The ability to accurately forecast annual changes in the likelihood of damaging hurricanes, cyclones and typhoons occurring has been a challenge in the past.
Now, with advances in technology, research and the use of AI, new companies are emerging with techniques to provide insurers with the evidence they need to confidently change their assessment of potential catastrophe risks over the short term (less than one year).
Australia-based reask is rethinking how catastrophe models are built, with a global solution that has caught the attention of leading insurers and investors.
Co-founders Nick Hassam and Thomas Loridan join Matthew on Podcast 127 to discuss how they are helping insurers understand climate change, and why it is driving a greater frequency and intensity in natural disasters.
Talking points include:
- Why today’s models need a global focus
- Seasonal forecasting
- Machine learning and neural networks
- Integrating new approaches with historical views
- Tracking storms after landfall
If you like what you’re hearing, please leave us a review on whichever platform you use, or contact Matthew Grant on LinkedIn.
Sign up to our newsletter for a fresh view on the world every Wednesday morning.
Continuing Professional Development – Learning Objectives
InsTech London is accredited by The Chartered Insurance Institute (CII). By listening to an InsTech London podcast, or reading the accompanying transcript, you can claim up to 0.5 CPD hours towards the CII member CPD scheme.
- Claim 0.5 hours for listening to Episode 127 of the InsTech London Podcast
Short term climate variability – forecasting with confidence – Episode 127 highlights
Matthew: Can you start by explaining what reask does? What challenges are you looking to solve?
Nick: We started reask to address two key challenges in natural catastrophe modelling.
The first goal was the need to build natural catastrophe models that are smarter in their understanding of climate signals. Current models don’t include climate forcing, which makes them not as useful for understanding risk in a warmer climate or getting a view of seasonal trends. Our solution was to rethink the way to build hazard models and to make sure they’re explicitly connected to global climate data from the start.
The second goal was to ensure global coverage within a single model. There’s no reason to silo the model building process on a region-by-region basis. We put a lot of effort into designing a more automated and scalable process for creating models that gives us the consistency to model across different regions and get a better understanding of global risk.
Matthew: Is reask focused solely on climate? You’re not looking at hazards like earthquakes or non-climate perils?
Nick: No, we’re looking in the atmosphere and, to a certain extent, to the hydrosphere. What we’re interested in understanding is how climate impacts the extreme perils that are phenomena of that particular environment.
Matthew: Most of our listeners will know what you mean by atmosphere, but can you explain what you mean by hydrosphere?
Nick: That’s the segment of the Earth that’s impacted by water. The impact of waves and ocean modelling on storm surges and flooding. With the geosphere then we’re talking about earthquakes and those types of phenomena. We’re focused above the ground, but primarily in the atmosphere.
Matthew: Are you offering different levels of granularity, modelling or analytics to customers in different countries?
Nick: No, we’re building a view from the ground up as a global solution that still enables us to have high-resolution views of risks in individual territories. We go all the way down to one kilometre of resolution, which is considered a good level of detail for cyclone risk.
Our goal is to have a consistent view, whether that’s of tropical cyclone risk in the South Pacific in Australia, or in the north Indian Ocean in Bangladesh, or in the Gulf of Mexico and the Atlantic. Our approach is to have a consistent methodology that uses consistent data sources and provides that same view, irrespective of where the exposure is on the Earth.
Matthew: The reask website talks about ‘a globally connected framework and machine learning approach to climate pattern extraction’. What does that mean?
Thomas: It relates to the goal of building a global solution and we believe that machine learning is the only way we can get there. The way models have been built in the last 30 years involved a team of scientists spending time building one model for one country at time, then switching to another region. There’s never really time to build full solution coverage. We are changing that.
The field we’re interested in is pattern recognition and what’s often referred to as unsupervised learning. We are looking for algorithms that can go through vast volumes of data and figure out what’s worth using. We still have an expert screening the data at the end to make sure we understand the physics being picked up, but it’s almost a fully automated process.
Matthew: How do you incorporate natural variability in climate cycles and seasonality into the models for the short term.
Thomas: A lot of what’s been done in the last 20 years has been around experts looking at data and picking up global patterns. The typical example would be the ENSO (1) patterns which look at things like water temperature. They’re usually very simplistic. Unsupervised learning allows us to use machine learning algorithms to do that job instead. We use neural networks (2) a lot.
