Agenda

The Financial industry has been one of the most successful to adopt machine learning. FICO's fraud detection systems played a pivotal role in establishing a long lasting victory for supervised neural networks 25 years ago in the credit card fraud detection space. Since then, we have continually improved behavior and unsupervised machine learning techniques, including methods from collaborative filtering and density-based models. Also important is explainable AI, which has been a part of fraud models for the last 20 years. We will discuss new innovations which are having dramatic impact on the financial systems for detecting and preventing fraud, abuse and money laundering.

This case study details how reconstruction costs following major weather events can be rapidly estimated by predicting damage to infrastructure at a granular geographic level. There is an increasing severity of major weather events impacting communities and infrastructure coupled with rising reconstruction costs. In Australia, funding models for disaster recovery relief are changing. To help get communities back on their feet, reconstruction funding is required sooner following such events, meaning public agencies need to find efficiencies in the existing machinery of government processes for disaster recovery. Implications for financial institutions and insurers will also be covered.

Breaking technical requirements into small enough increments that allow data scientists to deliver something of value to stakeholders every few weeks is a difficult challenge.  In this session, we cover how Northwestern Mutual used an Agile Analytics approach to develop a system of predictive models that helps automate risk assessments in life insurance underwriting.  We briefly introduce aspects of Scrum and Kanban along with an overview of the Standard Methodology for Analytical Models.  The session concludes with an illustration on how to assess the incremental information gain from improved models within a framework of data monetization.

One of the key needs of reinsurance providers who provide insurance to commercial auto businesses is on how reinsurance premia can be aligned with the risk profile of the loan portfolios. This session will focus on how Maiden Re developed an integrated insights portal that can profile customer risk by analyzing companies they underwrite, monitor loss ratios, and identify the best prospects to pursue, using predictive analytics techniques.

In this digital age of banking, a large part of financial institutions' time and resources is spent in fraud analysis and credit monitoring, yet knowing and engaging the customer based on their preference and ways of interacting is the new normal. It's all about knowing how to help the customer improve their standing, and even predicting when they may have challenges or knowing which product they could benefit from in the future. Predictive analytics are being more relied upon to help agents and customer service representatives service and engage the customer more effectively. Successful machine learning-driven prediction and recommendations, coupled with ML-driven bots can inherently differentiate your products and services to your customer and create a lasting relationship. This session will discuss the pieces and parts behind using ML and bots to change the customer engagement model of your financial institution.

Many Paychex clients are new businesses, with little or no established credit record. The Paychex payroll business is essentially extending short credit to our clients from the time they submit a payroll, until the time the funds complete transfer. This is a case study on dealing with the issues around Credit Risk Modeling for these new businesses. 

Injured workers being away from work constitutes one of the greatest claim cost drivers. In this case study, a Return-to-Work model is introduced to predict the likelihood that injured workers will stay away from work for longer durations. This model serves to target early interventions, such as proper treatments and treatments that assist in returning to work. Such measures reduce claim costs. We explore millions of claims and hundreds of millions of payment records across 2000+ of our clients, including myriad variables. Cutting-edge statistical and machine learning algorithms (GLM, Decision Tree, Random Forest, Gradient Boosting, Neural Network, etc.) are utilized, in combination with business knowledge.

In order to understand customer behaviors and provide better services, wealth management firms are investing greatly in data analytics. During this session, Meina Zhou will cover several examples of financial institutions that gained value with in-house predictive analytics solutions to improve wealth management services. She will also address multiple use cases built for different stages of the customer journey, including customer acquisition, customer personalization, and customer retention. She will discuss both the core analytics components and the challenges involved throughout the implementation process.

The sooner that incoming insurance claims can be identified as high risk of becoming  a long latency claim, the sooner they can be prioritized and targeted with specialized resources and services for better outcomes. And improving worker outcomes for our most challenging claims is where we can have the biggest impact on the overall success of the insurance system.  The Workplace Safety and Insurance Board (Canada) has used only data available at the time of claim registration to segment insurance claims by risk, as they come in. Our custom model combines random forest and conventional logistic regression techniques and is continually refined over time through machine learning algorithms.

We developed a Payment Defect Model for the John Hancock Long Term Care Insurance – Claims department. Long term care is different from most other insurance products, as it has much longer relationship with the clients, and the payout is recurring for several years. Since the reimbursement amount varies from payment to payment, detecting when over- or under-payments occur is especially challenging. Prior to this project, the Claims Audit department reached high detection levels only by monitoring nearly 50% of all payments. In comparison to a rule-based decision model, our machine learning model increased the detection rate, yet decreased manual effort by more than 60%.

In this session, we describe this project's full journey, from the problem statement to the data gathering, from the initial model and testing to a complete shut down and restart of the entire project. We break decades-old perceptions and solidified the necessity of involving subject matter experts in the model-building process.