How to Sanity-Check Loss Ratios for an Insurtech Idea

Loss ratio — the share of every premium dollar that goes back out as claims — decides whether an insurtech idea is a business or a subsidized hobby. Before you build product, pricing, or a go-to-market plan, model losses (L) against premium (P) as LR = L / P, add an expense ratio (ER), and check whether the resulting combined ratio (CR = LR + ER) can plausibly sit near or under 100% at a price your target customer will actually pay.

Quick Answer: Estimate claims frequency and severity for your specific risk, divide expected losses (L) by realistic premium (P) to get LR = L / P, then add your expense ratio to get a combined ratio, CR = LR + ER. If CR cannot plausibly land near or under 100% at a price the market will bear, the idea needs a different risk, price, or business model — not better marketing.

This matters most to founders coming from software, where a "good" business model rarely turns on a single ratio computed before a single customer signs up. In insurance, it does. You can have real demand, a slick quoting flow, and a partner willing to distribute for you, and still be building a business that loses money on every policy it writes. The loss-ratio check below is the fastest, cheapest way to find that out before you spend on licensing, integration, or a launch.

Prerequisites: Define the Risk and the Covered Event

Before any loss-ratio math means anything, you need a precise definition of exactly what event triggers a payout, to whom, and under what conditions — a vague risk produces a vague, unusable estimate.

"Small business risk" or "gig worker protection" is not a defined risk; it is a category. A defined risk looks more like "physical damage to a rented e-bike caused by theft or collision, for riders operating within city limits, capped at replacement cost." That level of precision is not bureaucratic overhead — it is the only way the frequency and severity numbers you estimate later will mean anything, because "how often does this happen" only has an answer once "this" is specific.

Work through four questions before you touch a spreadsheet:

• What is the insured peril? Name the exact event — theft, collision, equipment failure, cancellation, liability claim — not the general category of worry it belongs to.
• Who is covered, and under what conditions? Define the eligible population, any underwriting filters you'd apply, and the exclusions that keep the risk from ballooning (pre-existing damage, commercial use on a personal policy, geographic limits).
• What triggers a valid claim, and what's the payout structure? Lump sum, repair cost, per-incident cap, or a deductible-and-limit structure each change your expected payout very differently for the same underlying event.
• Over what period is the risk measured? Loss-ratio math is normally expressed per policy-year (or another fixed exposure period), so pin that down before you estimate anything.

If you plan to route the risk to a carrier rather than hold it yourself, this exercise is not optional homework — it's the first thing any capacity partner will ask you to produce, in writing, before they'll discuss terms. The choice of who actually holds the risk changes how much precision you owe them upfront, which is exactly the trade-off covered in our comparison of insurtech distribution versus becoming a carrier. A distributor routing risk to an established carrier can get away with a looser initial estimate than a founder trying to hold risk on their own balance sheet — but "looser" still means defined, not vague.

This step is also where loss-ratio work fits into the broader shape of insurtech validation. It is one gate among several — model choice, customer pain, risk economics, distribution, and capacity — laid out in our guide to validating an insurtech startup idea. Loss ratio is arguably the sharpest of those gates, because unlike demand or distribution, it can kill an idea on the math alone, before a single customer conversation happens.

Steps: Model the Loss Ratio Before You Model Anything Else

Work through four steps in order — frequency, severity, loss ratio, then combined ratio — because each one depends on the step before it, and stopping after loss ratio without adding expenses gives you a falsely optimistic number.

Step 1: Estimate frequency

Frequency (f) is the probability the insured event occurs, expressed per exposure unit — typically per policy-year. It answers "out of every 100 policies we write for a year, how many produce a claim?"

You will rarely have your own historical claims data this early, so triangulate from proxies instead:

• Adjacent product data. A related, already-written line of insurance (a broader category your niche risk sits inside) often has frequency data you can adjust down or up based on how your risk differs.
• Operational data from the underlying activity. If you're insuring e-bike damage, a repair shop or fleet operator's incident logs tell you more than a general insurance benchmark ever will.
• Conversations with reinsurers, MGAs, or brokers who already write near this risk. They have seen frequency data you cannot get any other way, and are often willing to sanity-check a rough estimate even before a formal relationship exists.
• A contained pilot. If the risk is cheap enough to self-insure briefly, running a small, real cohort produces the only frequency data with no proxy error at all.