What we found is that we gain a lot in terms of predictive skills and we can trust the machine to pick up the right signals. There’s still a need for expert screening, because not all those patterns are physically meaningful, but it does speed up the process a lot.
Matthew: What are your sources of information and how often are you updating the data in the model?
Thomas: It depends on what we’re trying to do with the data and what part of the model we’re looking at. The global climate products available from either the US agencies or the European centres are fantastic. We ingest their data into our system every month, which gives us the ability to look at hurricane activity in the coming season.
For other purposes, like tropical cyclone wind structure and what happens to the winds when they travel overland, we had to build new datasets. We use a weather forecasting model and run it at very high resolution on hundreds of historical cases. This has been the training data we use for machine learning.
Matthew: Anyone wanting to launch new modelling tools has to think about the regulatory impact. If someone wants to add in reask, do they have to go through some kind of regulatory approval?
Nick: The responsibility for this is divided between us and our clients. We do a lot of validation from a scientific perspective and a lot of our methodologies are peer-reviewed. We put that work out to the scientific community to evaluate our methodology and process.
Often, our information is being used to augment or supplement existing risk management processes, rather than price a homeowner’s risk on its own. It’s a case of leveraging our information to augment an existing risk management process. The clients who use our information need to be able to field questions from regulators, so we provide them with documentation and evaluation to make sure we’re transparent in how we provide information.
Matthew: How do you make clients confident about your forecasts in an area where there isn’t absolute certainty?
Nick: The market determines what is an appropriate approach to take about assessing risk, and that encompasses all of their internal risk appetite and approach to risk management.
We speak to many people in the markets regularly to update them on the different parts of our methodology. We also write publications that are peer-reviewed and we make that information available to them. Where clients need further information, we guide them along that process.
We’ve very fortunate that all our clients are sophisticated users of risk assessment, catastrophe modelling, and hazard modelling tools, and they’re all very unique. From ILS firms to reinsurance intermediaries and global insurers, the way they incorporate and utilise our information is unique.
Matthew: Can you give us some examples of how clients are combining your views with what’s already in their models?
Nick: What we’re trying to do is make sure that organisations can augment and have a collaborative view of risk that incorporates both ours and their existing approach. For example, where they don’t have coverage in certain regions, we can provide a consistent view across all territories.
We can also provide different solutions to some of the other providers. We do probabilistic modelling, but we also do event response, seasonal forecasting, and we deal with climate change as well.
Matthew: How closely integrated can your solutions be to other model providers?
Nick: Organisations consume our information through traditional sources like exposure management and aggregation management tools. They can take the information at a very raw level and perform their own analysis in that respect.
When companies are conducting research and trying to incorporate new findings into existing views of risk, we’ve developed methodologies to enable them to understand the statistics at a reasonably high level. They can compare our unit metrics to their other views of risk, whether they’re historical views or commercial catastrophe modelling views.
Thomas: One way to think of the operational side is that our model answers questions that other models don’t because of the global approach. A client can look at relativities in the long-term view of risk we have versus a particular season in the Caribbean or US for instance.
Because our system is pretty consistent, they can look at relativities between the basins and apply that as loadings or however the relativity applies to current views.
Matthew: Can you talk about some of the clients you’re working with?
Nick: We’re fortunate to have AXA Group as a client and they have been with us since our inception. We’re talking to reinsurers and we have insurance intermediaries at the insurance and reinsurance level.
There are other areas like securitised insurance funds where we’re working with Twelve Capital and Securis on the ILS side.
Matthew: I’d like to come back to seasonal forecasting as it’s such an important topic. Are we at the point where you can provide a seasonal forecast that allows people to base underwriting and reinsurance decisions on it?
Thomas: What we’ve found is that we’re quite comfortable providing a forecast in early June, or maybe May, for the season ahead, but we would be reluctant to go much further than that.
We’ve provided forecasts to Twelve Capital for the last three years in a row and they’ve been good. The way we can test that is by asking “can we beat the climatology for the long-term average of hurricane landfalls? Can we do better than this for the last three years?” The answer is yes, we can.
The second point is then “can we do better than the other forecasts out there?” That’s also something where we’ve performed well and we’re pretty happy with how the model has behaved.
