The Great Hunt—A brief history of automobile rating variables

By James P. Lynch

Pull that old Buick Roadmaster to the pump in 1944 and you would get service: An attendant, maybe a team of them, would check the oil, clean the windshield, make sure the tires had air.

But you might not get much gasoline, particularly if the windshield bore a sticker with a mighty letter A. There was a world war going on, so these were the days of gas rationing. The armed forces needed petroleum to fight. Every drop that fell into your tank wasn’t fueling the Army’s Sherman Tanks or the Navy’s Flying Fortresses.

All that sticker got you was three gallons of leaded. (It was all leaded then.) That rationing sticker and the dollop of fuel it represented marked the birth of rigorous auto insurance rating in the United States. It’s a fascinating story in its own right, replete with false starts that show how even the earliest rating variables needed the characteristics that actuaries recognize as critical to what makes a rating variable effective.

Today, private passenger auto makes up about 40% of U.S. property/casualty premiums, according to the Insurance Information Institute, so it’s hard to picture the early 20th century, when automobile insurance was, to a certain degree, an actuarial afterthought. Most of the action was in workers’ compensation insurance, the government-born system designed to modulate awards to individuals shredded by the industrial era. Pore through the first two decades of the Proceedings of the Casualty Actuarial Society and you quickly realize the early casualty actuaries were sorting through how to do their job, teaching each other how to price and reserve workers’ comp.

 Casualty actuaries knew, of course, that some jobs were riskier than others, and they knew some employers provided safer workplaces than others. They had to learn to quantify the risk—create the data sets, and calculate and correlate before there were computers.

And there was less math to bring to bear. Probability concepts had been around for a couple of hundred years, but mathematical statistics was relatively new. Key concepts in experimental design were just being developed, and statistical credibility—a cornerstone of casualty pricing—was a fresh area of inquiry. Actuaries had to discover, then teach each other what today we’d call the basics.

It was hard to gather data. Everything was recorded on paper. It was hard to create a data set. There were no computers. It was hard to know what to put into a data set. What information seemed to predict the likelihood of a claim?

Auto insurance was an upstart line then, though it was growing fast. The number of vehicles registered in the United States grew to 10.4 million in 1921 from 3.5 million five years earlier.[1] Insurance wasn’t required—that didn’t start until the 1920s and really didn’t look like what we see today until the 1960s—but the concept of liability was well in place, so insurance was a good way to protect your other assets.[2]

Still, it was a simpler time. As proof, Table 1 represents the rating card for the entire United States for private passenger auto liability in 1919. There are two rating variables. Rate schedules 1-7 represent rating territories. The four groups represent the list price of the auto.

The two-variable system of rating pictured above persisted for decades. One variable was always territory. The other variable … well, it varied. Actuaries would try one. When it didn’t work, they’d try another, then another. They tried horsepower, then list price of vehicle. Eventually manufacturers were lumped into one of four rating symbols, labeled W, X, Y, and Z. Then Z got dropped, and W and X were merged, then the whole scheme was abandoned.

All the while, actuaries knew a source for more accurate rates: the driver. As early as 1922, actuary A.L. Kirkpatrick recognized the driver “is by far the greatest hazard to be considered in automobile underwriting.” Unfortunately, “there has been no satisfactory solution presented for getting the information which will enable the underwriter to determine which drivers are good risks and which are bad.”[3]

Not that they didn’t try. A 1929 plan offered a 10% credit to drivers who went accident free for two years. It failed, too.

The breakthrough came in Connecticut. The state DMV noticed that from 1932 to 1936, drivers under 25 were involved in 37% more accidents and 62% more fatal accidents than the average driver.

In 1939, a new rating scheme was born. Policies covering low-mileage autos with no drivers under 25 and that were not used for business were Class A-1. Other non-business autos were Class A. Vehicles used for business were Class B. Here are property damage pure premiums by class from 1939 to 1941.[4]

Note how Class A-1 (low mileage, no young drivers) consistently outperformed the other classes.

Of course, you can’t tell why it outperformed: Was it the requirement for low mileage or the absence of young drivers?

Also note the final policy year: 1941, the year the United States entered World War II.

Gas rationing soon followed. There were myriad rationing stickers, but we’ll focus on three. A-stickered vehicles were limited to three gallons a week. B-stickered vehicles were for commuters, and C-stickered vehicles were for business use.[5] B- and C-stickered vehicles were allotted more fuel.

