Corporate
Analysis & Inference assists corporate clients with a wide range of issues, including risk assessment, customer response modeling, and forecasting. Our services have been sought by corporate marketing, legal, information technology, regulatory compliance, finance, and accounting. Among the corporations we have worked with are:
- Blue Cross Blue Shield
- Digital Equipment Corporation
- Kellogg Brown & Root
- New York City Health and Hospitals Corporation
- Packer Engineering
- Pfizer
- Public Broadcasting Corporation
- State Farm
- Time Warner
Case Histories
- Accident prevention through statistical modeling
- Accuracy of restatement of earnings
- Advertising strategy modeling
- Brittle fracture probability for engineering application
- Classification of insurance risk using "shrinkage" estimation
- Client response modeling in direct mail
- Discount rate in valuation of General Dynamics shipyard
- Evaluation of data biases and product marketing for a pharmaceutical company
- Finding "changepoints" in stock prices
- Nuclear power plant risk assessed using insurance premiums
- Predicting claim payouts for grocery chain using a statistical model
- Probability of a catastrophic failure of an undersea pipeline
- Risk analysis in rear axle safety in Saturn models
- Sampling medical insurance claims files for Blue Cross Blue Shield for an audit challenge
- Statistical machine learning algorithm classifies customer accounts for revenue gain
- Statistical modeling and analysis of Public Broadcast Corporation share of TV royalties
- Survival and reliability analysis for medical device implant
- Tracking insurance costs for global construction company
Case: Probability of a catastrophic failure of an undersea pipeline
In a litigation matter involving a global engineering and construction company and one of the world's largest energy companies, Analysis & Inference was asked to determine the probability of bolts failing along an undersea pipeline, causing leakage of oil or gas.
Using several years of data detailing inspections and breaks found for thousands of bolts along hundreds of joints along the pipeline, Analysis & Inference constructed sophisticated "maximum likelihood" models to determine the probability of failure over the lifetime of the field.
Analysis & Inference was asked to write an expert report setting forth its findings. Dr. William Fairley testified on these findings at the international arbitration proceeding related to the litigation, which is still pending.
Case: Statistical machine learning algorithm classifies customer accounts for revenue gain
A major service company wanted to classify its accounts by value of revenue generation in a certain business situation. Only a few pieces of information were available on each account.
Analysis & Inference was asked if statistical methods could be used to classify the accounts. We applied state-of-the-art methods from statistical and machine learning theory and data mining to the information available.
Assignments were made with substantial accuracy that was demonstrated to be unobtainable by chance, and were validated on data not used to construct the model. The technique resulted in considerable new revenue.
Case: Client response modeling in direct mail
How to lower test marketing costs for a direct market publisher. Client used client response modeling where each title was marketed. Of the more than 55 million pieces of direct mail offers over a two-year year period; 10 million were test marketing pieces. The company rolled out only those titles with higher response rates in the test marketing, so about half the potential titles were never published.
An initial testing revealed that the costs of using outside lists in the test marketing phase far exceeded the benefits, and recommended using outside lists only in the "rollout" phase of direct mailings. In the second phase, we combined data from different models in order to improve overall response rate.
The publisher had performed statistical modeling list by list, without correlating the different models. The models we proposed first chose which books to market, then chose the clients, and finally chose the proper advertising vehicle. These selections were done on an individual client basis based on client profiles and the results of logistic regression modeling.
By combining models, while still retaining the list-specific information, we helped the publisher improve the response rate while lowering test marketing costs.