All Models Are Right: Tom Philp on Cat Model Uncertainty, the Multi-Model View of Risk and Insurers Who Are Too Cautious

Tom Philp did his PhD on hurricanes and European windstorms, then spent his insurance career building the scientific view of risk for atmospheric perils at one company under three names: Catlin, XL Catlin and AXA XL. Along the way he walked up the street to the philosophers at the London School of Economics to ask how decision makers should handle uncertainty that large.

Today he is the Founder & CEO of Maximum Information (Maxinfo), which he started in London in 2020. Maxinfo does not build catastrophe models. It sells what Tom calls vendor-agnostic middleware that sits on top of the models insurers already license, so they can compare, adjust and stress-test them (MagniPhi), plus a platform for reacting to live hazard information (PsiClone).

In Episode 177 of InsurTechTalk, Tom and I covered why “all models are wrong” is the wrong way to think about cat models, what a view of risk built from a suite of models looks like, how Maxinfo plans to go from London to Florida and Bermuda, and why he thinks insurers are often too risk averse.

About Tom Philp

Tom’s PhD was mostly on hurricanes and European windstorms, the inner-core wind structure and the physics of extreme events. He got frustrated with the academic side, he says, because he wanted to help build resilience and, sitting on the science side, you never see where your data goes. He also spent short stints living in Florida doing hurricane research.

During the PhD he presented his weather data to reinsurance catastrophe modelers and met the chief science officer of a syndicate called Catlin. He took a job there expecting to stay a year and then go back to academia. Instead, he found the questions in catastrophe modeling more foundational than the ones he got to ask in science. He thinks catastrophe modeling should be a research discipline in universities in its own right, and that it has largely only existed inside insurance.

At Catlin he was put in charge of building the scientific view of risk for atmospheric perils, US hurricane in particular. The job was not only to build a view, but to adjust a model for a specific portfolio’s uncertainties, for a company’s risk appetite, or for science that had gone out of date. Every adjustment, he found, made the uncertainty space grow. The true range of what a 1-in-100 hurricane could be was very large, and he was not sure that was reaching the decision makers at the end of the chain.

So he went to philosophers at the London School of Economics who had written on how climate policy experts define frameworks like the IPCC’s, to see whether the same thinking could help catastrophe risk decision makers. He also worked on micro-insurance in Africa and Latin America, building pricing models for drought and excess-rain covers for smallholder farmers, where he saw insurers, humanitarian agencies and disaster responders reacting to the same events in isolation.

Those two frustrations became Maxinfo’s two products.

All Models Are Right

I opened with Tom’s own target: George Box’s line that all models are wrong, but some are useful. Tom has argued that it is misleading. I told him the plan was to start hard and move to the easy part.

His answer: the quote hints at a belief that models should be a perfect reflection of the truth, and they are not. “A model is always right because it is telling you exactly how you’ve built it,” he says, and it tells you about the information that goes into your system. So the industry should stop searching for an abstract “best model” and instead ask what a given model design is telling you about the system you work in.

He knows people in other modeling disciplines push back on this. In weather forecasting, you make a prediction and you are told quickly whether it was right. In catastrophe risk, he says, it is almost by definition impossible to know whether you are right or wrong, because the events are so far out in return period that you will never see enough of them to find out. Tom thinks Box himself was probably arguing something similar. If you start from “all models are right,” the useful question becomes which models in your suite are suitable for the question you are trying to answer. That multi-model suite, he says, is where resilience to catastrophic events will come from.

Why the C-Suite Should Care

I asked why a CFO or any other executive at an insurer should care. Tom’s answer turns a scientific principle around. Science, he says, works on using the simplest model that fits the data best. That is fine in science, but his customers are risk managers dealing with massive uncertainty in loss model output. The better question for them is: what are all the different models that don’t fail different statistical tests and could plausibly represent our system? The result is a suite of models and a suite of answers that captures the uncertainty space better than a single best model would. He admits it is harder to work with operationally, and that is the problem Maxinfo’s tools are built for.

