Presentation Abstracts
Registration and Summit Schedule
| Speaker | Presentation Details |
|---|---|
| Ron Richman | Title: The Actuary and IBNR Techniques: A Machine Learning Approach Abstract:Actuarial reserving techniques have evolved from the application of algorithms, like the chain-ladder method, to stochastic models of claims development, and, more recently, have been enhanced by the application of machine learning techniques. Despite this proliferation of theory and techniques, there is relatively little guidance on which reserving techniques should be applied and when. In this talk, we present recent work on reframing traditional reserving techniques within the framework of supervised learning, with the goal of selecting optimal reserving models. We show that the use of optimal techniques can lead to more accurate reserves and investigate the circumstances under which different scoring metrics should be used. |
| Stephen Mildenhall | Title: The Promise, Pitfalls and Process of AI Abstract: The insurance industry is primed to realize the promise of AI. But, looking beyond the hype, it is important to understand how and why AI works in order to differentiate promising use-cases from potential pitfalls. This session will begin by placing AI within the broader context of big data, predictive analytics, machine learning, deep learning, data science, and even humble statistics. Understanding context helps identify common features of problems and data where AI has been successful and, in-turn, highlights new business problems that could be solved by AI. It will end by framing some issues around ethical model use and model bias. |
| Mark Shoun | Title: Bayesian Model Stacking Applications in Loss Development Abstract: The actuarial literature is full of models for loss development. The problem of model choice in loss development is typically framed as the task of selecting one of these options as the optimal model for predicting ultimate losses in a given situation. In reality, it is near-impossible to say a priori that one model dominates all others for a particular application. In this talk, we consider stacking of Bayesian models as an alternative strategy to the problem of model selection in the context of loss development. Stacking is a theoretically sound way of blending predictions from multiple independent models based on their relative goodness of fit. We apply stacking to a panel of Bayesian loss development models in a corpus of triangles from statutory filings and compare the performance of the stacked model to the constituent models. |
| Steve Guo | Title: Patterns and Anomalies of Loss Development in P&C Insurance Market Abstract: Loss development, in terms of incremental loss ratio (ILR) and cumulative loss ratio (CLR) in P&C insurers' business lines, can be viewed as functional curves across a ten-year development period. Regulators, reinsurers and other parties may wish to learn from a large number of triangles to find out patterns and anomalies of loss development in the market. Relying on principal component analysis (PCA), we study the ILR and CLR of workers' compensation line across hundreds of companies over years and identify patterns and detect outlying loss development curves. Our analysis shows that companies with different business focus and regional focus have distinctive development patterns. We also propose an approach of PCA with penalized least square to forecast the incomplete loss development curves. |
| Garth Johnson | Title: Pricing Data Science at UberEats
Abstract: A presentation on the application of data science in marketplace pricing at UberEats, in particular for optimizing user promotions to generate short term business growth. Algorithms to target the most elastic users and frameworks to run production experimentation are outlined. Observations and recommendations for entering tech data science are also discussed. |
| Kevin Kuo | Title: Unblocking Insurance Analytics R&D with Open Data Abstract: The lack of data availability is commonly cited as a hindrance for conducting insurance analytics research or prototype development. In this talk, we discuss recent open source efforts for improving the discoverability of publicly available data and synthesizing realistic data. |
| Guojun Gan | Title: Self-Paced Probabilistic Principal Component Analysis for Data with Outliers Abstract: Rapid advancement of technology has led to ubiquitous high-dimensional data in almost all fields, including insurance. Telematics data in care insurance is an example of such data. High dimensionality poses great challenges for data analytics in terms of methodology and computation. Dimension reduction techniques become an indispensable tool for handling high-dimensional data. In this talk, I will give an overview of dimension reduction techniques and present a particular dimension reduction method called Self-Paced Probabilistic Principal Component Analysis (SP-PPCA), which incorporates the Self-Paced Learning mechanism into Probabilistic Principal Component Analysis (PPCA) to improve the robustness of the traditional PCA. I will also present some numerical results based on synthetic and real data. |