Whatever the source, write frequency down as an explicit number — f = 0.08 per policy-year, not "claims happen sometimes." An estimate you can't state as a number is not yet an estimate.

Step 2: Estimate severity

Severity (s) is the average cost per claim once it happens. Two risks with identical frequency can have wildly different economics if one pays out $200 per claim and the other pays out $20,000.

Resist collapsing severity into a single average too early. Most insurance losses split into two very different populations: attritional losses (small, frequent, predictable — the bulk of your claim count) and tail losses (rare, large, and the reason reinsurance and capital reserves exist at all). A single blended average can look perfectly reasonable while hiding a tail event that would wipe out a thin startup balance sheet. At minimum, ask what your worst plausible single claim looks like, separately from your typical one.

Step 3: Model the loss ratio

With frequency and severity estimated, expected losses per policy are simply:

L = f × s

Loss ratio compares that expected loss to the premium you plan to collect:

LR = L / P

where P is premium earned over the same exposure period. This is your pure cost of risk — what you'd need to charge just to break even on claims, with nothing left over to run the business. It says nothing yet about whether the company survives; it only says whether the risk itself is priced honestly.

Step 4: Add expenses for the combined ratio

Loss ratio alone flatters every insurtech idea, because it ignores the cost of actually operating one. Expense ratio (ER) captures everything else: customer acquisition, claims handling and adjustment, technology and servicing, compliance and licensing overhead, and the cost of any reinsurance or capacity arrangement.

ER = operating expenses / P

CR = LR + ER

A combined ratio above 100% means every policy you write loses money before you count anything else — no growth rate fixes that, it only loses money faster. Large, established carriers can sometimes tolerate a CR near or slightly above 100% because they earn investment income on the "float" between collecting premium and paying claims. An early-stage insurtech usually has neither the reserve scale nor the balance-sheet permission to lean on that offset, so treat CR under 100% as close to a hard requirement, not a nice-to-have.

The table below summarizes the variables — use it as a checklist, not a source of numbers; your own frequency, severity, premium, and expense figures are the ones that matter.

Worked Example: Feasibility for a Niche Cover

Run your own numbers through a simple worked example before trusting a spreadsheet someone else built — the mechanics matter more than any specific figures here, since your real frequency and severity will come from your own research, not a generic template.

Suppose you're evaluating a theft-and-damage add-on for e-bikes used by gig delivery couriers, sold as an embedded add-on at the point a courier platform onboards them. The figures below are illustrative inputs for the walk-through only — substitute your own researched numbers; they are not industry benchmarks.

• Premium: you believe couriers will pay P = $180 per policy-year for this add-on.
• Frequency: conversations with two e-bike repair shops that service delivery fleets suggest roughly one claimable incident per six policy-years of exposure, so f ≈ 0.167.
• Severity: average repair or replacement cost for a claimable incident is estimated at s = $650.

Expected loss per policy: L = f × s = 0.167 × 650 ≈ $108. Loss ratio: LR = L / P = 108 / 180 ≈ 60%.

Now add expenses. Say acquisition through the courier platform costs $25 per policy, claims handling averages $10, and technology plus compliance overhead adds $8 — total expenses of $43 per policy, so ER = 43 / 180 ≈ 24%.

Combined ratio: CR = LR + ER ≈ 60% + 24% = 84%. On these illustrative inputs, there's headroom — but the number that matters more than the point estimate is how it moves under stress.

Test sensitivity before you believe the base case. If your frequency estimate is off by half — a real possibility this early, and a directionally likely one if couriers who buy the add-on ride harder or in riskier conditions than the shops' general customer base (a form of adverse selection) — frequency doubles to roughly 0.33, expected loss per policy roughly doubles to $216, LR jumps to 120%, and CR lands near 144%. The same inputs that looked comfortable at the point estimate now describe a product that loses money on every policy. That swing, not the single headline number, is the real output of the exercise: does the idea survive a plausible miss on your weakest estimate, or does it only work if your first guess happens to be exactly right?

Signs the Numbers Kill the Idea

A handful of patterns in your loss-ratio model are reliable warning signs that no amount of good marketing, growth, or a better app will fix. Catch them before you commit capital, sign a distribution deal, or file for a license.

