hurricane season in August and September. As we look ahead to 2025 it’s worth taking a moment to step back and think about how seasonal hurricane forecasts are made.
Two of the primary metrics used to predict seasonal hurricane activity are the sea surface temperatures (SST) in a region of the Atlantic Ocean called the Main Development Region (MDR) and a region of the Pacific Ocean called the El Nino Southern Oscillation (ENSO). There are a few ways to measure ENSO, but the most common is by using the SSTs in the Nino3.4 region (see map, Figure 2). When the MDR is warm and the Nino3.4 region is cool hurricane risk is elevated, while a cool MDR and warm Nino3.4 region generally reduces hurricane activity. Both of these regions have decades worth of scientific literature highlighting their importance and were the centrepieces of the elevated 2024 forecasts. Other variables such as vertical wind shear, Saharan Dust, the Madden-Julian Oscillation, and atmospheric humidity levels are important, but forecasting them months in advance is not currently possible. As a result we end up relying quite a bit on variables that are predictable, such as SST. While the MDR and the Nino3.4 SSTs are important for predicting seasonal hurricane activity, they are less accurate than desired. Even if we knew exactly what temperatures would be in these two regions during hurricane season, a simple leave one out multivariate regression model using data since 1980 can only explain 34% of the variability in Atlantic hurricane activity (Figure 1). More sophisticated modelling frameworks can improve this, but the problem still remains. Adding to the difficulty is that we don’t know what temperatures in those two regions will be during hurricane season as we rely on seasonal forecasts released by C3S to predict what the SSTs will be. This means that seasonal hurricane forecasts rely heavily on imperfect forecasts of two variables (MDR and Nino3.4) that can only explain part of the variability in hurricane activity. No wonder seasonal hurricane prediction is so difficult! As we get closer to hurricane season the C3S seasonal forecasts become more accurate and we can begin to use other variables in addition to SST. Both of these help create more accurate forecasts and are a big reason why the June forecasts from Colorado State University are much more accurate than their April forecasts (see page 45 of CSU Hurricane forecast verification).
Figure 1. While the MDR and the Nino3.4 region are helpful predictors of hurricane activity, they fail to capture the full range of observed variability.
Given the historical difficulty of making seasonal hurricane forecasts in spring, what is Inigo doing to try and make our forecasts more accurate? Over the last few months we’ve been exploring what other regions of SST might be useful to include. We’ve done this by leveraging large sets of climate model simulations, using data driven ML approaches, and continuing our ongoing partnership with Reask. In addition, we’ve spent time thinking about what the ‘right’ variable to predict is. As an insurance company with significant US interests what we really want to know is the specific landfall information for every storm in the upcoming 2025 season. Unfortunately, this level of detail is impossible to know this far in advance so we have to settle for alternatives. The two complementary approaches that we highlight below predict different things, but help us form a more complete view of hurricane risk in 2025.
The first approach uses a machine learning model called I-SPARK (Inigo Seasonal Prediction and Analysis of RisK). This model uses current values and future predictions of dozens of regions of SST to forecast if the accumulated cyclone energy west of 60° longitude (ACE 60W) will be in the lower, middle, or upper tercile for the season (referred to as Low, Medium, or High risk). We have trained our I-SPARK model to predict ACE 60W as we have found that of all hurricane related variables, ACE 60W strikes the right balance between predictability and correlation with insured US hurricane losses. This approach is also qualitative, we are only predicting a risk category of ACE 60W rather than its exact value. While this has limitations, we’ve chosen this approach due to the historically low skill of hurricane predictions in the March and April timeframe. We’re currently working on a more quantitative prediction method to complement this categorical ML model.
The second approach focuses on seven key regions of SST (Figure 2) and leverages Reask data to create quantitative estimates of what the upcoming season may bring given the forecast SST conditions in each of those seven regions. Reask provides us with hundreds of thousands of stochastic hurricane tracks tied to environmental conditions. By selecting only the Reask data that is connected to the environmental conditions we expect in 2025, we can create a 2025-specific stochastic hurricane set. This builds upon our approach from last year, and we’ve found the inclusion of seven SST regions instead of the two used last year (MDR and Nino3.4) yields more accurate forecasts over the last 20 years. Seasonal forecasting is always a humbling exercise, and we will no doubt learn something in 2025 that we can apply to 2026, but we feel confident that our methodology is moving in the right direction.
