This course introduces programming techniques to craft and feature engineer meaningful inputs to improve predictive modeling performance. In addition, this course provides strategies to preemptively spot and avoid common pitfalls that compromise the integrity of the data being used to build a predictive model. This course relies heavily on SAS programming techniques to accomplish the desired objectives.

The self-study e-learning includes:

• Annotatable course notes in PDF format.
• Virtual Lab time to practice.

• Extract data from a relational data table structure.
• Define population qualifications and create a target sample.
• Use feature engineering techniques to transform transactional data into meaningful inputs into a predictive model.
• Transform low-, mid-, and high-cardinality categorical input variables into meaningful predictive modeling inputs.
• Use ZIP codes and latitude/longitude points to calculate great-circle distance, driving distance, and estimated driving time.
• Use Bayes' theorem to estimate meaningful predictive modeling inputs, impute missing observations, and partition the target sample into training and validation data sets for honest assessment of the predictive model.

Who should attend

Formats available	Standard Duration (duration can vary, see event schedule for details)
Live Web Classroom:	6 half-day session(s)
e-Learning:	21 hours/180 day license

This course assumes some experience in both predictive modeling and SAS programming. Before attending this course, you should have:

• Exposure to DATA step programming equivalent to the SAS Programming 1: Essentials course.
• Exposure to programming in SQL or the SQL procedure.
• Exposure to querying data in PROC SQL and building and deploying a predictive model.
• Familiarity with the analytical process of building predictive models and scoring new data.

Familiarity with the SAS macro language is helpful but not required.

This course addresses Base SAS, SAS/STAT software.

• Data difficulties.
• Assessing available data.
• Accessing available data.
• Drawing a representative target sample.
• Drawing an uncontaminated input sample.

• Advantages and disadvantages of transactions data.
• Common transaction structures.
• Defining the time horizon.
• Fixed and variable time horizon methods.
• Implementing common transaction transformations.

• Definitions and difficulties of nonnumeric data.
• Miscoding and multicoding detection.
• Controlling degrees of freedom.
• Geocoding.

• Exploring input variable distributions.
• Detecting data anomalies.
• Creating custom exploratory tools for candidate input variables.
• Missing value imputation.
• Data partitioning.