The following section gives examples of machine learning applications for
banks / insurances in more detail.
1) Fraud detection
Fraud takes place in most insurance categories. For an insurance company it is difficult to know which claims a clerk should investigate more carefully. Or: Knowing which claims are those with a high probability of being fraudulent prior to investigating the claim in more detail.
The insurance company constructs a predictive model for the probability of a claim being fraudulent using machine learning and data from the past. Data like age, income, family status, data describing the claimant's financial history and data describing the claim together with the information whether a claim in the past
was fraudulent or not is taken to construct the predictive model.
By applying such a predictive model to future claims to estimate the probability of fraud the clerk's administrative efficiency is usually increased by a multiple.
2) Credit scoring
Banks use past customer data to estimate the credit worthiness of new credit applicants. For this task linear scoring models are widely used. However, today there exist much better machine learning algorithms for this.
A bank has data about past credit applicants describing the applicants' personal status, financial history and behavior together with the information which application was approved and which customer payed back the credit. Using this data and machine learning one can construct a predictive model for the probability of a customer paying back a credit.
Then this predictive model can be applied to all future credit applications in an automatized way.
With reinforcement algorithms it is possible to optimize the strategy to approve a credit to customers to which in the past the bank would not have considered giving a credit.
3) Estimating the probability of a case of damage
Insurance companies (like car insurances) face the task of correctly predicting the individual probability of a case of damage for each client. So far insurance companies mainly use simple linear scoring models and have not yet
exploited the potential improvement by using more sophisticated machine learning algorithms.
A car insurance company wants to build a model to predict the probability of a case of damage within the next year for a new customer. The insurance company uses the data of customers in the past that was available at the time the insurance policies were underwritten. Family status data, data about the customers prior financial history, data about the car type is taken together with the information whether a customer had a case of damage within the next year or not. With this data a machine learning model for the probability of a case of damage is built.
Then this model is used in the future to predict the risk class of every new customer.