Your Strategic Assumptions Might Be Wrong

Causal Discovery in Insurance Reporting

Every insurance organization operates on a set of inherited beliefs. “Claims inflation is driven by medical cost trends.” “Geographic concentration is our biggest risk.” “Underwriting quality is stable year to year.” “Urban drivers have higher claims frequency.” These aren’t guesses. They’re based on years of experience, industry benchmarks, and historical patterns. They’ve probably served you well.

But here’s the uncomfortable truth: they might be wrong. Or at least, they might have been right five years ago and wrong today.

Markets shift. Competitive pressure changes. Underwriting standards drift. Customer mix evolves. Regulatory environments shift. The relationships that held true in your historical data might not hold true anymore. And if you’re basing strategic decisions on assumptions that no longer match reality, you’re flying blind.

The question is: how do you know when your assumptions have become liabilities?

Last week, we talked about building causal models when you know the causal structure—when you can sit down with domain experts and say, “Here’s how we think the system works.” But what happens when you’re not sure? What happens when the market has shifted, or when you’re entering a new domain, or when your inherited wisdom feels shaky?

That’s where causal discovery comes in.

The Assumption Trap

Let me give you a concrete example. Imagine a mid-sized regional insurer that’s been in business for 30 years. They have strong beliefs about what drives profitability in their motor insurance business. They believe:

1. Claims inflation is primarily driven by external factors (medical cost trends, repair costs, etc.)

2. Mix shift is a secondary driver—they’re confident in their underwriting

3. Reserve adequacy is stable because their reserving methodology is sound

4. Geographic concentration in their home region is a manageable risk

These beliefs have shaped their strategy. They price for inflation. They don’t aggressively tighten underwriting because they trust their standards. They maintain reserves at historical levels. They’re comfortable with their geographic concentration because they know the region well.

Then something happens. Maybe it’s a market downturn. Maybe it’s a shift in competitive dynamics. Maybe it’s regulatory pressure. Whatever the cause, their loss ratios start deteriorating. Not dramatically, but noticeably. Over three years, they’ve gone from 65% loss ratios to 72%.

The leadership team meets. The CFO asks the obvious question: “Why?”

The actuarial team pulls together their analysis. They look at claims inflation indices. They compare their experience to industry benchmarks. They analyze mix shift. And they conclude: “It’s about 60% external inflation, 30% mix shift, and 10% unexplained.”

The board nods. They adjust their inflation assumptions. They tighten underwriting slightly. They move on.

But what if the real story is different? What if the problem isn’t external inflation at all, but internal? What if their underwriting standards have gradually loosened over the past five years—not deliberately, but through a combination of competitive pressure and staff turnover? What if they’re attracting a riskier customer base than they used to, and that’s the real driver of the loss ratio deterioration?

The strategic response to these two scenarios is completely different. If it’s external inflation, you adjust your assumptions and move forward. If it’s internal underwriting quality, you need to fix your underwriting process, retrain your team, and potentially reprice your entire book. You’re looking at a much bigger intervention.

The data they have—claims, underwriting characteristics, customer profiles, loss history—contains the answer. But they need to ask a causal question to find it. And they need to be willing to challenge their own assumptions.

What Causal Discovery Does

Causal discovery is a set of techniques that lets you learn causal relationships from data, rather than assuming them based on domain knowledge. Instead of saying “here’s how I think the system works,” you let the data tell you what relationships are actually present.

This sounds like it could be dangerous. Isn’t domain expertise supposed to guide analysis? Shouldn’t you start with your assumptions?

The answer is: yes, but. Domain expertise is crucial. But it can also blind you. If you encode your assumptions too early, you stop looking for alternatives. You see what you expect to see. You miss the relationships that contradict your worldview.

Causal discovery is a way to challenge that. It’s a systematic method for asking: “Given this data, what causal relationships are most likely?” It doesn’t replace domain expertise; it complements it. You use discovery to generate hypotheses, and then you validate those hypotheses against your domain knowledge.

Here’s the key insight: causal discovery is most valuable when your assumptions are uncertain. When the market has shifted. When you’re entering a new domain. When you’re seeing results that don’t match your expectations. In those moments, letting the data speak can reveal relationships you didn’t know existed.

The Real Example: Multi-Line Insurer

Let me give you a more realistic example from the financial reporting world.

A multi-line insurer with motor, home, and commercial lines has been seeing volatility in their reserves across all lines. They’ve attributed this to claims volatility and external inflation. But the volatility is larger than they’d expect, and it’s not consistent with industry benchmarks.

The traditional approach: they’d analyze each line separately. They’d look at claims development patterns. They’d compare to industry data. They’d adjust their reserving assumptions.

