Вы находитесь на странице: 1из 4

Case Study

About The Client


The client is one the largest and profitable insurers of India. As a joint venture between one
of India biggest banks and a European financial conglomerate, it had over 3 million active
policies at the end of 2014-15.

The Challenge: Claims Fraud Prediction and


Growth of New Business by
improved Cross Sell
The client has been in business since 2005. With a persistency rate of 69%, the client has an
above industry average retention rate.
However, over the last few years they have suffered from claims fraud which has impacted
their revenues. In the past couple of years, the client has been able to identify fewer than 100
cases of early claims fraud. An early claim is defined as a claim which has been intimated
within 3 years of its issuance.
The client receives about 700 early claims a month, and they wanted a predictive model which
would help them predict the probability of a claim being a fraudulent claim. On the other
hand, the client was also looking at decision models that would help them cross sell better by
identifying closer prospect product match.

Approach: Identifying Fraudulent Claims


Aureus took two approaches to build a Claims Fraud Model one using Logistic regression
and the other by using decision trees. Both these models would run as soon as there was a
claim intimation and categorize the claim in one of three buckets Red (high probability of
fraud), Amber (Medium Probability) or Green (Low Probability).
The approach used to develop both models was similar:

Business
Understanding
Deployment

Fraud
Prediction

Data
Understanding

Data
Preparation

Evaluation
Modelling

www.aureusanalytics.com

Five main data sets were used to for the modelling purposes Customer Data, Policy Data,
Agents Data, Products Data and Claims Data. From these sets a few derived variables were
also created that would help with the claims fraud prediction. A total of 29 different variables
across these buckets were used for modelling.

Customer

29
Significant
Variables

Claims

Product

Policy

Agents

The base data set was divided into three main parts training data set, validation set and test
data set. Models were built for both the training and validation data sets and only those
variables which were significant on both sets were retained.
Both models delivered nearly 90%+ accuracy on the test data set.

Key Benefits
The model built by Aureus identified nearly 10-13% probable fraud cases. This is
a substantial improvement over manual identification of fraud cases

The model will save the client nearly USD 400, 000 in terms of effort
spent in investigation of genuine cases.

Overall turn around time will improve as fewer cases will


now be referred for further investigation.

www.aureusanalytics.com

Making Cross Sell Effective


The other challenge that the client wanted to address was how to make cross selling more
accurate and effective. The cross sell model was designed to answer three primary questions:

WHOM TO
SELL?

WHAT TO
SELL?

HOW MUCH
TO SELL?

The model developed would attempt to answer all the three questions together for large and
medium ticket size product categories. The approach to make an effective cross sell model
was straightforward.

Input = Data Used

Process - Random
Forest Algorithm

1. Policy
2. Customer Information
3. Family History

Predictive Model

Output
For each customer,
the model gives the
probability of the
16 CATEGORIES

Figure 1: High level Approach


Developing the model was an iterative process and used randomly selected samples for training and testing.

Dened
Veriables

Back Testing
& Turning

Derived
Veriables

Model
Development

Insights
Development

The significant variables identified in the modelling runs were used to further finetune the model. The model
accuracy ranged from 90% for high risk products to 98% for low or medium risk products.

www.aureusanalytics.com

Key Benefits
The cross sell model has reduced the efforts wasted in miscalling by nearly
60%. This will translate into substantial cost savings.

The precision of pitching the right product to the right person has improved
by 5x.

The number of products that are now pitched to a customer have been
reduced by nerly 70% thus giving the agents a more focused scope to work on.

Web: www.aureusanalytics.com
Stay Connected With Us:

Email: info@aureusanalytics.com

@AureusAnalytics

/company/aureus-analytics

Вам также может понравиться