Understanding Risk with Z Loss Inspector

The main objective is to take better decision in Underwriting risk, by that we mean: risk appetite, wording, deductibles, etc.

Also, as secondary objectives we wanted to:

• have a complete property loss map: by claim type, claim cause, activity, segment, customer and region;

• improve Risk Engineering Efficiency by focusing on claim reduction strategies.

• provide insights to our customers: mapping typical claims by activity, geographical location, segment.

Now lets take them one by one as I believe there was quite a lot of information on the last slide.

What is Intelligent Document Capture? Intelligent Document Capture is a software we developed in-house. It takes data from Loss Adjuster Reports (for IT geeks: in different image formats and pdfs) and together with claims & policy information we extract the claim type, claim cause and location as it is. This extraction is only possible after a step which is involves a software called OCR or Object Character Recognition, which translated an Image into it’s textual representation (here we used an open source solution - Tesseract OCR -, although we could also use Google Cloud Vision OCR or any of the other paid OCRs available on market) so that a machine can therefore understand the information inside. Also, without advanced NLP techniques or Natural Language Processing it would have been close to impossible to extract relevant information from text.

In order to standardize the information extracted we classify it (according to predefined types and causes by risk engineering and portofolio teams). This classification step is done using a hybrid approach: both using an expert rule based system and also using machine learning. The machine learning step involved having close to 600 manually classified reports by the risk engineering team and continuous feedback in order to improve accuracy.

Last but not least the location needed to be normalized, this last step was done using Google Maps Cloud functions which return us the address in a standard form including the geolocation.

We are proud to tell you that we finally reached an accuracy in both the claim type and cause of more than 90% and of 65% in the case of the location. If the results sound to Good, well they do sound very Good to us as well as previous research shows us that getting an accuracy of more tan 70-80% is considered very Good for text mining.

So far we learnt how to convert unstructured data to structured data and performed classification in order to understand the type and cause of the claims.

With this valuable information we moved to the Phase 2 of the project which is to understand risk and we do so by using two methodologies:

I will explain first methodology here which is 'Risk Insights'. Here we gathered all the information related to the claims : in the data sources we took data from the Phase 1 of the project which is Intelligent Document Capture . From here as I explained earlier we took the geo-location of the claim, the type and cause of the claim.

Along with this we also integrated many other data sources such as structured policy data and some auxiliary tables in access database. These additional databases provided enriched data about the policies, customers, customer segmentation and basic information about the claims.

There were many challenges in integrating these different data sources regarding the format and granularity of the information stored in these databases. Also we faced a lot of data quality issues most of which were fixed using data cleaning techniques and implementing some business logic.

Once we had all required information , we build a Tableau Dashboard to visualize data in an easily understandable format. The dashboard contains multiple reports each giving a risk insights along different dimensions.

This tool helps the Risk Engineers to explore risk data and get some valuable insights.