Maximize Revenues on your Customer Loyalty Program using Predictive Analytics 27 th Feb 14 Free Webinar by
Before we begin... www Q & A?
Your Speakers @parikh_shachi Technical Analyst @tatvic Loves js and data analysis @kushan_s Web Analyst @tatvic Loves R, Pandas, ML
Agenda Background and Economics of Customer Loyalty Defining the Business Question A Primer on Predictive Analytics Defining the data sources Logistic Regression Model Accuracy Improving the Model
Customer Retention Why should you Care? Customer Acquisition Costs are on the rise Repeat Customers Create higher value (both in AOV & Revenue) Evangelize your brand Have Lower Service Costs Retailers can achieve tremendous revenue gains by shifting their marketing budgets to better target these customer segments Attributed from (http://www.practicalecommerce.com/articles/63459-seek-repeat-customers-to-drive- Ecommerce-Profits)
Real Life Example Sample Size: 5000 Consumers
Contribution to Revenue 750 (repeat) customers drive 40% of the total Revenue
Contribution to Revenue If 5% of these customers become repeat buyers after Discount Targeting, what are the implications for revenue?
Conventional Approach to Customer Loyalty Send Discount Coupons to all Customers either via email or some other medium Problems Non Targeted Campaign hence suffers from Low Conversion Rate Sending Discount Coupons to all customers erodes your sales margin
Revenue Leakage: What If Analysis Size of Email List 100,000 Click Through Rate of Email List 5% Visits 5000 Conversion Rate 2.5% Transactions 125 Average Order Value $250 Discount Provided 20% Discount $50
Revenue Leakage: What If Analysis Size of Email List 100,000 Click Through Rate of Email List 5% Visits 5000 Conversion Rate 2.5% Transactions 125 Average Order Value $250 Discount Provided 20% Discount $50 Persuadables (Customers Who bought after discount was provided) Sure Things (Customers who would have bought anyway) Loss in Revenue $2,500 75 50
Summing up Target your Loyalty Campaign to this segment Image Courtesy : Dr. Eric Siegel (http://www.predictiveanalyticsworld.com/lower-costs-with-predictive-analytics.php)
Business Question for Predictive Analytics Predicting Customers who would make a repeat purchase within 2 months of their initial purchase Outcome/Response Variable: Whether the customer would make a repeat purchase within 60 days Using Data of Past Customers who have made purchases on the site
Until Now Repeat Customers are valuable and we need more of them Sending out discount coupons to all customers w/out segmentation leads to a loss in your Revenue Use a Predictive Model to find out those customers who would not make a return purchase without a discount coupon Target your Discount Coupons only to these customers
Data Sources and Features Google Analytics Data CRM Data Transaction Date Product Category Item Quantity Shipping Cost Incurred Medium Customer ID Is Newsletter Subscriber? Discount Coupon Redeemed? Account Creation Date
An Aside: Extracting Google Analytics Data into R User performing data extraction Google OAuth2 Authorization Server Google Analytics API Access Token Request Image adapted from: Google Analytics Core Reporting API Dev Guide
An Aside: Extracting Google Analytics Data into R User performing data extraction Google OAuth2 Authorization Server Google Analytics API Access Token Request Access Token Response Image adapted from: Google Analytics Core Reporting API Dev Guide
An Aside: Extracting Google Analytics Data into R User performing data extraction Google OAuth2 Authorization Server Google Analytics API Access Token Request Access Token Response Call API for list of profiles Image adapted from: Google Analytics Core Reporting API Dev Guide
An Aside: Extracting Google Analytics Data into R User performing data extraction Google OAuth2 Authorization Server Google Analytics API Access Token Request Access Token Response Call API for list of profiles Call API for query Image adapted from: Google Analytics Core Reporting API Dev Guide
Intuition behind Supervised Learning Example courtesy : Trevor Hastie, Rob Tibschirani (Statistical Learning, StanfordOnline)
Supervised Learning Generates a function that maps inputs (labeled data) to desired outputs (e.g. Image Classification) Training Data Labels Variables Machine Learning Algorithm Supervised Learning Model Labels are right answers from historical data e.g. Image of Car/Bike Input Data: Contains Images of Bike and Car Image Courtesy: Olivier Grisel https://speakerdeck.com/ogrisel/machine-learning-in-python-with-scikit-learn
Supervised Learning Generates a function that maps inputs (labeled data) to desired outputs (e.g. Image Classification) Training Data Labels Variables Machine Learning Algorithm Supervised Learning Model Labels are right answers from historical data e.g. Image of Car/Bike Input Data: Contains Images of Bike and Car Test Data Variables Predictive Model Predicted Outcome labels Image Courtesy: Olivier Grisel https://speakerdeck.com/ogrisel/machine-learning-in-python-with-scikit-learn
Logistic Regression Model Algorithm used to predict categorical labels In our problem Categorical Labels are 0 : Did not carry out repeat purchase 1 : Carried out Repeat Purchase within 60 days Using the algorithm we predict the probability of a Customer ID belonging to either class
Checking Model Accuracy Split Data Randomly into Train and Test 80% Train Data 20 % Test Data Fit glm model on Train Data Predict labels for unseen Test Data
Model Accuracy Predicted Labels (Predicted by running Model on Test Set) Confusion Matrix Actual Labels (From Test Set) Not a Repeat Purchaser Repeat Purchaser Not a Repeat Purchaser 5271 4 Repeat Purchaser 1209 1 Labels 0 : Customer didn t make a repeat purchase in 60 days 1 : Customer made a repeat purchase in 60 days.
Model Accuracy Predicted Labels (Predicted by running Model on Test Set) Confusion Matrix Actual Labels (From Test Set) Not a Repeat Purchaser Repeat Purchaser Not a Repeat Purchaser 5271 4 Repeat Purchaser 1209 1 Accuracy = (Number of Correctly Predicted Labels) / Total Number of Labels = (5271 + 1) / (5271 + 4 + 1209 + 1) ~ 81.3 %
Improving Model Accuracy Adding New Features to the model Difference b/w Account Creation Date and Transaction Date Checking for Transactions occurring during Weekend (based on Date) Adding Days To Transaction, Location, Device Type as Features from Google Analytics Trying out additional models Random Forests Gradient Boosting Support Vector Machines
Q&A Round
Next Webinar How to Perform Churn Analysis for your Mobile Application March 19 th 11:00 AM PDT Key Takeaways Predict the Segment of Mobile App Users who would uninstall your app Remain Inactive and Churn over a period of Time Register Now: www.tatvic.com/webinar
Thank you! Kushan Shah kushan@tatvic.com +1 276-644-0456 Drop us a line on Twitter @tatvic