Predictive modeling predicts events or quantities (outcomes) using any of a wide variety of machine learning algorithms. It affects every sector of the economy and our personal lives. It forecasts weather, assists with medical diagnoses, approves loans, identifies fraud, predicts movements of enemy combatants, and estimates outcomes depending on the decision made.
Models work when the information on which they are trained applies well to current and future scenarios, as their decisions are shaped by the wisdom embedded in the historical data. But if the training data don’t reflect current situations well, or the modeling is not of high quality, expectations will outstrip actual performance. We have all experienced bad navigation advice from a GPS, inaccurate COVID-19 projections, or financial projections that were quite far off. We believe is vital to master the skills of data science to harness its power and avoid its pitfalls to arm organizations to successfully extract insight from their data and put into production predictive models that deliver exceptional value.
Primary Techniques
Predictive modeling automatically learns a relationship (model) between a set of inputs and an output according to a training data set. The model (an equation or set of rules) is inferred from training data cases where the output (outcome) is known and will be used for new cases where the outcome is unknown.
A high-quality machine learning model is essential, but is only one of several factors vital to project success. Problem formulation, data preparation, feature selection, feature engineering, and testing design – which all require intensive analyst intervention – must also be done well for models to become useful and trustworthy.
The utility of a model’s decisions must be compared to that of the current “baseline” process. The expected errors in its estimates must be well understood and communicated. A robust modeling process can uncover previously unknown relationships that add to the intellectual property of the organization. Trust in a model comes from its out-of-sample accuracy and also from its transparency. It is important to understand how the inputs interact to provide the predictions. Modeling ideally provides important descriptive insights that can confirm or deny theories about the drivers of the outcome, and lead to new ideas.
General Applications
A predictive model may identify or diagnose a condition or state, such as early-onset Alzheimer’s, a mechanical condition that warrants maintenance, or a transaction requiring investigation for fraud. Or, a prediction may be used to plan for a future condition, such as tomorrow’s weather, a future stock price, the likelihood of a loan going into default, or the anticipated hospital caseload for COVID-19 cases. Sophisticated predictive models can identify the expected effects of alternative actions, such as incentives to reduce the likelihood of customer churn, agricultural actions to maximize organic crop yields, or what will best encourage patients to pay their medical bills. We have experience developing solutions for a wide variety of such problems, which involves building new models or improving existing models.
Case Studies
- Fraud, Waste, and Abuse at USPS and DentaQuest (Dental insurance)
- Pharmacy fraud
- Subrogation likelihood on insurance claims (MunichRe)
- Impact of public health policy on COVID-19 cases
- Gas well failure in next four to six months (BP)
- Dental provider performance vis a vis outcomes (Delta Dental)
- Determine actions likely to increase patient payment of their bills (ACS Canton)
- Predict wireless customers most likely to drop their phone contract (churn) (NTelos)