What Maxinfo Sells

I asked for plain language. Tom describes Maxinfo as “the first vendor agnostic middleware for catastrophe modeling.” Catastrophe model data flows into its tools, and users build views of risk: tailoring a model to the company’s risk appetite. If they are risk seeking, that means understanding what the model is doing in the tail and checking that reinsurance and portfolio structure fit. If they are risk averse, it means seeing quickly where uncertainty blows open. He says this takes internal research teams a lot of time, and the software makes it efficient and upskills those teams.

A Cat Model, Explained for an 8-Year-Old

I asked Tom to explain a cat model as he would to an 8-year-old. First, the probability that a hurricane hits the region where you live in any given year. Then the impact: if you live in a wood-frame house, does the roof come off or do only the windows break? Then the cost of replacing what is damaged. Do that across millions of exposures in every country you model, aggregate it, and it informs insurance products across the risk transfer chain.

Maxinfo does not create catastrophe models, he stresses. The vendor market was historically dominated by two or three firms and is now proliferating. With exposures and technology changing fast, having access to many models that let alternative information surface is something the market wants, Tom says, and is operationally hard. The software is meant to take that operational work away.

Where the Name Comes From

I told Tom I liked the names MagniPhi and PsiClone and the Greek letters in them. He said there was a story behind the company name, too.

As Tom tells it, “Maximum Information” comes from the debate between Niels Bohr and Erwin Schrödinger over what quantum theory meant after Newtonian mechanics, where measuring more precisely always gave a more accurate answer. Quantum theory said there was only a maximum state of knowledge you could reach. His analogy for insurance: catastrophe modeling has become exceptionally sophisticated, and some questions need all that complexity. “But there are a large number of others that really don’t. And actually, we might be kidding ourselves that we are adding information when really we’re not.”

The point is not that people can’t make precise decisions from cat models, he says. It is to build in the awareness that a very precise answer may carry inaccuracies, and that the chain may already have exceeded the maximum information that exists in it. I thanked him for the Bohr, Einstein and Schrödinger conversation and gave a shout-out to everyone trying to do quantum computing with one or two cats or a few qubits.

Who Buys

Maxinfo’s primary customers are large reinsurers, Tom says, with smaller syndicates coming on.

Large, sophisticated reinsurers use it for efficiency. One use case is automating the evaluation of models they don’t have capacity to review in house, such as non-peak perils. His example: Chile earthquake, where a reinsurer may have a model implemented but spends all its time on US hurricane. The main use case is testing models from different vendors against each other, seeing where they align and where they don’t, and having a constructive dialogue with the vendor.
Smaller syndicates that haven’t had catastrophe modelers in house and now want their own view of risk, because investors or regulators ask them to defend one rather than borrow one from a vendor.
Beyond insurance, it was always his goal to make the modeling available to insurers available to other disaster response sectors. Maxinfo is starting pilots with humanitarian agencies and with county-level emergency managers in the US.

A Thread Across Guests

At this point I gave a shout-out to earlier guests. Rick Zullo of Equal Ventures, on Episode 170, talked about a super El Niño and a softening cat market followed by more catastrophes year after year. Sid from Arbol talked about it, and here in California we feel the fires, which Kitt from WISH talked about. I also raised cyber catastrophes: with cyber wars around the world, including the wars in Ukraine and Iran, they are not kinetic, but they have a kinetic impact one way or another.

Tom’s answer to that thread came later, and from the other side of the same weather (see “The El Niño Test” below).

Lloyd’s Lab, Cohort 12

Tom had worked with Lloyd’s on consulting projects after leaving AXA XL, bounced early Maxinfo ideas off the Lloyd’s team, and was encouraged to apply for cohort 12 in 2024. The work: Lloyd’s syndicates spend a great deal of time on regulatory reporting to Lloyd’s, which oversees how they model. Over 10 weeks, with Lloyd’s teams and syndicates that volunteered, Maxinfo worked on what should be reported, which questions should be answered, and how to structure data so syndicates can respond to Lloyd’s quickly.

From Consultancy to Product

Tom set the company up in 2020. He had been at one company under three names, so he was aware that the problems he wanted to fix might not look the same elsewhere. He spent the first years consulting across the market, including for Lloyd’s, to learn other firms’ processes. He started bringing people on in 2022, and after a couple of years of development Maxinfo launched MagniPhi in 2024.