• Combined ratio only clears 100% at a price nobody will pay. If CR dips under 100% only when you price above what your target segment tolerates, you don't have a viable product — you have a viable spreadsheet.
• Your frequency estimate rests on one thin, unverified source. A single conversation or one comparable dataset is a starting point, not a basis for a launch decision. Triangulate from at least two independent sources before trusting the number.
• A small severity miss swings you from profitable to underwater. If doubling your severity assumption — a routine margin of error this early — pushes CR well past 100% with no cap, reinsurance, or exclusion to blunt it, the risk is undercapitalized for a startup balance sheet regardless of how it looks on paper.
• You can't explain who is adversely selecting into the product. If you can't articulate why the customers most eager to buy aren't also the riskiest ones, assume they are, and re-run the model with a worse frequency assumption for that segment.
• The CAC/LTV math looks fine but never touches loss ratio. Software-style unit economics — the CAC, LTV, and payback period framework most startups default to — can show a healthy customer lifetime value while completely ignoring that the product loses money on claims before acquisition cost is even in the picture. In insurance, loss ratio and combined ratio have to clear their own bar independently; a strong LTV does not offset an unprofitable underwriting result, it just means you're acquiring unprofitable customers efficiently.
• No honest input set gets you under 100%. If your most optimistic, still-defensible frequency and severity assumptions can't produce a combined ratio with real headroom, the idea isn't underpriced — it's structurally unworkable at the risk you've defined, and needs a different risk, segment, or model rather than a better spreadsheet.

Key Takeaways

• Loss ratio is LR = L / P — expected losses divided by premium earned — and it's the first number that can kill an insurtech idea outright, independent of demand.
• Combined ratio, not loss ratio alone, is the real bar. CR = LR + ER; a startup usually needs real headroom under 100%, since it rarely has the investment float larger carriers lean on.
• Frequency and severity are estimates from proxies, not guesses. Adjacent product data, operational logs from the underlying activity, and conversations with reinsurers or MGAs beat an unverified single source.
• Model severity as a distribution, not one average. Attritional and tail losses behave differently, and a blended average can hide the tail event that would sink a thin balance sheet.
• Stress-test your base case. A model that only survives if your first guess is exactly right isn't a validated economic case — check what happens if frequency or severity comes in meaningfully worse.
• A defined risk is a prerequisite, not paperwork. Precisely naming the peril, population, trigger, and period is what makes every number after it meaningful — and it's what any capacity partner will demand before backing you.
• Good CAC and LTV don't rescue a bad combined ratio. Loss-ratio economics and customer-acquisition economics are separate gates; both have to clear, and neither substitutes for the other.

Frequently Asked Questions

What is a good loss ratio for an insurtech startup?

There's no single good loss ratio — it varies enormously by line of insurance, distribution model, and time period, and a benchmark from one line can be actively misleading applied to another. Rather than anchoring on an industry-wide figure, model your own expected loss ratio from your specific frequency and severity estimates, add your expense ratio to get a combined ratio, and judge feasibility from whether that combined ratio has real headroom under 100% at a price your segment will actually pay. An actuary or a comparable filed rate in your state can tell you what's typical for your specific line far more reliably than a general rule of thumb.

What's the difference between loss ratio and combined ratio?

Loss ratio (LR = L / P) measures only claims paid against premium earned — it's the pure cost of the risk you're underwriting. Combined ratio (CR = LR + ER) adds your expense ratio — acquisition, claims handling, technology, and compliance costs divided by premium — to give the full picture of whether a policy is profitable to write. A low loss ratio can still describe a losing business if expenses are high enough to push the combined ratio over 100%.

How do I estimate claims frequency without any historical data?

Triangulate from proxies rather than guessing: data from an adjacent, already-written insurance line; operational records from whatever activity generates the risk (repair shops, fleet operators, service providers); direct conversations with reinsurers, MGAs, or brokers already writing something close to your risk; and, where the risk is cheap enough, a small contained pilot that generates your own real data. Use at least two independent sources before trusting the number, since a single proxy can be systematically off in a direction you won't notice until it's expensive.

Can a good loss ratio still mean a bad business?

Yes. Loss ratio only measures whether premium covers claims — it says nothing about acquisition cost, claims-handling overhead, technology spend, or compliance burden, all of which live in the expense ratio and the combined ratio, not the loss ratio alone. A product can also clear a healthy combined ratio and still fail commercially if distribution is too expensive or too slow to reach scale, which is why loss-ratio feasibility and customer-acquisition economics need to be validated as two separate, independent checks rather than assuming one implies the other.