A smarter approach to analysis refines insurance predictions
There's an improved model for understanding the main drivers of losses
New technologies are delivering a deluge of data, with many companies struggling to harness useful insights that can deliver a business benefit.
The insurance industry is particularly affected by the sheer volume of data, some of which is entirely new, and the challenge is on to leverage this data for improved outcomes.
One area that has potential is the prediction of cause and effect, but to achieve this the industry needs new approaches to data modelling.
A team of scholars in the school of risk and actuarial studies at UNSW Business School has taken on this challenge, and has created an improved inference and prediction model for insurance losses. An initial modelling tool is now publicly available, easy to digest and available for download.
The model is set out in a paper by Benjamin Avanzi, Greg Taylor, Bernard Wong and Alan Xian, and was awarded the Taylor Fry Silver Prize at the 2018 General Insurance Seminar in Sydney.
The Business School team worked with three of Australia’s largest general insurance companies to produce the paper: How to proxy the unmodellable: Analysing granular insurance claims in the presence of unobservable or complex drivers.
The companies provided data on a daily basis, whereas other prediction models use aggregated data that tends to be either quarterly, half-yearly or annually.
“Few data analyses have taken place at this level of detail,” says Wong, an associate professor and head of the school of risk and actuarial studies. “It was a unique collaboration with industry partners.”
While companies have this data at their disposal, its full potential has yet to be harnessed.
“There are lots of data pools that haven’t been linked up with different business data – that’s where we come in,” says Wong.
Sometimes, aggregate models are sufficient, but often the analysis benefits from more sophistication.
'It is much easier than doing a deep dive into the data and trying to figure out any small interactions'
ALAN XIAN
Residual information
The new model enables insurers to improve their risk management systems and that in turn will improve calculations which can determine pricing of insurance claims.
“[The model] will provide better insights on insurers’ processes by giving them a better tool to understand what is important and what they can’t model. Rainfall, for example, certainly has an impact [on claims] but it is hard to model. This is where our methodology can help,” explains Wong.
The model contains two parts. The first is what the practitioner has chosen to model explicitly, while the second part includes a proxy, which serves for any residual information. Residual information may be unidentifiable or highly complex.
The idea is to model whatever is possible, and to proxy the rest. By contrast, residual information has previously been left out because it wasn’t possible to model it.
“Our idea is that you don’t have to neglect that information. You can proxy it instead, and hopefully that will give you more insights into your modelling predictions,” says Xian.
He adds that this approach provides a better understanding of the random processes that drive whatever observations an insurer is trying to understand.
“It is much easier than doing a deep dive into the data and trying to figure out any small interactions,” he says.
Time being wasted
The model can be used to analyse something such as severity – that is, the amount insurers are liable for from a claim. Take El Nino, for example. In periods of El Nino, claims tend to become significantly higher.
“Severity has a strong correlation with El Nino periods. That’s a motivation for insurers to look more closely into climate change and its effects. The fact that we were able to infer this without actually putting it explicitly in the model is interesting,” says Xian.
Ultimately, the model helps insurers understand what is worth modelling and what isn’t, and prevents time being wasted on unhelpful data analysis exercises.
“The deeper into data you go, the more trends that come out. But the problem is that if the interaction doesn’t result in a material change for the final result, it wasn’t worth going through all that effort,” says Xian.
In the prize paper, the model has focused on frequency, which is whether something occurs or not. In the case of a car accident, for example, frequency as well as severity is important. Severity relates to the seriousness of the crash and the claim amount.
More research is being carried out so that extensions such as severity can be added to the model, and this will make it possible to analyse the full impact of a car accident. There is also the potential for the modelling to be applied to a variety of different events outside of the insurance industry.
'There is no one solution to any of this, but we think this is a valuable part of the toolkit and we would like insurers to try it'
BERNARD WONG
Predicting the future
“While much energy has already gone into ensuring the model has a high degree of practical usefulness, further work is required”, says Xian.
“We put a lot of effort into calibration times – meaning that it is possible to extract results in a reasonable amount of time. Currently it takes a couple of hours, and that is reasonable for a practical implementation. However, when the extensions are added it may take a bit longer,” he adds.
Xian says this will depend on how the team frames the model and handles some of the underlying computation. The intention is to keep timings to a minimum so that decisions can be made in a timely manner.
There is no doubt the new prediction model significantly advances the actuarial profession’s general knowledge, but it is not a panacea. More research will need to be undertaken to obtain a fuller picture for insurance companies, who are in the inherently complex business of predicting the future.
“There is no one solution to any of this, but we think this is a valuable part of the toolkit and we would like insurers to try it,” says Wong.