Like most insurance companies, we rely on predictive models to calculate the lifetime value (LTV) of customers. After all, that’s basically what insurance is—trying to predict the future, and pricing risk accordingly.
At Lemonade, though, we’ve taken this much further. For each and every customer, our AI predicts their likelihood to file a claim, churn, or cross-sell—assigning a lifetime value (LTV).
Recently, we unveiled the sixth iteration of our proprietary model: LTV6. A synthesis of several existing models, it’s our most advanced and credible to date, and one that will continue to evolve.
Why should you care?
At the risk of sounding too nerdy, it’s a splashy moment at Lemonade when a new LTV model drops. Adopting each generational improvement makes us feel like we’ve been gifted the latest NASA telescope, allowing us to see further, and with greater fidelity, than anyone has ever been able to before.
LTV1 through LTV3 were humdrum in comparison. For instance, they allowed us to estimate claim frequency, but not claim severity. We were able to differentiate high-risk users, but overall the models weren’t as granular or accurate as we’d like.
In June 2021 we were able to take a quantum leap forward with LTV4, which leveraged AI models that were credible enough to be allowed into production. That iteration of our LTV model had the benefit of data culled from the almost four years we’d been in business at that point.
Machine-learning modules thrive on data, and it’s a steady supply of new information that keeps them sharp and worth listening to. We’ve had a surplus of data over the past 12 months—over the past year alone, we’ve handled twice as many claims as we did in those first four years of business combined. And during those same 12 months, we pushed 12,000 software builds into production.
LTV6’s lessons, put into practice
Once they reach statistical significance, each product, each market, and each campaign will be scored using LTV6. Lemonade Car’s forthcoming application of LTV6, for instance, will greatly benefit from the wealth of data we’ve inherited from our acquisition of Metromile.
After adding the customer acquisition cost (CAC) into the formula, the resultant LTV/CAC, together with the CAC Payback Period, determines which product, market, and campaign receives the incremental dollar spend.
This flexibility equals strategic optionality. It means we can adapt quickly, shifting resources from one product or market to another, and from one marketing campaign or channel to the next.
What makes LTV6 special, and worth getting excited about? Well, first off, it marks a genuine and dramatic improvement over LTV5. It comprises machine learning and deep learning models using almost a million parameters, incorporates state-of-the-art CAT (natural catastrophes) models, and it was trained on Lemonade’s ballooning, proprietary, and highly textured data.
One fairly dramatic insight from LTV6 was the extent to which certain homeowners insurance business in California was riskier than originally expected. LTV6 flagged that many policies in the state—which had been scored favorably by LTV5, and indeed previously appeared to have compelling marketing efficiency—would actually prove loss-making over time.
What each new iteration of our LTV model gives us is a greater degree of magnification and granular perception. So while LTV5 operated at a certain level of magnification, LTV6 is able to zoom in, subdivide an area, and spotlight good pockets and bad pockets in terms of risk.
So suddenly, LTV was revealing to us that within the mix of profitable pockets within California, it was really made up of better and worse. There’s white blood cells and red blood cells—it’s not just one monolithic blob of blood.
Again, countless parameters went into LTV6, but in this particular case, with California, it was arguably a greater granularity around catastrophic event modeling—particularly fire events—that helped sharpen our predictive model.
Another illuminating LTV6 takeaway is that our overall Pet business promises to be more profitable than LTV5 had assessed—welcome news, for sure.
Having this real-time data and predictive modeling—and being a multi-product, multi-market, multi-channel business—means that we can pivot and shift priorities quite quickly. So in Q2, we dramatically slowed our California homeowners business in several locales, while accelerating our Pet business across most states.
The systems we have in place both capture the data needed to train our neural networks, and are built to act on their outputs in real time. Each new policyholder brings fresh data into the model, making its predictions better and sharper. We’re not aware of any other insurer who has these capabilities. And it’s not something that an old-fashioned company could simply adopt and adapt; these tools and techniques are difficult to graft onto a company that wasn’t built with them as a core design principle.
A less distorted signal
Another point that’s worth digging into: For a fast-growing and developing business like Lemonade, lagging indicators and leading indicators often commend opposing actions, and the leading indicators are almost always right.
Consider loss ratios. One of the component models of LTV6 predicts the lifetime loss ratio of each customer. As the term suggests, this anticipates how much each customer will pay in premiums, and how much they will cost in claims, dividing the one by the other.
As you’d expect from a probabilistic model, the outputs can be hit-or-miss for any individual customer, but when you total them all up they become highly predictive.
It is this predicted loss ratio, rather than the ‘rear-view mirror’ kind, that we use in managing our business. Why the preference for a prognostic lifetime loss ratios over the traditional kind? Simply put, our company is evolving too fast to rely primarily on a feedback loop that lags as much as statutory loss ratios do.
So far, so unremarkable. What is remarkable, though, is that 73% of Lemonade’s premiums in Q2 were from customers who have been with us for fewer than 24 months. For incumbents it’s the other way around: the overwhelming majority of their customers are comfortably beyond their second anniversary. A health check that fails to account for the relative newness of our customer base will yield an unduly grim prognosis.
It follows that, for companies like Lemonade, traditional loss ratios can send a distorted signal, an outdated one, or both. We’re not saying to shrug and disregard loss ratios. But it’s important to look at them with a bit more nuance—and that’s where our machine learning capabilities kick in.
By using big data and machine learning to predict lifetime loss ratios, we timeshift a ‘lifetime’ of future data into the present, allowing us to optimize our current business for the long term. (Admittedly, we also really enjoy saying the phrase ‘timeshift a lifetime of future data into the present.’)
This is in stark contrast to traditional loss ratios. These also timeshift data to the present, but they do so from the past, rather than from the future.
Indeed, it’s this deep understanding of our loss ratios that gives us the confidence to reiterate our expectation that our business will operate on a multi-year average loss ratio below 75%. In a deep sense, it already is, once we pay attention to the superior modeling capabilities of LTV6.
What’s next? We’ll continue to improve and elevate our LTV models, leveraging AI and machine learning with a nuanced understanding of the responsibility those technologies require.
Lemonade is well aware of the pitfalls and limitations of AI, and we keep the ethics of artificial intelligence at the forefront. It’s why we’ve hired Tulsee Doshi as our AI Ethics and Fairness Advisor. She’s also the co-host, along with CEO Daniel Schrieber, of Benevolent Bots—our podcast series investigating the ethical implications of AI.
If you’re interested in how Lemonade is thoughtfully applying cutting-edge tech to one of the most conservative industries, there’s no better place to start.