Shilpi Bhabhra, head of the Analytics and Data Sciences at Acko General Insurance, was one of the guest speakers on the day 1 of the Deep Learning DevCon 2020 (DLDC 2020). Like the topic of her talk, ‘Who Priced My Insurance’, the content was equally engaging. Through her talk, she shed light on how digital insurance was quickly changing the way risk is calculated and how the insurance amount ranged from known factors such as age and experience to colour of the car and the brand of the customer’s phone.
The speaker’s talk was broadly divided into three main parts:
- Background information on insurance
- How machine learning is changing the game in the insurance sector
- Disruptive new-age technology.
Insurance & Ways to Calculate Premium Price
Insurance, as we know, is the contract between a customer and the insurance company where the former pays a premium to the latter on a regular basis. In case of a loss, this amount is reimbursed to the customer. Bhabhra mentioned an interesting point about how the insurance sector operates, that is, unlike other sectors, the concerned company has to decide the cost against possible losses first, and the losses happen later.
Explaining a few traditional ways of calculating the insurance pricing, the speaker told about the methods to calculate the price tag. The most basic break-even method, as Bhabhra explained, is the ‘no profit-no loss’ method. The issue with this method is that it is very simplified and may not be an entirely fair method. Bhabhra illustrated with an example of auto-insurance where both small and large car owners pay the same premium under this system, which is unfair.
To overcome this challenge, most insurance companies adopt cohort-based pricing. This type considers additional factors such as car model and age, showroom price, etc. However, even this method doesn’t correctly determine the most appropriate pricing as it does not consider individual-level risk factors such as the age of the user, how frequently the car is used, among other factors.
“This method of pricing leads to fair insurance, cheaper premium, and competitive advantage,” Bhabhra said as she explained about the true personalised pricing method that depends on the ‘individualised risk exposure of the customer’.
Explaining further, the speaker then gave the mathematical denotation of calculating the pure or the risk premium amount, given by,
Loss cost = Incident rate X Severity
Where the incident rate, also called the frequency of claim, is how likely the customer is to get into an accident and severity denoted the amount that is likely to be claimed after an accident.
Very interestingly, the incident rate can depend on seemingly unrelated factors such as whether you are taking a planned or sudden trip, whether you are married and have kids or even your road challan history. Similarly, severity factor can also depend on factors such as the colour of the car or the brand of phone you use.
Machine Learning in Insurance Sector
Next moving to the types of modelling techniques currently in use in the insurance sector, speaker Bhabbhra first spoke about the generalised linear model (GLM), which as per her is followed by 95% of the industry. Further, the following observations were made:
- GLM is the technique that has been in use for the last few decades
- Its popularity lies easy implementation and simpler interpretation
Highlighting the gradual move of the industry towards machine learning-based models, the speaker listed out techniques such as random forest, gradient boosting machine (GBM), generalized additive model (GAM), and neural networks. Citing a paper titled Boosting Insights on Insurance Tariff Plans With Tree-Based Machine Learning Methods, Bhabhra said, “While comparing different models — Tree, random forest, GBM, GLM, and GAM, for the same data, classical methods GLM and GBM outperformed tree and RF, however, GBM outperformed all of them.”
Apart from these techniques, Bhabhra also explained in detail about the SHAP analysis, which is quickly emerging as the preferred method for interpretation.
Reflecting on the implementation of machine learning-based models at her parent organisation, Bhabhra shared the following learnings:
- Concentrate on collecting data from day 1
- In the beginning, implement simpler models with limited variables
- Build several models and choose the best among them
In the last section of her talk, the speaker spoke of new-age technologies which are increasingly finding applications in the insurance industry. In particular, she spoke of telematics and usage-based insurance.
Telematics is a method through which a vehicle is monitored using GPS and onboard diagnostics to record its movement on a computerised map. “Many insurance companies are offering discounts to people with telematics devices installed in their vehicles. The other advantage of using telematics is that if you fall under the high-risk category, using telematics can help reduce the risk course,” said Bhabhra.
On the other hand, the usage-based insurance (UBI) system is a ‘pay as you drive and how you drive’ system. In this case, the risk is calculated in real-time that depends on distance travelled, where and how.
In her ending note, Bhabhra revealed that governments around the world are pushing for individualised pricing based insurance as risk-based pricing as in the case of UBI leads to careful driving and in turn, higher road security.