# Reading the actuarial data stack that prices every policy
A driver in Ohio pays $1,240 a year to insure a 2021 Honda Civic. That number is not a guess. It is the output of a data pipelinedata pipelineETL (Extract, Transform, Load) is a data integration process that pulls data from sources, reshapes it into a consistent format, and writes it into a target system.Voir la définition complète → that starts with tens of thousands of past claims and ends with a single figure on a renewal notice. This lesson traces that path.
By the end, you will be able to read the three data structures that do the heavy lifting: exposure tables, loss triangles, and development factors. These are the actuarial data stack, and every priced policy sits on top of them.
An insurer cannot price "a driver." It prices exposure: a standardized unit of risk over time. In auto, the standard unit is the earned car year (one vehicle insured for one full year).
Two six-month policies equal one earned car year. This matters because all the math downstream is expressed per unit of exposure, not per policy or per customer.
So the first table in the stack is an exposure table: a count of earned car years, sliced by rating variables.
| Territory | Vehicle age | Driver age band | Earned car years |
|-----------|-------------|-----------------|------------------|
| Urban | 0-3 yrs | 25-40 | 4,820 |
| Urban | 0-3 yrs | 41-60 | 6,110 |
| Rural | 0-3 yrs | 25-40 | 2,340 |
The exposure table is the denominator for everything. Losses only mean something relative to how much risk you were carrying.
The actuary's target is the pure premium (also called the loss cost): the expected claims cost per unit of exposure.
Pure premium = Total losses / Earned car yearsIf the Urban / 0-3 yr / 25-40 cell had $3,180,000 in losses across 4,820 car years, the pure premium is about $660. That is the raw cost of risk before any expenses, profit, or taxes.
But there is a problem hiding in "total losses." Claims are not fully known when you need to price. That is where triangles come in.
Here is the uncomfortable truth of insurance: when a car crash happens, you do not know what it will cost. A fender bender might settle in three weeks. An injury claim might take four years, and the final medical bill can climb the whole time.
So losses develop. The amount you know today for an old year will grow as claims mature. Actuaries track this with a loss triangle.
Read it like this. Each row is an accident year (the year the loss event happened). Each column is a development period (how many months have passed since that year began). The cells show cumulative losses reported so far.
Cumulative reported losses ($000s)
| Accident year | 12 mo | 24 mo | 36 mo | 48 mo |
|---------------|-------|-------|-------|-------|
| 2022 | 1,000 | 1,400 | 1,540 | 1,586 |
| 2023 | 1,050 | 1,470 | 1,617 | |
| 2024 | 1,120 | 1,568 | | |
| 2025 | 1,180 | | | |
It looks like a triangle because recent accident years have had fewer months to mature. We know 2022 at 48 months, but 2025 only at 12 months.
Notice the same pattern in every row: the number grows over time, then flattens. That growth is what we need to measure and project.
For a clear visual walkthrough of building one of these:
🎬 [VIDEO: "How to Create a Loss Development Triangle" — youtube.com — a step-by-step build of a claims triangle from raw loss data]
To fill in the blank corners of the triangle, actuaries calculate age-to-age development factors (also called link ratios). Each factor answers: from one development period to the next, by what multiple do losses grow?
Factor(12 to 24) = losses at 24 mo / losses at 12 moUsing the triangle above:
The 12-to-24 factor is consistently about 1.40. Do the same for later columns:
| Period | Age-to-age factor |
|--------|-------------------|
| 12-24 | 1.40 |
| 24-36 | 1.10 |
| 36-48 | 1.03 |
| 48-ult | 1.01 (tail estimate) |
The last factor is a tail factor: an estimate of the remaining development beyond your data, because a few claims settle even after 48 months. Tail factors are judgment calls informed by industry benchmarks, and actuaries flag them as estimates, not certainties.
Multiply the factors forward to get a cumulative development factor (CDF) that takes any immature year to its ultimate (final expected) value.
For accident year 2025, known only at 12 months, multiply all the factors ahead of it:
CDF = 1.40 × 1.10 × 1.03 × 1.01 = 1.60
Ultimate 2025 losses = 1,180 × 1.60 = 1,888 ($000s)This is the chain ladder method, the most widely taught loss-reserving technique. The Casualty Actuarial Society publishes free foundational material on it and other reserving methods.
The key insight for a non-technical reader: the $1,180 you observed for 2025 is not the cost of 2025. The cost of 2025 is $1,888. Pricing off the raw, immature number would badly undercharge.
Ultimate losses tell you what past accident years cost. But you are pricing a policy that covers a future period. Two things change between then and now:
Actuaries apply an annual trend factor to bring ultimate losses to the future policy period. If loss trend runs at an estimated 5% per year and you are projecting two years forward:
Trended losses = Ultimate losses × (1.05)²Trend is one of the most scrutinized assumptions in a rate filing, because a small change compounds into large premium swings.
Vérification des acquis
1. Why do insurers price 'exposure' (earned car years) rather than pricing individual drivers or policies directly?
2. Two six-month policies are treated as one earned car year. What does this convention primarily ensure?
3. The pure premium is calculated as total losses divided by earned car years. Why is it described as the cost 'before any expenses, profit, or taxes'?
4. Select ALL correct answers about the role of the exposure table in the actuarial data stack.
Sélectionnez toutes les réponses correctes.
5. Select ALL correct answers about why 'total losses' is a problematic input for pure premium if taken at face value.
Sélectionnez toutes les réponses correctes.
Now stack the pieces. The indicated pure premium for our rating cell:
Trended ultimate losses / Earned car years = expected loss cost per car yearSay that lands at $700 per car year. That is not the price yet. Insurers add a loading for costs the loss cost does not include:
A common way to convert loss cost into premium is the permissible loss ratio. If an insurer expects 70 cents of every premium dollar to go to losses and LAE, the remaining 30 cents covers expenses and profit.
Premium = Loss cost / Permissible loss ratio
= $700 / 0.70
= $1,000 per car yearLayer on the rating variables specific to our Ohio driver (territory, vehicle, prior claims, credit-based insurance score where legally permitted) and you arrive at the $1,240 on the renewal notice.
In most US states, auto rates must be filed with and approved by the state insurance department before use. The filing shows the triangles, factors, and trend assumptions. Regulators check that rates are not "inadequate, excessive, or unfairly discriminatory," the standard language in most rate laws.
This is why the actuarial data stack is not just internal plumbing. It is the evidence trail that justifies a price to a regulator and, ultimately, to the customer.
Traditional triangles use aggregate data. Insurers increasingly layer in granular signals: telematics (driving behavior from an app or device), individual claim characteristics, and machine learning models for frequency and severity. But even the most advanced pricing model must reconcile back to the same reserving discipline. If your fancy model prices below your developed, trended loss cost, you lose money. The triangle keeps everyone honest.