But what if the real driver is something they haven’t considered? What if it’s not claims volatility, but mix shift? What if it’s not external inflation, but internal underwriting loosening? What if there’s a correlation between underwriting decisions and reserve volatility that they haven’t noticed because they’ve never looked for it?

Causal discovery can help them find these relationships. By analyzing their reporting data—underwriting characteristics, claims outcomes, reserves, loss ratios across all lines and segments—they can discover which variables are actually causally related to reserve volatility. Maybe they discover that geographic mix shift is a bigger driver than they thought. Maybe they discover that underwriting tightness in year N predicts reserve adequacy in year N+2. Maybe they discover that certain product combinations have unexpected interactions.

These discoveries don’t come from intuition or industry benchmarks. They come from the data. And once they’re discovered, the insurer can validate them against domain expertise. Do they make sense? Are they actionable? Should they change their strategy based on what they’ve learned?

The Power and Peril of Discovery

Causal discovery is powerful because it can find relationships you didn’t know existed. It can challenge your assumptions. It can reveal non-obvious drivers of profitability, risk, and volatility. It’s a way of staying humble in the face of complex systems.

But it’s also dangerous if you’re not careful. Discovery algorithms can find spurious relationships—correlations that look causal but aren’t. They can find relationships that are artifacts of your data, not reflections of the underlying reality. They can lead you down rabbit holes that waste time and resources.

This is why validation is so critical. When a discovery algorithm suggests a causal relationship, your first instinct should be skepticism. Does it make sense? Is it consistent with domain knowledge? Can you think of a plausible mechanism? Can you test it on holdout data?

The best use of causal discovery is as a hypothesis-generation tool, not a hypothesis-testing tool. You use it to say, “Here are some relationships I didn’t know about. Let me investigate them.” Then you validate those hypotheses using domain expertise, statistical testing, and business logic.

When to Use Causal Discovery

Causal discovery isn’t always the right tool. If you have strong domain knowledge and stable market conditions, you’re probably better off with the approach we discussed last week: define your causal structure based on expertise, fit a Bayesian network to your data, and run inferences.

But there are situations where discovery makes sense:

When your assumptions are uncertain. You know the general shape of the system, but you’re not confident about specific relationships. Discovery can help you narrow down the possibilities.

When the market has shifted. Your historical relationships might not hold anymore. Discovery can help you identify what’s changed.

When you’re entering a new domain. You don’t have strong prior beliefs about how the system works. Discovery can help you learn the structure from data.

When you’re seeing unexpected results. Your models are predicting poorly, or your results don’t match your expectations. Discovery can help you find the relationships you’re missing.

When you’re doing exploratory analysis. You’re not trying to answer a specific question; you’re trying to understand the system better. Discovery is a great tool for that.

The Bigger Picture

Here’s what’s really happening when you use causal discovery: you’re admitting that you don’t know everything. You’re saying, “I have domain expertise, but I’m also willing to let the data teach me something new.”

This is uncomfortable for many organizations. There’s a lot of ego invested in inherited wisdom. If you’ve been running a business for 30 years based on certain assumptions, admitting that those assumptions might be wrong feels risky.

But it’s also liberating. Because the moment you’re willing to challenge your assumptions, you open yourself up to discovering things that could transform your business. You might find that your biggest risk isn’t what you thought it was. You might find that your most profitable customer segment isn’t who you think it is. You might find that the interventions you’ve been making aren’t actually addressing the root causes of your problems.

On Friday, we’ll get into the technical side: how to actually run causal discovery algorithms on your reporting data using tools like pyagrum. But the real insight starts here, in the boardroom. It starts with recognizing that your data might be telling you something different from what you’ve always believed.

The question is: are you willing to listen?

---

Reads of the Week

• In this sharp, clear-eyed piece, Michelle Bothe explains why you don’t need to build a full-stack insurance company to tap into the economics of insurance. For readers interested in innovation, startups, or platform strategy, this is more than just interesting—it reframes the booming MGA space and highlights the emerging role of program administrators as a more flexible, focused path. Bothe’s insight: the next generation of insurance programs will be built not just by industry veterans, but by outsiders with edge—and PAs make that possible.

• In this standout piece, Aaron Prather argues that insurance—not regulation—may be the decisive force shaping the future of robotics. Insurers are already influencing robot design, deployment, and business models by asking the only question that truly matters—who pays when things go wrong? As robots leave the lab and enter the world, Prather shows how insurability becomes a quiet but powerful gatekeeper of scale.

• This proposal from InnSure Insights reframes two crises—housing affordability and insurance availability—as a single, interconnected challenge. Whether it’s climate resilience, urban equity, or local governance, it all falls under a pioneering framework, which involves using Total Cost of Risk (TCOR) to embed insurance into affordable housing strategies at the regional level. The idea is bold and actionable—protect vulnerable communities not just with policies, but with better policy design.