Figure 2. In 2024 the Inigo team focused on SSTs in two regions, Nino3.4 and the MDR, but have expanded this year to using the seven highlighted in this map.
As evidence of the improvement from increasing the number of SST regions used, if we repeat the example from Figure 1, where we used MDR and Nino3.4 values during hurricane season to predict the number of named storms that season, but now include our seven regions of interest, we see a noticeable increase in the variability explained from 34% to 45% (Figure 3). Not all historical years see improvement with the inclusion of seven variables instead of two, but most do. While not shown here, we see similar improvements using seven regions instead of two when using the Reask data.
Figure 3. Comparison of using two versus seven regions of SST to predict the number of named storms in the Atlantic Ocean. Overall, using seven regions yields noticeable improvement. Similar improvement is seen forecasting other variables such as ACE 60W.
Getting to the primary question of interest, what does Inigo think the 2025 hurricane season has in store for us? Our I-SPARK model predicts that ACE 60W will fall in the upper tercile, which would be indicative of increased hurricane risk, particularly near the US coastline (Figure 4). However, the likelihood of an upper tercile risk year has been decreasing in recent months. Using Shapley values we can discern that the model has backed off its earlier confidence for an active year primarily due to cooling in the Atlantic Ocean from January to early March, which have brought temperatures from near-record warmth to above average. While a below average year is still unlikely, the odds of an average year have been increasing.
Figure 4. Our I-SPARK ML model predicts the likelihood of a risk category for the upcoming season. While the odds of a ‘High Risk’ season are diminishing, the model predicts this is still the most likely outcome.
Our second approach uses the predicted SSTs in seven different regions and Reask data to create a stochastic hurricane set for the 2025 season. This method produces a similar result to the ML model, above average risk, but not nearly as dire as the 2024 seasonal forecasts were. This above average risk is primarily driven by warmer than average SSTs that are persisting in the Atlantic Ocean, but is also influenced by a number of other SST regions. To better understand how each of the seven SST regions are influencing risk we can visualize what the forecasts are for hurricane season compared to the observations since 1990 (Figure 5). The forecasts are from the C3S seasonal forecasts issued by a number of leading national modelling centres and we exclude modelling centres with exceptionally poor historical performance. A few items that stand out for the SST forecasts: the MDR is forecast to be below the record-breaking 2023 and 2024 values, but still above the long-term average and with a tail risk for very warm temperatures in 2025. This tail risk is partially supported by ocean heat content values (a measure of the amount of heat in the upper layer of the ocean) that are still quite high, particularly in the western MDR. The Iberian Coast SSTs might surpass 2024 for the warmest year, potentially raising risk via connections with the North Atlantic Oscillation (Elsner et al., 1999). The forecasts for the Nino1.2 and Nino3.4 regions quite uncertain and we are just emerging from the spring predictability barrier that makes ENSO forecasting difficult in winter and early spring. There is some indication that the Nino1.2 region could tilt towards the El Nino (warmer, and lower hurricane activity) phase and it has rapidly warmed since January, but whether that persists through hurricane season is an open question. Given these uncertain forecasts for the Nino3.4 and Nino1.2 regions we are not heavily relying on the either region in making our seasonal forecast.
Figure 5. Historical SSTs and the forecast distributions for this upcoming hurricane season for the 7 different regions we are primarily focusing on in 2025.
So, where does this leave us overall? Unlike 2024 the signals are not all pointing towards the same outcome. While some regions, the MDR and the Iberian Coast, point towards elevated risk, others, Nino1.2 and the Atlantic Nino, point towards average or diminished risk. Overall our I-SPARK model and our Reask-supported methodology both point towards slightly elevated risk. While our current best estimate for the 2025 season is slightly elevated risk, we also have reasons for cautious optimism thanks to recent cooling trends. Here’s to hoping for a quiet 2025 hurricane season!
We’re genuinely interested in our clients’ needs, using untapped data-backed insights and empowered underwriting to explore uncharted territory, exceed expectations, and provide innovative protection where others hesitate.
