Anyone who relies on automated vehicle navigation systems whiledriving in congested traffic knows to disregard the principlethat the shortest distance between two points is a straight line.One-way streets, missed turns, and dubious directions often makereaching a destination by car seem to take far longer than byfoot.

The same could be said of traditional insurance ratemakingapproaches. Like a car navigating heavy traffic, the traditional ratemaking approach, otherwiseknown as multivariate regression, considers numerous ratingvariables to estimate risk. Sure, they may accurately reflect apolicyholder's loss propensity. But at times, it seems the path ismore circuitous than necessary.

Sophisticated data science

Now consider applying an alternative approach to usage-based insurance (UBI) programs, whichtypically offer discounts for safe driving behaviors confirmed byvehicle-generated data. UBI programs can be fertile ground to testvarious data science techniques that constitute the field ofadvanced pattern recognition, commonly known as machine learning.By using an alternative approach that relies on sophisticated datascience techniques and considers potentially more powerful variablecombinations, one could conceivably arrive at similar, if notbetter, results — and avoid unnecessary twists and turns in theprocess. Think of the swift-footed traveler.

This approach applies such sophisticated decision-making toolsas decision trees, random forests, and neural networks. These toolscan be applied to help the inquirer quickly discern the set ofvariables that most closely "maps" to each policyholder's estimatedloss propensity. Such an application would reduce the need to useevery significant variable that could potentially apply to eachpolicyholder — which occurs with traditional approaches even thoughsome variables might not always be relevant for all cases. 

To see how machine learning is applied, in contrast to a moreconventional method, consider the following example involving ahypothetical auto risk:

• An insured vehicle is garaged in a traffic-congestedmetropolitan area.
• Its primary operator is a 19-year-old who recently obtained adriver's license.
• The vehicle is usually operated between midnight and 4a.m.
• The operator makes several sharp right turns during eachtrip.

It's easy to see from a traditional underwriting perspective whyan insurer might assess a high premium for this hypothetical risk.Typically, the territory base rate may be elevated due to anincreased likelihood of crashes and higher cost of living in somemetropolitan areas. Also, younger or inexperienced operators mayexhibit higher accident rates, justifying a surcharge. Finally,driving during hours when there's lower visibility and taking turnsquickly may both demonstrably amplify risk and potentiallydisqualify policyholders from more significant UBI savings.Considering each risk factor more or less in isolation, an almost"perfect storm" of risk surfaces. This analysis reflects what datascientists might refer to as multivariate regression and resemblesthe way many insurers set rates.

Continue reading…

In this May 23, 2014 file photo, traffic moves on theInterstate 495, the Capital Beltway, in Hyattsville, Maryland,outside Washington. (AP Photo/Carolyn Kaster,File)

The same insured risk characteristics can be viewed differentlythrough the lens of machine learning, where we attempt to createthe optimal map to the policyholder's estimated risk rather thantaking a series of winding streets. This perspective suggests thatsince the vehicle is garaged in a traffic-congested area, it's bestto drive at night because fewer cars are on the road. Thepolicyholder's youthful inexperience could be viewed favorably withnighttime driving when correlated with improved eyesight andreaction time.

Finally, making sharp right turns might be expected on narrowcity streets that intersect at tricky angles — and also farpreferable to left turns, which some generally regard as riskier. Looking at ahomogenous cohort of policyholders who exhibit these exact samerisk factors, it might be observed that this hypotheticalpolicyholder, on average, experiences lower losses than thetraditional approach would predict.

Decision tree application

If we were to describe the previous paragraph as an applicationof machine learning, it would essentially be a decision treeapplication. A decision tree algorithm generally helps the inquirerto discern which risk factor, such as local traffic density, bestexplains losses for a particular cohort and then divides the sampleinto two (or more) groups — such as high versus low.

Within each subgroup, the method similarly identifies the mostsalient explanatory factor, such as whether the primary operator isyoung. Then within the youthful subgroups, the method may identifythe most salient explanatory factor, such as whether the vehicle istypically operated late at night. Note that the risk factor neednot be the same for each subgroup. The automated process continuesin this manner until a human intervenes or until the automatedprocess cannot continue splitting information in this manner.

The elegance of this approach is that, in a low-traffic-densitysubgroup, predictions may hold true about youthful inexperienceamplifying risk, with risk increased by late-night driving —corresponding with the predictions of the traditional multivariateregression. But for the higher-traffic-density subgroup, youthfulinexperience may prove less of a factor, and (when combined withlate-night driving) may significantly reduce loss propensity.

In other words, changing one risk factor, even with otherelements being equal, can mean a world of difference. By defininggroups based on risk characteristics, decision trees have thepotential to save us from making that one "wrong turn" that couldconceivably have us traveling quite a distance in the wrongdirection.

With respect to UBI, machine learning approaches can revealperceived drawbacks as potential pluses. For example, a demonstrably risky driving behavior, such asapplying intense braking pressure, may occur infrequently andperhaps even be misinterpreted when it does occur. Consider if avehicle's brakes are applied to avoid a pedestrian collision in anarea where there's a high volume of foot traffic. One could arguethat defensive braking under such circumstances should not bepenalized.

More variables into risk equation

A decision tree may identify braking as less significant inareas with high pedestrian traffic and, in turn, bring othervariables into consideration under these circumstances, such as howfast or often one drives during congested times of day or whenpedestrians are more likely to enter the road unexpectedly. Takingsuch an approach has the potential to bring more variables into therisk equation, which, from a UBI perspective, may be one potentialbenefit.

There will likely always be some people who prefer to use theirown feet (and intuition) as opposed to relying on vehiclenavigation. Similarly, data scientists may have valid reasons forelecting multivariate regression over machine learning — or mightopt for some combination of both. In any case, applying machinelearning to UBI data demonstrates that, in some contexts, theoptimal path to ratemaking is the one less traveled, however thetraveler gets there.

Related: Here's what you don't know about today's personallines business

Jim Weiss is director of analytic solutionsfor Verisk Insurance Solutions, a Verisk Analytics(Nasdaq:VRSK) business. Udi Makov is head ofactuarial research, Verisk Telematics Innovation Center, Tel-Aviv,and professor of statistics, director of the Actuarial ResearchCenter, University of Haifa. Opinions expressedin this article are their own.