On growth, Tom says the team went from himself in 2022 to three or four more people in 2023, and is now 16 people. In its seed round, raised last year, Maxinfo took investment from LMSV (Liberty Mutual Strategic Ventures), Convex, Insurtech Gateway and Portfolio Ventures, a mix of VCs and CVCs. Most of them know the insurance market well, which mattered to him: from the outside, he says, the market can look peculiar, and you have to know your stuff and be trusted to sell into it. As Maxinfo moves into other sectors, he expects to look for broader investors.

He says the company has a healthy runway and expects to raise a Series A at the end of 2027, starting conversations with investors around mid-2027. London has been “a really nice sandbox,” and the next step is geography: Bermuda and certain US states. With it comes a new narrative. MagniPhi was positioned as the first vendor-agnostic middleware for catastrophe modeling. Now, as Maxinfo connects sectors and geographies, Tom describes it as “the first shared ecosystem for disaster and catastrophe risk management.”

The El Niño Test

With the seed money, Tom says, Maxinfo hired to build out PsiClone without losing focus on MagniPhi. If MagniPhi is the baseline view of risk, PsiClone is about what you do when new information arrives that might nudge you away from that baseline.

El Niño is his example. It is devastating for many communities worldwide, he says, but for Atlantic hurricane-exposed communities the risk drops quite a lot. In his words, this season has seen “basically only one hurricane land falling” and very little development in the basin. “I don’t see insurers responding to that live information,” he says. He doesn’t see them taking on more risk or offering new products when this happens, but he does see them getting cautious in years when high activity is predicted. He calls it a potential over-cautious bias, and PsiClone is built around that question: should you be more cautious, or more risk seeking, in a given year with the information coming in?

PsiClone is in pilot with humanitarian and government agencies, and Tom expects to partner with other sectors, such as utilities and potentially finance, to take catastrophe modeling beyond its insurance roots. He says he feels Maxinfo is the only one trying to build the engineering between all the data and models now available.

Going to the US: Fifty Markets

I asked what separates the US from the London market, whether Maxinfo would go directly into states like Florida and California or through Cayman and Bermuda, and how it would put people on the ground, because most of what happens in insurance happens in person.

Tom says he is reluctant to call the US a country from a market perspective. “It’s fifty,” I said. “It’s 50 countries, right?” he answered, each with its own regulation and market nuances. So this is not a US expansion, he says, but a targeted one. Florida is likely to be the base. He points to reinsurance links being formed between the state of Florida and the London and Lloyd’s market, and to Florida’s emergency response community, “born out of necessity,” which he thinks London can learn from. Bermuda is the other target: an excellent and historically innovative place for risk transfer, where Maxinfo has close ties to reinsurance and ILS firms and is starting to serve those clients.

The 35-Person Rule

Tom once asked someone who had led PwC’s expansion into Eastern Europe and Russia about taking Maxinfo to the US. The first question back was how many people Maxinfo had. About 11 at the time. The advice: don’t contemplate it until you’re about 35, or you’ll exhaust yourself. Tom took it as a good lesson and thinks a state-by-state approach can still work.

My challenge to that statement: it is really a function of the people, not the number. Do you need ten salespeople or five, and how many engineers? The number on its own doesn’t tell you much without the context.

Rapid Fire

Favorite hurricane to study: Katrina. A category 5 in the ocean that became a category 3 at landfall. “Didn’t matter.” It was enormous, it showed something wrong with intensity scales, it showed how much flood defenses change losses, and the litigation touched every part of the insurance chain. Anyone who wants one storm to learn the insurance impact of hurricanes, he says, should start there. I added that Mark Anquillare of Verisk, on Episode 81, described how Hurricane Andrew changed the industry.
Most overrated metric in cat modeling: AAL, average annual loss. Cat events matter for their 1-in-100 and 1-in-200 contribution and the reinsurance layering, Tom says, and looking at an AAL distorts that information “quite grossly.” Move away from it, keep as many tail metrics in view as possible, and compare them rather than settle on one.
One book on uncertainty: anything by Richard Bradley, a philosopher at the London School of Economics who co-developed the “confidence approach” to decision theory. Tom says it lets a user input what’s at stake and their attitude to uncertainty, and narrow the uncertainty space on that basis, instead of chasing a single best estimate.