Matthew: Is there an independent body out there tracking forecasts? A kind of leaderboard of forecasting abilities?
Thomas: There is a very good website from the Barcelona Supercomputing Center that correlates every forecast for the North Atlantic basin. The limitation with that is most forecasts only look at how many named storms there are and how much accumulated cyclone energy there is and don’t forecast landfall hurricanes.
What we’re trying to do, and what Twelve Capital was interested in, is moving those forecasts on to include where we could see more risk. Hurricane landfall is essentially the next challenge.
Matthew: So, the first challenge of forecasting is information. Then there is a whole second order of inquiry around which hurricanes have formed and which ones will make landfall?
Thomas: That’s exactly the challenge. Some years see a lot of activity in the basin but a lot of storms don’t make landfall. They could be very strong storms, but they stay out at sea.
In other years, there’s not much activity, but one or two storms become what we call straight movers and go straight into Florida or the Gulf. That’s what people we talk to are worried about.
Matthew: You mentioned event response after hurricanes, cyclones or typhoons make landfall. How are you helping people understand what the intensity has been?
Thomas: We are trying to apply the same logic, which is this probabilistic modelling approach. We don’t believe there is enough information at landfall, or even days afterwards, to make a confident, deterministic view of what the risk is.
If we take tropical cyclone winds as an example, we know some things pretty well, like where the cyclone has been and the intensity, but other things like the size and shape of the storm are very hard to determine because everything changes at landfall.
The approach we take is to build a catastrophe model on the fly at landfall and create 100 different views for each event. We can summarise the information as an expected risk and provide it as one map of the expected wind speed at the surface level. We’ve provided that for the whole of 2020, both for landfall in the US and Japan to all our subscribers.
Matthew: Can you talk about any specific modelling providers, or other platforms, that you’re collaborating with?
Nick: We’ve begun conversations with several exposure management providers about adding our solutions to their platforms. Those are the traditional exposure management tools used in large insurance and reinsurance organisations. What that enables us to do is put our products directly into platforms that can start the process of understanding and calculating expectations of exposure, and potentially of loss as well.
Another key area we’re excited about, in terms of the use of our information, is in the parametric insurance space. It appeals to us as it’s much more focused on the hazard and what the potential for loss is. That suits the approach that we take.
Matthew: Anybody who wants to know more about that can take a look at our parametric insurance report. Taking a step back from the technical side, it would be good to understand what building a new business in Australia is like. What’s been your experience?
Nick: There’s a great start-up atmosphere here in Australia. We have a much smaller insurance market serviced by local and big global players as well, which means there is a smaller pool of funding to draw from. Having said that, there are organisations like Insurtech Australia and others like Insurtech Gateway, who set up an office here, so there is plenty of support for organisations looking to push into the insurance technology space.
The vast majority of our market has prospects based outside of Australia. We’ve managed to pick a place that is asleep when the rest of the world is awake, but we’ve managed to use that to our advantage. During the day, when everyone else is asleep, we do a lot of our technical research and development, product management and development. Then the next day we get to follow up on all of those things without being interrupted.
Hopefully, at some point this year Thomas will be returning to Europe to start the process of developing our presence in the European market.
Matthew: You mentioned funding. How are you funding the business?
Nick: Currently, we’re bootstrapped by revenue. We’re fortunate to have enough clients so we can afford to continue servicing them and push our research and development forward. We’ve had some help from government grants in terms of research and development, and tax incentives, which are effectively retrospectively applied to tax payments.
In terms of external funding, we had looked into a capital raise towards the middle of last year, but we decided against raising funds in the middle of a pandemic. We’ll push forward with those plans this year and we’ve continued our conversations with investors. With that funding, we’d be looking to scale our business development, and grow into Asia, Europe, and then eventually the US.
For more information about reask, go to reask.earth
InsTech London notes
1. ENSO
The El Nino Southern Oscillation – or ENSO for short. For the last couple of decades, an increase in temperature of the water in the central and eastern Pacific Ocean around the tropics has been linked to a reduction in the frequency of hurricane formation. But it’s a rather imprecise tool for hurricane forecasting – and one of the most devastating hurricanes, Hurricane Andrew, occurred in 1992 during an El Nino year.
2. Neural networks
These are advanced types of machine learning that learn as they go along – a bit like the human brain.