Yiannis Parizas and Phanis Ioannou weigh up the benefits of open-source and commercial solutions for general insurance pricing
Actuaries have been data scientists since long before the term became popular, and the pricing process has been a complex machine learning pipeline for decades. Infrastructure ranges from simple multi-way rate books, crafted using expert underwriting experience, to data-driven, nested deep learning models. During the past decade, we have seen more insurers and products moving towards advanced infrastructure architectures and using more complex models. This has created business opportunities for vendors of deployment infrastructure and modelling solutions that are specialised for non-life insurance. We spoke to several of these companies and compared the merits of commercial solutions to an open-source route.
Current market practice
Developed insurance markets are dominated by commercial deployment vendors. An insurer’s choice of pricing implementation solution is driven by the market’s availability of people who have expertise in the proposed solution. As such, the software with the largest trained community in the local market has the advantage. Insurers seeking custom solutions and use of newly developed algorithms such as machine learning sometimes opt for open-source solutions. These insurers are usually insurtech start-ups that want to obtain a strategic advantage and build the skillset necessary to support such solutions.
Less-developed markets make less use of commercial vendors for pricing deployment. Those that use commercial vendors are usually the largest local insurers, or multinationals that can afford or understand the solutions’ benefits. Most of the insurers in less-developed markets provide a simple-to-code algorithm to the policy management system vendor to deploy it. Alternatively, a simple Microsoft Excel Workbook is uploaded and called by the policy management system, providing an easy and transparent deployment method.
Non-specialised policy management systems cause slow deployment. The infrastructure cannot support all types of models and is not centrally controlled by pricing or underwriting teams. When talking to different vendors, it was clear that clients deemed deployment speed and control as necessary to react to market changes quickly.
Commercial vendor solutions
We reviewed some of the popular commercial solutions for modelling and deployment.
All the vendors provide modelling solutions; most either provide deployment solutions or are working on providing automated deployment. All provide data pre-processing and visualisations, and support running future scenarios for testing pricing strategies and potentially plotting strategy frontiers. Most provide cloud deployment and version management control.
Different vendors have different business models. Some are more focused on their modelling algorithm. Others offer their platform as a data science pipeline to be used alongside their internal models – models to be built externally and then imported from commercial modelling vendors such as DataRobot or open-source tools. Most solutions support machine learning model algorithms: mainly gradient boosting (GBMs) and generalised linear regressions (GLMs). A benefit of GBMs and other automated machine learning models over standard GLMs is that they materially reduce the manual work required for model selection; as a result, the time and labour cost needed for modelling decreases. The drawback of GBMs is reduced transparency.
All the providers offer training for their products. Some offer consulting services, where they can build the models and infrastructure for clients. For some, modelling and deployment infrastructure is just one of their offerings, and they can cross-sell their products among other arrangements. For example, reinsurers or reinsurance brokers sometimes offer tools to their clients, promoting economies of scope.
We reviewed the software of different vendors and asked them to summarise the strategic advantage they offer over their competitors. The aim was to maximise the sample of reviewed external providers for this exercise, not to promote any particular vendor. The goal is to educate the market on what exists, to enable insurers to identify and fine-tune strategies and make informed decisions about their pricing infrastructure.
One contributor, Akur8, provides an integrated pricing platform that allows risk assessment, price sensitivity, pricing scenario testing, forecasting and optimisation of pricing strategy. Guillaume Beraud-Sudreau, founder and chief actuary, says: “The platform offers fully transparent machine learning capabilities to allow actuaries to accurately predict and understand the risk of their clients. This method increases modelling speed and accuracy without sacrificing the transparency of the models, effectively ensuring management of the adverse-selection risk.”
Another, Earnix, offers real-time artificial intelligence-driven rating, dynamic pricing, product personalisation, and fully operationalised telematics solutions. CEO Udi Ziv says: “With composable offerings, insurance carriers are empowered to break away from the status quo of legacy systems, transitioning to nimble solutions that can be rapidly deployed and respond to the needs of the market in real-time.”
Addactis embeds the pricing process in a single transparent solution, helping pricing actuaries to gain more agility and control over business objectives. Partner Stéphanie Dausque says: “Our end-to-end approach guarantees a shallow learning curve and quick adoption of rating models in production. Concretely, Addactis Pricing integrates all data processing, modelling and publishing capabilities to deploy actuarially sound sales strategies and ultimately embrace the global value chain.”
Finally, Quantee provides a ‘next-generation’ pricing platform that helps insurance carriers to improve model accuracy and enables instant deployment and real-time monitoring. Founder and CEO Dawid Kopczyk says: “Our platform is reducing the time-to-market for any pricing updates. More concretely, we support building a competitive, flexible and end-to-end pricing pipeline from data processing, through advanced modelling, price optimisation, open-source integration and exposing pricing to sales channels instantly.”
The open-source route
The two most popular open-source solutions for non-life insurance modelling and deployment are currently R and Python. They offer the latest modelling innovations before the commercial vendors, as well as the option for full customisation. Modelling and impact analyses can be run fast by using central processing unit cores in parallel, graphics processing units or cluster machines.
R can be deployed on Linux environments with Plumber, pm2 and nginx. Likewise, Python can be deployed with Flask, Gunicorn and nginx. Both would enable parallel load balanced deployment for high performance in servicing requests. Vendors for ‘cloud platform as a service’ such as Heroku are available to seamlessly manage and deploy the pricing web services. In higher demanding deployments, Kubernetes or Docker containers could be deployed in clusters. Library versions can be managed using Git, through GitLab, GitHub and others. Visualisations are possible in both R and Python and can be published as dashboards on Shiny (R), Dash (Python) and Streamlit (Python). Visualisations tools such as Microsoft Power BI or Tableau can be used alongside open-source modelling and deployment solutions to enhance the pipeline’s speed.
Open-source solutions have a larger online community than commercial vendors, offering similar examples from other industries, from which knowledge can be transferred. They allow for customisation solutions and speedy uptake of the latest methodologies. However, they need different professionals to support the pipeline and have a steeper learning curve than commercial software. As such, it would be more difficult to source people who have both the necessary advanced coding skills and business knowledge. This creates a key person’s risk.
Commercial software, on the other hand, is easy to use through graphic user interfaces, so the challenge is more about what to do than how to do it. As such, the cost of commercial software may be offset by the cost of diverse teams and skills necessary for the open-source route.
The decision for the pricing deployment route depends on the insurer’s strategy.
In our view, larger insurers that prefer a streamlined process, fast changes, accurate pricing and process control, and want to avoid any key people’s risk, should opt for commercial vendors. Insurers seeking full pricing customisation, implementing the latest technologies first and having technology skills available, should opt for the open-source route. Insurers with a simple pricing architecture and infrequent changes could deploy their prices within their policy management system as the most cost-efficient method.
Insurers sometimes use more than one solution, either for parts of the pricing process, for different products or for different distribution channels. To enrich their work opportunities, pricing practitioners should be up to speed with the current solutions. Decision makers for the pricing infrastructures should be aware of what is available so that they can choose the infrastructure that fits the business best and generates the most value.
Yiannis Parizas is an actuarial pricing consultant
Phanis Ioannou is a quantitative risk manager at Grant Thornton
Image credit | iStock