One Thing to Change: Lean Into the Fear

My last question usually comes from the folks at TRIGGERnometry: what should people talk about more? For Tom I changed it: if the industry did one thing differently in cat modeling over the next three years, what should it be?

Tom said he had become cautious about how he answers this, and then gave his genuine answer. Because of very real and systemic risks, climate change among them, there is a lot of conservatism built into the modeling, and many people think that conservatism helps. Tom says a lot of research shows the opposite: being over-conservative leads to less protection, more people on the poverty line, insurance that is unaffordable, and structural problems in certain US cities. More nuanced modeling and more careful unpicking of the models, he says, can show where the industry is too cautious, drive pricing down and create risk transfer products “we might be, I don’t know, scared of, but we should be leaning into this fear.”

His one-sentence version: “I believe that insurers are often too risk averse and actually we should be leaning into that to drive innovation.”

My take

Across this year’s episodes, the catastrophe thread keeps coming back: Rick Zullo’s super El Niño after a softening cat market, Sid at Arbol, and Kitt at WISH on the fires we feel here in California. When I raised it with Tom, he came at it from the other side. When El Niño lowers Atlantic hurricane risk, he says, he doesn’t see insurers respond to that information.

In October I wrote that Stand is valued at $1 billion on a catastrophe model, and the next fire or storm season is the audit. Tom’s point is worth holding next to that line: in catastrophe risk, he says, it is almost by definition impossible to know whether a model is right or wrong from what you observe.

Where I pushed back was the PwC advice to wait until you’re 35 people before going to the US. It’s the function of the people, not the number: how many salespeople, how many engineers. And the US is fifty markets, where most of what happens in insurance happens in person, so the question is who you put on the ground, and where.

Tom closed on insurers being too risk averse. As I said on the recording, that’s the industry, and that’s the culture of the industry.

Key Takeaways

Tom inverts George Box: a model is always right about how it was built and what went into it, so the useful question is which models suit the decision, not which one is “best”
In catastrophe risk, events are too rare to prove a model right or wrong, so a suite of models captures the uncertainty better than a single answer
Maxinfo does not build cat models; it sits on top of licensed vendor models so users can compare, adjust and stress-test them
Large reinsurers use it to evaluate non-peak perils and to test vendors’ models against each other; smaller syndicates use it to defend their own view of risk to investors and regulators
More model complexity does not always add information, Tom says: “we might be kidding ourselves”
PsiClone is about reacting to live information, and Tom’s example is El Niño: risk drops for Atlantic hurricanes, and he says insurers don’t respond
Maxinfo’s next markets are Florida and Bermuda, treated state by state rather than as one US expansion
AAL is the metric Tom would drop first; tail metrics such as 1-in-100 and 1-in-200 matter more for cat
His one change for the industry: insurers are often too risk averse, and over-conservative models can make cover unaffordable

Tom’s figures and descriptions, including 16 people, the customer and pilot groups, his account of the 2026 Atlantic hurricane season, and the “first vendor-agnostic middleware,” “first shared ecosystem” and “only one” descriptions, are as he described them and were not independently verified.

Stand’s $70M Series C at a $1B valuation on a property-level disaster model (weekly report, October 17, 2026)
Kin’s $335M Hestia Re catastrophe bond, priced below guidance (weekly report, May 10, 2026)
Kettle × RLI, and the MGAs using third-party catastrophe models (weekly report, March 2, 2026)
Convex and the Onex-AIG deal (M&A quarterly report, Q4 2025)
Rick Zullo, Equal Ventures: a super El Niño after a softening cat market (Episode 170)
Sid Jha, Arbol: El Niño, La Niña and a wider distribution of outcomes (Episode 158)
Kitt Doucette, WISH: wildfire risk and mitigation (Episode 151)
Mark Anquillare, Verisk: how cat models changed the industry after Hurricane Andrew (Episode 81)
James Bradbury, Lloyd’s: inside the Lloyd’s Lab (Episode 90)

Connect

Maximum Information: maxinfo.io
Maximum Information on LinkedIn

If your team is working through cat model choices or underwriting technology, let’s talk strategy.

Tom Philp, Founder & CEO of Maximum Information