Using the stickers, insurers could determine, at least roughly, how many miles a car was driven. Those three gallons only let A-stickered vehicles go maybe 50 miles a week. B- and C-stickered vehicles could go much farther.

Here are pure premiums for private passenger cars in Massachusetts for two war years, by type of rationing sticker:

The stickers—proxies for miles driven—were remarkably predictive. And the stickers didn’t consider the age of the driver.

Take a moment to consider, from the perspective of, say, 1946, the known facts:

  • Connecticut showed that a class made up of older drivers who didn’t drive much had favorable experience.
  • Massachusetts showed that, regardless of age, folks who didn’t drive much had favorable experience.

The conclusion, at least to some folks: Mileage was predictive; age was not. There was no need to surcharge young drivers.

There was just one thing those folks forgot: There had been a war going on!

That war had been fought, overwhelmingly, by young men. The United States had 6 million males between 20 and 24 at war’s end. Just more than half, 3.1 million, were in the armed services. By contrast, about 0.5% of females in that age group were in military service.[6]

When those men were fighting at Iwo Jima and Bastogne, they weren’t driving in Iowa City and Baton Rouge. With fewer young male drivers on the road, the age variable lost its predictive power.

With the end of the war, of course, gas rationing ended. Insurers remained thirsty for mileage data but needed a new proxy.

Massachusetts led the way. The state instructed insurers to collect information on driver age and expected mileage of cars by having brokers and agents obtain a signed questionnaire. Consider how difficult this must have been to administer. Agents meeting with customers, mailing out forms, mailing reminders, collecting the returns and sending them or their summary to a home office that recollated them and fed them into (non-computerized) data collections.

All for a topsy-turvy result, because the situation changed again. Age regained its predictive power. Miles driven lost its.

The former is easy to understand. The war ended. Young men came home. The ones under 25 drove the way that young men had driven before and, to a greater or lesser degree, continue to drive.

It’s also not too hard to see what happened to the mileage variable. The old variable—the sticker—was objective. Unless you could obtain prodigious amounts of gas from the black market, the sticker revealed how much you drove. After the war, mileage was a self-reported estimate of a number that most people don’t think much about. Even today: How many miles do you drive in a year?

Drivers, even of the best intent, didn’t estimate accurately. And not all had the best intent. If you underreported your mileage, you would save on insurance.

The experience drives home important characteristics of rating variables.

  • They must be objective. Both age and mileage are objective. You are how old you are. You drive how far you drive.
  • Just as important, though: The variable must be verifiable. Age was easy to verify: Check a driver’s license or take last year’s answer and add one. Mileage was more difficult. People gave wrong answers. Whatever their reasons, the inaccuracy in self-reporting confounded the variable’s predictive value.
  • And a final lesson: A rating variable must be inexpensive to collect and easy to administer. Again, it’s easy to administer the age variable. Gather the information once, and you are set. Mileage is trickier. The driver’s estimate of miles driven is hard to verify. You need two odometer readings a year apart, and you need them every year. And consider all the work in Massachusetts, all those records, all that collecting, collating, transcribing, reporting. It was clearly expensive, yet it wasn’t enough.

The concept of rating by mileage ebbed, with occasional attempts to revive it. In the 1990s, for example, California considered raising the price of gasoline to purchase insurance—pay-at-the-pump. Gas consumption would have acted as a proxy for miles driven. The proposal never went into effect.

Today, telematics have made monitoring mileage much more straightforward. Policyholders can allow a device in their car—usually the driver’s smartphone—to report travel information to the insurer: where the car is driven, how far and how fast, among other things. Even so, the insurer may occasionally ask the customer to report their odometer reading—to make sure no one is underreporting miles driven by shutting their phone off or leaving it at home.

The search for other rating variables continues. Laws that mandated auto liability insurance grew the marketplace and created larger datasets to analyze. The emergence of computers and their ever-compounding power have made the search easier. Mathematical insights have made it easier to isolate the predictive power of each variable, both on its own and in combination with other variables. Social concerns have thrust some variables, notably gender- and credit-based insurance score, into controversy. But the characteristics that make all rating variables work—they must be objective, verifiable, and easy to administer—remain as steadfast as the day more than 75 years ago when the homefront gas attendant peered at a windshield and saw a sticker bearing a capital A.

JAMES P. LYNCH, MAAA, FCAS, is owner of James Lynch, Casualty Actuary.


