thumbnail

Who responds to marketing campaigns?

The overall response rate for a marketing campaign by a Portuguese bank was 11.27%. In contrast, the 198 campaigns (among 40K+) to a person who purchased a product the last time on a date when the Euribor rate (the Euro Interbank Offered Rate) is less than 0.7 had 80.30% response rate. In fact, there are hundreds more that are similar to the group. How could we detect ALL these good groups with abnormally high response rates?

The overall response rate for a marketing campaign by a Portuguese bank was 11.27%. In contrast, the 198 campaigns (among 40K+) to a person who purchased a product the last time on a date when the Euribor rate (the Euro Interbank Offered Rate) is less than 0.7 had 80.30% response rate. This is not the only group; the 1,270 campaigns with contacts via cellular phones to customers who purchased products in last campaigns had 65.2% response rate. In fact, there are hundreds of groups with substantial sample sizes and abnormally high response rates. How could we detect ALL these "anomaly" groups?

Data

The data is from direct marketing campaigns by a Portuguese bank between May 2008 and November 2010. It is publicly available data at the UCI Machine Learning Repository. The data contains 41,188 campaign activities with 20 input variables, such as age, job, marital status, education level, previous campaign outcome, consumer price index and etc.

(We are leaving the data source's appropriate citation here because it was requested: [Moro et al., 2014] S. Moro, P. Cortez and P. Rita. A Data-Driven Approach to Predict the Success of Bank Telemarketing. Decision Support Systems, Elsevier, 62:22-31, June 2014)

Example solutions

Here are some of segments with substantial volume and high response rate.

  • Group 1

    • Response rate: 80.3%
    • Group size: 198
    • Characteristics

      • The 3-month Euribor rate is less than 0.7
      • Campaigns to customers with successful previous outcome
  • Group 2

    • Response rate: 65.2%
    • Group size: 1,270
    • Characteristics

      • Contact method is cellular
      • Campaigns to customers with successful previous outcome
  • Group 3

    • Response rate: 51.39%
    • Group size: 1,076
    • Characteristics

      • Consumer price index was in mid-range (between 94.0 and 94.26)

There are hundreds of other groups with enough records (more than 50) and response rates higher than 50%.

How did we find all these?

Extremely simply put, we identify ALL anomaly (in a good sense) segments by following the below two steps:

  • Step 1. Generate all combinations of values

    • For example, marital status is married is one group; marital status is married and job is admin is another group
  • Step 2. For each combination, decide whether the combination has enough records and these records contain a high percent of anomaly cases.

Though the method sounds easy, the number of combinations can increase exponentially. For example, the bank marketing data's two variables job (12 unique values) and education (8 unique values) generates 116 combinations (12 + 8 + 12 x 8 = 116). If we want to consider another variable month as well, then we have 1,520 combinations (12 + 8 + 12 + 12 x 8 + 12 x 12 + 8 x 12 + 12 x 8 x 12 = 1,520). See how exponentially the number of combinations increases!

Furthermore, if data is large (i.e. BigData), calculating the number of records and event rate for each combination themselves can be easily daunting tasks.

To solve the issues, we use a combination of BigData and Cloud technologies to distribute and process workload in many powerful computers. Moreover, we borrow existing ideas from the computer science field (such as caching) to run the computations much more efficiently.

Can we apply the algorithm to our data set as well?

If outliers (e.g. fraudulent transactions, high-revenue products and customers, marketing campaign responses) mean significant opportunities or risk to your business, our solution will work fabulously. Find how it is applied to find high-risk segments in credit data: Find exceptionally bad performance groups in your loan portfolio. If you are interested, please contact us by clicking the button below.

Tags

Comments