Supporting the Office of the Chief Financial Officer (OCFO)

Modeler for the Data Cooperative project at GuidewireCompleted redesign and implementation of a risk model for cyber catastrophe events for the Cyence product. This model aligns risk of attack to a combination of historical experience, cyber expert judgment, and customer feedback. Improved the computational efficiency of this model from a runtime of 12 hours for 10k simulations to 40 minutes using coarse-grained parallelism and vectorized numpy calculations in AWS EMR.Lead modeler and prototype developer for Merger and Acquisition cyber-risk model intended for a leading financial services client. This model is used to quantify the cyber risk of merging two organizations. Adjustments in tail risk and average annual loss are based on the combined risk profile of the two companies. Provided direction to contractors on the report and chart design.Designed and developed an original algorithm to quantify supply risk in a Bayesian network. Algorithm is implemented in python with a polynomial time complexity. This algorithm replaced the initial proposal with an exponential time complexity, which was too slow for a production application. Developed the original set of inputs or priors for this model based on cyber risk data, industry studies, original research, and associations of sources of revenue between companies (correlation matrix).Led second-place team in Guidewire hackathon. Developed NLP model to mine cyber attack data for new ransomware gangs and categorizing types of attacks.

Supporting the newly formed Wealth Management Insights and Analytics group for National Accounts.

Developed machine learning and statistical models to re-design North American Distributionsales territories. These models optimized territory coverage for over 200 sales professionalsacross the country.Leads a team of 4 consultants to build analytics for three new departments of the NorthAmerican Client Group. These departments work with the home offices of financial servicesfirms for product placement and sales initiatives.Developing a model to estimate the sales lift for a major coverage redesign initiative. Thismodel is based on historical and projected marketshare, based largely on activity data reportedthrough salesforce.com.Extensive experience with using Salesforce data to drive business decisions related to salesanalytics. Worked with senior business leadership to build the foundations for reporting andanalytics.Fit a Generalized Linear Mixed Effects model (GLMM) to Salesforce.com data to identify themost significant drivers of work for Internal Wholesalers.Developed the forecasting methodology and implementation for the American Funds Ser-vice Company budget process. This process has been adapted and reproduced for multipledepartments in the organization.Developed two statistical computation packages in R, available on CRAN:- hmclearn: https://cran.r-project.org/web/packages/hmclearn/index.html- bayesGAM: https://cran.r-project.org/web/packages/bayesGAM/index.htmlPUBLICATIONSThomas, S. and Tu, W. (2020). Hamiltonian Monte Carlo. In Wiley StatsRef: Statistics Reference Online (eds N.Balakrishnan, T. Colton, B. Everitt, W. Piegorsch, F. Ruggeri and J.L. Teugels). doi:10.1002/9781118445112. stat08243Thomas, S., & Tu, W. (2020). Learning Hamiltonian Monte Carlo in R. arXiv preprint arXiv:2006.16194.In ProgressThomas, S. and Tu, W. 2020. An R package for Bayesian Multivariate Generalized Additive Models.

- Developed an R package called mlts (multi-level time series) to automatically develop forecasts and perform cross-validation for bottoms-up forecast models. This package is in current use to forecast more than 400 individual forecasts each year. - Developed a variety of long-range and short-range forecasting models for business and capacity. These dynamic regression models (ARIMA-based) have been accurate within 1% to 5% per year. - Developed a churn predictive model in R for retirement plan de-conversions using the C5 algorithm. The model is currently in use to identify plans at risk for de-conversion as part of a plan retention effort. Back-testing has demonstrated that plans with high risk scores are twice as likely to de-convert compared to other plans- Estimated the long-range Fund Account / Unique Registration ratio using queuing theory (i.e. Poisson Processes)- Developed control charts to automatically identify dealer behavior changes, likely as a result of the Department of Labor fiduciary rule. - Has provided internal statistical consulting for Business Management, Work Management, Legal/Compliance, Human Resources, Quality, Retirement Plan Services, and AFS. - Developed a Python program to estimate the number of Shareholder Services associates needed if processing work was outsourced. This linear program leveraged 25 different associate schedule types and leveraged mathematical assumptions on call handle time distributions. - Developed a Cox Regression predictive model in R for associate attrition based on internal and external variables (e.g. unemployment rate). Regression model has been accurate within 3%

Forecasting and capacity planning; statistical modeling

Developed a regression model to match the automated pricing results from a website. The model was developed using statistical software and translated to Excel for the client's use. Developed a predictive model based on machine learning algorithms for an environmental sensor application with over 400 variables. Created an automated forecasting model to predict fuel demand in various locations in the UK. This model was used to anticipate geographical fueling needs for a trucking company.

Taught Calculus, Finite Math, and Algebra

Used text mining to analyze open-responses to survey questions from a call center. This analysis was used to identify drivers of caller satisfaction. Evaluated search paths commonly used by CSRs to find scripts using sequencing analysis in R. This analysis was used to identify opportunities to improve script searching capabilities.Identified opportunities for quality and cost savings for the 1-800-MEDICARE call center through data mining and statistical methods.

Forecasting and regression modeling; health reserving

Health insurance reserving calculations