Deploying Machine Learning in claims reserving
15 August 2017
Many insurers are investigating how Machine Learning can be best deployed to both improve risk segmentation and enhance pricing models. Many insurers have already acknowledged the speed at which both supervised or unsupervised Machine Learning can be used to build new types of high-quality models, leverage Big Data, and identify new relationships between variables.
Relatively little has been made to date of the capabilities of Machine Learning or Artificial Intelligence to dramatically increase both the speed at which claims reserving can be undertaken, and the extent to which highly sophisticated automation can be introduced.
There are two key areas in which Machine Learning and predictive analytics can be used to enhance claims reserving:
- By using statistical case estimation, algorithms will generate estimated losses, taking away the need for claims handlers to manually set the case reserve for any given claim.
- By deploying advanced machine learning algorithms to determine the overall level of outstanding claims reserves required, leveraging traditional reserving techniques.
Machine Learning in case estimation
For types of liabilities for which there are very substantial volumes of data, such as motor bodily injury claims or household insurance claims, highly sophisticated algorithms can be fitted to predict the ultimate cost of a claim. This can be done by combining insurers’ historical in-house claims data with digitised, structured information from medical reports, loss adjusters’ reports etc. captured within claims files. The algorithms reflect both the characteristics of the claimant and the circumstances surrounding the event giving rise to the claim.
In relation to bodily injury claims, predictions of the eventual claims cost are largely driven by the nature and severity of injuries, cost of care, current earnings and age (and hence forecast longevity in the most severe cases). Costs are also driven by proximity of the claimant’s home to physiotherapy providers or other medical specialists, as needing to travel long distances to get specialist medical support can deter claimants from getting the necessary treatment to ensure a fast return to work. All of these variables are taken into consideration within the algorithms deployed. Following the Lord Chancellor’s decision on 27 February 2017 to reduce the discount rate used to determine lump sum payouts, claims handlers have had to manually revise case estimates for those claims impacted by the Ogden ruling. Using statistical case estimates (SCEs) would remove this need, as the SCEs can be refreshed instantaneously to reflect a new discount rate.
In other applications, AI can be deployed to teach algorithms how to determine the extent of property damage from multiple images showing various levels of damage.
Determining outstanding claims reserves
In relation to regular reserving studies for financial reporting or internal management information purposes, this is currently a significant burden on actuarial reserving functions. Many insurers segment their business into a high number of portfolios for reserving purposes, creating an onerous workload for reserving teams. In extreme cases, this can lead to limited time being spent sharing deep insights with the business.
The deployment of Machine Learning in the reserving process can deliver significant advantages as follows:
- Improvement of current methodologies to embrace robotics-based reserving
- New reserving methodologies can be introduced, using Machine Learning techniques
- More focused reviews, in turn allowing more time for deep dives on areas of greater uncertainty, while other, more stable lines are subjected to more automated assessments
- A significantly faster end-to-end process from data extraction to final results
- The ability to run the full reserving process each month or quarter as required.
To understand how these advances can be achieved, Machine Learning algorithms can be used to determine both selected claims development patterns and initial expected ultimate loss ratios (IEULRs) without human input, or with very limited input, depending on preferences.
Under traditional reserving methods, claims development patterns are selected by:
- Considering the stability of development factors over time, and often using volume-weighted averages over a three or five year period
- Focusing on minimising the residuals and achieving consistency where appropriate.
Under robotics-powered actuarial reserving, algorithms can pick up changes in the speed of claims development over time, for example, and split out specific underwriting or accident years for separate treatment. Under traditional approaches, this is currently manually intensive and involves significant subjectivity.
Additionally, Machine Learning methods such as an Age-Period-Cohort (APC) approach use Machine Learning algorithms to produce more sophisticated versions of traditional techniques, which can reflect trends and correlations across underwriting or accident years, for example.
In selecting the IEULR, under traditional methods, rate changes and claims inflation are considered to reflate historical ultimate loss ratios to current pricing conditions. Changes in the mix of business (including new business vs renewals), premium volumes, average premium and deductible/limits etc are considered, as is a historical analysis of plan loss ratio against eventual experience.
Under Machine Learning techniques, some of the more subjective elements outlined above can be automated. For example, real time analysis of differentials between technical/benchmark prices and the prices at which business is actually being put onto the books can be used to adjust -- in near real time -- the IEULR away from the plan loss ratio. The plan loss ratio itself would have been predicated on the expected level of pricing adequacy to be achieved.
Machine Learning offers multiple opportunities to the insurance industry to augment the roles currently being performed in each of pricing and reserving. Achieving the full extent of the potential benefits from the deployment of Machine Learning will, however, require organisations to embrace change and to invest in the liberation of unstructured data.