PRA Policy Statement 1/22: Insurance business transfers
We review the key amendments to the PRA’s Policy Statement 1/22 on insurance business transfers and consider the potential implications.
After I replicated a cup of coffee using my carbon nanotube molecular printer, I made myself comfortable in my virtual office.
Yesterday afternoon, I launched the daily reserve and solvency ratio assessment script for today’s official sign-off. Despite reserve and solvency assessments being derived each minute with updated information on claims activity and automatic payments (through the blockchain), the WIA (the recently launched Worldwide Insurance Authority) has been requiring formalized daily submission for three years now. The daily reports are made available to all customers and market participants for information and analysis.
This morning the algorithms have detected a number of fraudulent claims. The partnership between the WIA and Interpol has permitted database cross checks, effectively solving 73% of the claims that indicate fraud after having been flagged as such by the algorithm. Those claims have been excluded from the reserving exercise. As usual, claims about self-service flying scooters have also been segregated due to historical high fraud levels and limitations in the IA model which hinder its ability to provide reliable estimates - an area of current internal research and development.
To provide reserve estimates, a large collection of individual claim forecasting models are used that are directly sourced from algorithmically peer reviewed mathematical models and related algorithms contributed from the scientific community. If needed, three-dimensional holograms of the scientific contributor can be summoned in order to ask questions on the model or simply assist and provide training. The few insurance companies – the market has been structured into a small number of companies after a decade of M&As – pay on demand for this follow-up service, which has become the main funding source of public and private research labs as well as self-employed researchers in actuarial science, who constitute the most significant part of this community.
The claims and underwriting databases are continuously updated and processed to validate data quality – again, a regulatory standard – while another algorithm that is only used to control that the overall process is accurate also runs in parallel. Models are recalibrated using 10^9 parallel neural networks and the best calibration is retained. Auditors released to us their latest app for their pre-approval of our calculations in view of financial reporting use, which is now part of the production process.
Stochastic forecasts have been derived for 10^15 scenarios combined with a set of AI expert opinions. Those are based on recommendations from relevant algorithms about economic and behavioral assumptions, including hourly inflation outcomes and exogenous factors to claims magnitude, weather indicators, population density and infrastructure details in each 10×10x10 cubic meters. AI recommendations have become standard in the market since a recent survey paper showed the supremacy of the forecast over claim managers for most lines of business. The algorithms behind our AI recommendations, as well as the results of hourly back-testing, are available to all companies from the WIA website.
Like old fashioned airplanes in the 2000's that embedded duplicated electric systems for safety purposes, we have up to 10 (the minimum regulatory requirement is five) parallel and independent calculation streams for the financial statements production process in our company, involving separate server farms and independent AI ensembles (and actuaries). In a final step, an algorithm based on advanced Isolation Forests for anomaly detection helps the processes to converge in the final reserve position and solvency ratio. This is usually one of the last steps of my early morning routine.
I upload the report and sign off to the encrypted WIA folder for automated regulatory compliance validation. If the validation does not immediately pass, this is not blocking but increases the future regulatory checks required.
I now have virtual meetings to handle specific challenges for cyber, climate change and pandemics.
Today is a quiet day because only 254 cyber incidents have been identified among policyholders which trigger insurance cover. Because IT security and protections have progressed a lot, the human factor is even more sensitive in comparison; this is one of the biggest sources of attacks, which take the form of communicator spear phishing. In the last decades, cyber claim manager teams have grown in all insurance companies and ours is now the largest claims managing team in the market. We have continued to improve the way we equip clients with software to protect from attacks, as well as to identify/qualify those in case of occurrence, in order to automatize eligible claim payments. This also helps to intensify accelerated underwriting because the software is used to produce a cognitive mapping of the policyholder’s business and social networks, so that system entropy can be fully considered in setting the policy premium. While in the past cyber incidents originated from malware designed by human beings with bad intentions, there is now a collection of self-propagating malware, or SPM, that learn from network activity and “live their life” without human intervention while creating new malware themselves. Their existence is triggering responsibility issues; originally exclusion wording was used in the insurance contracts, but recently losses from SPM are increasingly being covered due to the high consumer demand and lucrative regulatory incentives.
The changed climate has profoundly transformed agriculture and home insurance. Agriculture insurance is now structured with refined parametric products that are triggered based on indicators (temperature, rain) collected via a network of polar orbiting environmental satellites. Each company now has its own meteorology service. This is, by the way, one of the remaining areas where structural mathematical models from physics are still employed and where better forecasts have not yet been achieved by AI algorithms. Our classical products now cover server crashes due to high temperatures, a recurring issue given the density of the network in our times and the need to optimize the use of air conditioning for preserving nutrition resources. Home insurance coverage for droughts have expanded as well. We are now tracking the personal ecological behavior (including sorting of waste and optimized use of water and heat) and some of the covers are now triggered only if a set of ecological behavior criteria have been met by the policyholder in the last year. This has been put in place by the government in a partnership with insurance companies in order to incentivize a behavior shock due to the urgent need for action based on ecological indications. This corresponds with a shift in the insurance paradigm, where the pooling of risks now holds between virtuous clients only.
My last meeting of the day is regarding the run-off of last year’s pandemic claims. The last decades have seen a particularly high frequency of such claims and the market has reorganized around this risk. In particular, insurance companies and governments have disengaged. The market is now merely structured with pandemic bonds, the first having been launched many years ago by the World Bank, following the Ebola pandemic that occurred between 2014 and 2016, with the objective to grant immediate access to life saving funds in case of future pandemics. In the current environment, pandemic bonds are issued by consortiums of professionals in several domains (restaurants, hotels, shops) looking to protect their interests, while financial investors aiming for high returns and diversification remain attracted by such products. A new prediction algorithm, revisiting the old « Google Flu Trends » approach, which ceased activity in 2015, helps in forecasting pandemic bond payoffs by tracking health behaviors and online searches.
I launch the script running the daily reserve and solvency ratio assessment for tomorrow because it is 3pm and time to « leave » the office—if I can say that. The office days are increasingly shifted towards an earlier schedule in order to limit the use of air conditioning and server activity during the prevailing high temperatures.