[1] Morris Pike, “Some Aspects of the Compulsory Automobile Insurance Movement,” Proceedings of the Casualty Actuarial Society IX, no. 19 (November 17, 1922): 23–37.

[2] Robert E Helm, “Motor Vehicle Liability Insurance: A Brief History,” St. John’s Law Review 43, no. 1 (July 1968): 25–33.

[3] F. R. Mullaney et al., “Discussion: The Development of Public Liability Insurance Rates for Automobiles,” Proceedings of the Casualty Actuarial Society VIII (1922): 301–7.

[4] Lawrence W. Scammon, “Automobile Accident Statistics by ‘Age of Driver,’” Proceedings of the Casualty Actuarial Society 37, no. 1 (November 17, 1950): 43–56.

[5] “The ABC’s of Gas Rations,” The New York Times, November 13, 1973, sec. Archives, https://www.nytimes.com/1973/11/13/archives/the-abcs-of-gas-rations.html.

[6] Thomas Parran, “Vital Statistics of the United States 1945, Part I,” 1947, 201.

Twists and Turns

The first rating plans were a leisurely stroll compared to the gantlet of factors customers face today. The development of factors had one big actuarial success and several efforts that were less successful.

The success: rating territory. Even 100 years ago, actuaries understood that where you drove a car affected your likelihood of being in an accident.

But they struggled to find variables to pair with territory.

First up: horsepower. Actuaries understood that speed kills, even when the speed is under 30 mph. They calculated horsepower based on formulas from the Society of Automobile Engineers. That worked for a while, but some manufacturers changed how they constructed engines. The formula stopped working.

Next up: the list price of the vehicle. The more expensive the car, the reasoning went, the faster it could go. That didn’t work too well, either. Car prices were rising, so a newer car—substantially the same as its year-earlier predecessor—would cost more to insure.

A new taxonomy emerged: the rating symbol. Autos were classified into five groups, based upon the vehicle’s make. Each manufacturer—I counted 150 in a 1919 rating plan—was classified by size and weight into four categories, denoted by symbols W, X, Y and Z. W vehicles were the slightest. Z represented powerful luxury cars like the Rolls-Royce. There was a fifth group for Ford, which dominated the market back then. (Later, Fords were moved into symbol W.)

This worked … for a while, and the WXYZ system evolved. By the mid-1920s, it classified according to list price, shipping weight, number of cylinders, and wheelbase. The Z variable was dropped. Then the W and X variables were combined. It wasn’t long before, as actuary Lawrence Scammon wrote years later, “a car was a car for insurance rating purposes.”[1]

The focus turned increasingly to the driver. In early days, a good-driver discount was at the discretion of the underwriter, but under a merit rating plan introduced in 1929, drivers who had been claim-free for two years automatically earned a 10% discount.

That plan, too, was a bust. Rating by driving record had the same problem then that it can have now: Lots of not-so-great drivers qualify. It was estimated that a single customer then made a property damage claim about once every 12 years. Some simple math tells you that something like 85% of drivers got the discount. And if 85% of drivers get a 10% discount, the remaining 15% have to pay around a 50% surcharge to collect the premium the insurer needs to cover all risks. So underwriters and agents found ways to squeeze even more customers into the discount. It was the Lake Woebegon of underwriting: All the risks were above average, with predictable results. The merit rating plan was withdrawn in 1932.[2]

We still have rating symbols, though developing them requires quite a bit more analysis than in the old days. At Insurance Services Office (ISO), for example, physical damage factors start with the sticker price but then account for vehicle experience. (Some models are more likely to get in a wreck. Some cost more to repair.) Further adjustments reflect the results of a predictive model that incorporates factors like curb weight and chassis type. For liability, the symbol reflects experience and the results of a predictive model.

The early struggles held important lessons. Rating variables need to be retested; what worked last year might not work this year, or next. And actuaries must be willing to adjust and abandon variables that don’t work.


[1] Lawrence W. Scammon, “Automobile Accident Statistics by ‘Age of Driver,’” Proceedings of the Casualty Actuarial Society 37, no. 1 (November 17, 1950): 43–56.

[2] “Auto Coverage Focus” (Best Ed Insurance Continuing Education), accessed September 24, 2021, http://www.bested.com/StudyGuides/CA-ACF/CA-ACF.pdf.

Next article A Perfect Example
Previous article Using Data for Good—Settlement annuities & factoring negotiations

Related posts