Product Focus – Theft model rebuild

resonateABI figures show that theft from households accounts for 13% of all claims received. Although the volume of theft claims has been falling in the last decade, it is still significant and amounted to over £440 million in the UK over the last year on property related claims. Having an accurate perils rating model that can differentiate risk at a highly granular level can make a considerable difference to improving loss ratios and boosting profitability.

The Business Insight residential theft model ‘Theft Insight’ predicts the relative risk and variation of domestic burglary across the UK and is currently used across the industry by sixteen major property insurers.

Business Insight also has a commercial property theft risk model specifically for commercial property insurers.  Both models are based on extensive research into crime patterns using the latest available data and take account of the changing economic landscape of the UK. This covers a cross-section of inner cities, large towns and suburban neighbourhoods through to small towns and more rural areas.  Built from high resolution spatial and demographic data and calibrated using sophisticated mathematical techniques, the models produce estimates of risk on a street by street level across the UK.

At Business Insight, we know our products need to add value to insurance company pricing and they also need to beat insurers own in-house actuarial models for an insurance company to licence our products as external data feeds.  Consequently, we invest significantly in R&D to ensure that our products help insurers maintain a competitive advantage.

Some vendors build a peril risk model which is a static product with little or no further refinement. Once built, the predictive accuracy of a perils risk model degrades over time so the continuity of development and focus on improvement and refinement is very important.

We are currently working on rebuilding our theft models using AI techniques, refreshed data and experimenting with a new level of geography that ensures the anonymity of people residing in those locations but that is also more powerful than current postcode versions. This will provide a deeper insight into crime and theft patterns across the UK and a higher level of predictive capability.

Contact the sales team for more information on 01926 421408.

AI and machine learning: things to consider

Companies are investing heavily in artificial intelligence and machine learning techniques.  Harnessing the value from data available internally and externally has become a business-critical capability for insurers. 

Using sophisticated methods and algorithms, machine learning uses automation to find patterns in data, not always obvious to the human eye. Data can be mined from a variety of sources to help insurers build a fuller picture of their customers and machine learning can be used in all areas of an insurer’s business from claims processing and underwriting to fraud detection.

An advantage of machine learning is that algorithms can potentially analyse huge amounts of information quickly. Solutions can be recalibrated and redeployed rapidly by automating a process without introducing human error or bias. The desire to uncover hidden patterns and discover something the rest of the market is missing is a key driver for many companies though it is easy to be seduced by the technology and the fear of not wanting to be left behind. There are pitfalls to avoid and sometimes it is all too easy to concentrate on the technology and lose sight of other perhaps more important pieces of the jigsaw.

Neural Networks
Business Insight has been researching machine learning techniques and has developed its own AI platform that can take large volumes of records across many variables as data feeds before iteratively learning from the data, uncovering hidden patterns and forming an optimal solution. The software can take a vast number of input data points and hundreds of corresponding risk factors per case before constructing a more accurate estimate of risk. The main advantage of the neural network platform we have developed is that it can potentially offer significant improvements in predictive accuracy compared to statistical data models. There can also be significant savings in time to rebuild and redeploy by the reduction in human involvement.

Traditional statistical methods require intensive manual pre-processing of input data to identify perceived potential interactions between variables.  Whereas a neural network needs minimal data preparation and interactions between variables drop out automatically which saves a considerable amount of time in model building. That said, you do need to ensure that you are not blindly seduced by the technology as there are other issues just as important when carrying out analysis of large databases.

Pearls of wisdom
Here are a few observations from what we have learned over the years that may seem blindingly obvious yet often get ignored, specifically:

1) Focus first on data quality
The validity, veracity and completeness of the underlying data you are feeding into the system is paramount. Whether internal data or external data feeds, data quality is essential. The saying ‘garbage in, garbage out’  is often true if the data you are using is of inferior quality. Hidden patterns are not ‘gems’ of knowledge but costly blind alleys if the data you are using is riddled with inaccuracies or is out of date.  Quality external data is becoming more easily accessible to the insurance market and investing in the best quality data will pay dividends over the long term.

2) Ensure the relevance of your input data for what you are trying to achieve
If you are asking the system to predict a particular target outcome you should ask:  Is the data you are utilising fit for purpose, is it relevant or sufficiently meaningful and is it representative relative to what you are trying to achieve?

3) Ensure you have the relevant knowledge and expertise to maximise the results
Though the technology is readily available, having people with a deep knowledge base, domain expertise and experience in this area is not something that is easily accessible in the insurance market. A deep understanding and knowledge of the market, the data and experience of why certain risk drivers happen is often under estimated.

The winners in the market will be those able to address these points focusing not on the technology in isolation but also the data, both internal and external, as well as attracting the best talent with the relevant domain knowledge and expertise to maximise value. Those that make sure they invest in the technology as well as the people and the appropriate data assets to drive their business forward, will be the winners in the years to come.

 

Data quality and models

Whether it is collecting live information on drivers through telematics, the use of geographic risk mapping data to assess property underwriting risk or having access to weather and event data that includes claims and loss data, there is no doubt that data coupled with technology solutions is driving change in the insurance industry.

At Business Insight, we understand how important good quality, accurate data is. Data and models need to be regularly maintained and updated as the quality of data on which a predictive model is built and run will have an impact on the quality of the predictions it makes.  Without reliable and accurate data, you could be basing your underwriting and pricing decisions on old or out-of-date information.  The saying ‘garbage in, garbage out’ is certainly true and having the best technology, mapping or systems is pointless if the underlying data is substandard.

Effective data management and data quality are core components of Solvency II.  Article 121 which governs statistical quality standards sets out the requirements that insurers should perform regular data quality assessments.

With this in mind, there are a number of factors which you should take in to consideration before licensing software, data and models.  These include:

  • When the model was built
  • How often it is updated
  • Who built the model
  • Whether they have they done this before
  • Level of analytical experience and qualificationsWhat factors an insurer should consider Final version smaller.png