Description: Basic feature engineering uses input data to create dummy variables, log transforms, and other single-value transformations. This talk builds upon these univariate transformations and presents several strategies for creating complex multivariate features that can enhance your model's lift. Topics such as target encoding, dimension reduction/projection, row/column, and other feature engineering topics will be covered.

Description: This article discusses data mining and the statistics of coincidence, and provides “guidelines for backtesting integrity” to help ensure results have value beyond mere coincidence. Editor’s Note: The following article originally appeared in the Winter 1998 issue of the Horizon, a publication of BARRA, Inc., and is reprinted with permission.

Description: Comparison of healthcare costs for early retirees versus active workers in the same age band. When adjusted for age, there is not a significant difference in costs between each group.

Description: Neural networks have fallen only to rise again twice. The latest incarnation borrows much from the past but adds elements to support deep learning. This article demonstrates the state of the art in deep neural network programming.

Description: This dissertation by Tia Sawhney will help you understand this issue and start measuring it, or start promoting equitable public policy.

Description: Data requirements for the Society of Actuaries cancer experience study, for those submitting exposure and claim information.

Description: How do you make model output media worthy? This article explores considerations that may contribute to the mass production of model results for public consumption.

Description: A general introduction about predictive modeling in insurance applications. Some basic models are discussed, such as Generalized Linear Model(GLM), Calssification and Regression Tree Model. Other models are also briefly introduced, including data clustering, neural networks and support vector machine.

Description: The newer predictive modeling tools are powerful but they work best when supported by a lot of computing power. Parallel processing is very powerful but often very difficult to implement. Dihui and Richard take us through the maze of options with H2O, SparkNet/MLib and Mahout. Which do you use for a generalized linear model (GLM), and which is better for high-volume data manipulation?

Description: Description of the vast variety of predictive analytics topics covered by the Forecasting & Futurism section and introduction to the July 2015 issue of the newsletter.