预测分析的新前沿

Over 97% of businesses worldwide have invested in big data. However, only24% of these companies claimed they use the collected data to analyze andmake informed decisions.1Data management is an integral part of running a Today the insurance industry is experiencing a fundamental shift in how todefine, understand, and quantify risk. Recent technological advancementshave led to an explosion of data, which demands new processing and analysistechniques beyond traditional methods to make sense of it all. Consequently, Unconstrained models — those with few limitations — maximize data utilityand predictive power by leveraging advanced algorithms. These modelsallow for flexibility, have the ability to capture complex relationships, When used strategically, unconstrained models can analyze and enhancetraditional models to unlock new insights, even in highly constrained orregulated environments like insurance. For organizations that embrace them, In this paper, we examine the latest advancements in the insuranceanalytics landscape. We discuss how unconstrained models can strategically THE INSURANCE ANALYTICSLANDSCAPE UNLOCK NEW DATA SOURCES FOR MORE ACCURATE RISKMODELING The past decade has seen a data revolution characterized by the emergence of new datatypes as well as increased volume and velocity. Today, smartphones continuously transmittelemetry data for various applications, vehicles provide diagnostics and receive over-the-air updates, and smart refrigerators are poised to display advertisements. As a result, thevolume and growth of data have skyrocketed. Data created, captured, copied, and consumed The insurance industry has similarly transformed by leveraging this new data. Datasetsnow include telematics for scoring driving behaviors, Internet of Things (IoT) data for real-time leak detection, satellite imagery, and climate data for a more refined underwriting UNDERSTAND LIMITATIONS OF TRADITIONAL MODELS AND ALGORITHMSINTODAY’S DATALANDSCAPE The industry standard for data analysis, modeling, and product development centers onthe versatile generalized linear model (GLM) and its variants. Formally introduced by JohnNelder and Robert Wedderburn in 1972,2these models were built on years of mathematical While GLMs are highly useful analytical tools, their development predates modern datasets,which now contain millions of rows and hundreds of columns. Though modern GLMs handlelarger data, they still require additional algorithms to find the most predictive inputs. Other •Difficulty capturing nonlinearrelationships•Interactions must be explicitlyspecified•The required choice of distribution and link function introduces model specificationrisk Data scientists use various techniques to tackle these challenges, including gridsearch, regularization, and quasi-likelihood models. They also use advanced methods WORK WITHIN MODELING CONSTRAINTS DRIVEN BY BUSINESS ANDREGULATORYNEEDS Constraints are conditions that restrict the scope of a model’s inputs and outputs to explaina given phenomenon or process. By design, constraints hinder what we seek to understand;however, due to their natural intuitive limits or boundaries, they will always exist as we strive Constraints are often driven by historical business needs — the information required tooperate effectively. Data management is an integral part of running a business, for year-end reporting and tax purposes, and to comply with laws and regulations. Traditionally, the Model constraints can be grouped into several broad categories. Exhibit3 highlights the The constrained use of data arises from the specific use cases for which analyticalframeworks, or models, were designed. These constraints are not necessarily negative;in fact, they often enhance the model’s usability. For example, a complex model aimed at UNCONSTRAINED MODELS CAN HELPORGANIZATIONS OVERCOME TRADITIONALLIMITATIONS AND MAXIMIZE PREDICTIVEPOWER Unconstrained models place no limits on data, models, or algorithms. They are designed tomaximize predictive power and accuracy by identifying the most relevant data and selectingthe optimal analytical framework. It’s important to view these models as strategic tools that Advanced algorithms such as gradient boosted machines (GBMs) and neural networks(NNs) play a significant role in data analysis. GBMs operate by creating a series of trees that On the other hand, neural networks, inspired by the human brain, are composed of flexible,layered neurons that learn intricate relationships. In fact, NNs form the foundation of thesuccessful large language models (LLMs), like OpenAI’s ChatGPT, Google’s Gemini, and Traditional models typically require highly organized and standardized data, often referredto as structured data. In contrast, advanced algorithms can work with both structured andunstructured data, such as text notes, images, videos, and audio files. This ability allows Case Study PERFORM GAP ANALYSIS TO DETERMINE A TRADITIONAL MODEL’SPREDICTI