• Shounak Mondal

The confidence to predict - Applying regression to campaigns for profitability lift

Updated: Dec 31, 2019

Which of our customers will respond to the campaign resulting in sales ? Among those, who should we target for maximizing profitability ? How much will the additional profitability be compared to random targeting ?



The context of this real world project :


A B2B office supplies retailer has held campaigns previously by random targeting its customer base with poor results. Now it wants to improve next campaigns by learning from the data of the previous campaign.


Here is an outline of how I would tackle to answer these questions but before that lets take a quick preview of the process and approach. The steps to do this are as follows :


1. Exploratory Data Analysis

2. Data Transformation and Analysis

3. Model building and fitting

4. Building the gains or lift chart

5. Final recommendations


How can I stop wasting marketing dollars and how can I not miss any revenue opportunities. Every marketing leader running a campaign would love to get a grip on these questions.

Which of our customers will respond to the campaign resulting in sales ?

This question needs a binary outcome i.e sales or no sales. Hence I use Logistic regression to predict the positive or negative outcome. However, the data was unbalanced, meaning there were far too many no sales cases compared to positive sales cases, hence Logistic regression does not train well and hence the scores of the model was poor. Hence I used Random Forest Classifier and could get much better scores and hence and more robust model.




Among those, who should we target for maximizing profitability ?

To maximize profitability we have to target the campaign to only those customers where the sales is high and makes the overall profitability positive keeping into account the cost of the campaign i.e transaction costs and fixed costs. Applying Linear regression we can predict the amount of sales for positive sales customer cases. Additionally we also get a deep insight to which features or factors decide on the amount of sales.




How much will the additional profitability be compared to random targeting ?

If we arrange the customers in descending order of profitability using the above formula, then we get a gains or lift chart. From this we know that we should target the first 4 deciles of customers to maximize profitability. We also now know by adding the profitability increase over the campaignprofitability, the financial gains possible.



All rights reserved. 

© 2020 by SHOUNAK MONDAL

  • White YouTube Icon
  • LinkedIn - White Circle