Over the past week, I had the great opportunity to present at two events – Celent’s inaugural Asia webinar series and the Toronto Machine Learning Summit for Finance and Insurance (TMLS) – and hear from distinguised guests and speakers. Topics discussed include innovation, complementary partnerships between insurers and insurtechs, data management, and data protection.
Asia Webinar Series on Innovation Ecosystem and Complementary Partnerships
Through the discussion with Jamie Macgregor, Celent’s CEO, we hold the perspective that the adoption of API still has potential for widespread adoption in the insurance industry. However, the rate of adoption, along with priorities, differs from countries, regions, and markets. For example, the US is focused on ecosystem development now, while Southeast Asia is focused on distribution, and global insurers may be ahead of regional Insurers in terms of adoption, partly to the availability of greater resources. We should also learn more from academia and adopt their research into the industry. Through understanding research maturity and sound experimentation, it can lead to production in the organization.
From the discussion with Winnie Wong, CEO and Executive Director of Asia Insurance and Avo Insurance, there are growth opportunity in Hong Kong for insurance, and with the support from the regulator, it is still a market to consider. Asia Insurance and Avo Insurance showcase how the incumbent can mix with the modern to offer the best of both worlds – an established business with a strong customer base going digital with a growing diversified product range and a new digital insurance proposition promoting first in the world P&C insurance such as e-wallet, and work from home protection. Ultimately, it is understanding the nuances of the region we choose to operate in and to know what works now and for the future.
From these discussions, we can see the following global trends such as the adoption of a B2B2C model. B2B may serve well for workflow enhancements of the organization and B2C provides the potentially larger customer base. Hence, there is a continued momentum of B2B2C markets to tackle the undeserved segment and this has been the recent chosen model as compared to B2B and B2C.
Separation between the established and the greenfield, as in the case of Asia Insurance and Avo Insurance, can be a suitable approach to enable focus and speed in targeting separate target customer segments. This is strengthened by partnerships as well, which provide the skills outside of the insurance domain. Partnerships help foster greater collaboration between the insurtechs and insurers, with the sharing of knowledge and complementary expertise.
The role of the regulator would hold greater emphasis, a sentiment we see from TMLS too. We must do things right and be open to fostering a comfortable relationship with the regulator as an insurer.
TMLS on Data Management, Protection, and Regulation
The TMLS Machine Learning in Finance and Insurance Micro Summit showcase various topics concerning ML adoption in finance. The continued discussion of MLOps, data lineage and traceability, and data protection echoes the perspective I believe will have greater emphasis for data, analytics, and AI application in the insurance industry for the long term.
Data Traceability and MLOps Challenges to Fintech
On the topic of experiment tracking and data traceability hosted by Clear ML, explainability is understanding why a model makes certain decision and this will help avoid the black box. And traceability is to understand what went wrong when model output seems suspicious. Data traceability is examining the lineage of data and to know where the samples of data are coming from. Data lineage saved and organized versions of datasets, code, packages installed, model’s result, and evaluation metrics. This provides experiment tracking and to monitor how parameters and approach changes when training and validating models.
On the topic of AI implementation challenges to Fintech and the finance industry, hosted by Superwise, is the differing view by different stakeholders such as data science, ML engineers, risk and compliance, and the business when looking at ML results. Appropriate explainability is important to cater to stakeholders’ concern. There are also challenges such as delayed results feedback and data drift.
Providing timely feedback to measure use case performance can be delayed. For example, obtaining ground truth for a fraud analytics model can be delayed which can cause inaccurate loan fraud assessment in an algorithmic automation loan system which can result in loan approval with a delayed check in fraudulent action. Proxy measures to delayed model feedback are to consider data quality metrics (missing data, new data, outliers), data distribution stability (covariate shift, out-of-distribution data), and model shift affecting results (classification ratio change). We must consider performance metrics measurement, as accuracy rate could hold different meaning when place in different context.
Missing changes in data is known as data drift as it is observed that data scientists examine data from a macro lens and may miss changes that occur at the granular level. This cause uncertainty which usually have no deterministic approach. Resolution approach include looking for changes in data values and volume.
Privacy Law for ML in Insurance
On the topic of privacy law for ML in insurance, the usage of data and ML for complex underwriting would need consideration of personal information privacy rights and risks in dataset. ML identifies novel risk factor and label health risk more accurately for better policy servicing, but we must be mindful when using data with personal information. This includes compliance with privacy laws of government regulations, commercial and contractual terms of data use, and understanding the reputational risk when using personal data. We would require good data management to built trust and transparency of users’ whose data are being used. Some rules on privacy laws relates to having meaningful consent or obtaining fresh consent when working on new data initiatives and to take into context when anonymizing data. Such privacy laws and consideration show the importance of having explainability in AI models.
Regulations are placing more emphasis on data protection for data privacy and to request insurers to include responsible AI metrics as part of training and production model development. Models must be monitored carefully according to the domain knowledge and we should know what is going on “under the hood”. We need to also understand what metrics lead to fairness or bias constraints.
Application of Data and Machine Learning for Actuary and Underwriting
During my presentation on ML Trends, Opportunities, and Challenges in Insurance, I received some questions that explores the perspective shared. We will share these insights in the frame of using data and ML for actuary and underwriting.
The usage of the expansive data sources that are available today holds potential and opportunity for ML adoption in life and health insurance. The models developed can be directed into revenue-generating products that drive value creation and monetary benefits for the insurer. Data sources include using wearables for premium prediction for underwriters. However, we should ensure we do not use data in silos but to integrate wearables data with related data points such as a policyholder’s health status, lifestyle, and possible open health data recorded by a country’s government as part of nationwide healthcare provision, among others.
Once data are organized, we can apply ML techniques to augment an insurer’s function such as enhancing actuarial price prediction capability. This provides an added dimension towards the classical actuarial methods and develop ML models that are fit for purpose and use. To bridge the difference in evaluation metrics between the data science function and the actuary, we can look at loss ratios and incidence rates instead of F1-score and precision-recall metrics. In terms of vendors, platforms which provide the ability to view data and evaluation metrics on a dynamic dashboard can provide explainability and user interface management of model deployment/re-training. Such platforms also allow for better compliance to regulation and to monitor data drift. The goal is to overcome the challenges of transforming the traditional actuary approaches such as introducing ML tools that have lower learning curve for actuaries to embrace and with common performance metrics evaluation.
To learn more, Celent tracks this market and has research addressing it (list of recent reports here).If you would like to find out more, please feel free to get in touch with me.
Below are related reports/blogs contributed by my colleagues and I:
Securing Insurance Data: Confidential Computing and Data Lineage Use Case
From Data to Product to Customer Engagement: Delivering Good Distribution and Customer Experience with Alternative Data
Data, MLOps, and IoT for the Next-Generation Insurance Industry
The Data Force: Cultivating a Data-ready Organization