280 likes | 512 Views
Business Analytics. Database Marketing & Statistical Modeling Douglas Cohen, Director of Business Analytics @ Beachmint. Online Consumer Market. Why do companies bother with database marketing? Margin Players Online Gaming, e-Commerce, Lead Generation Buy low, sell high
E N D
Business Analytics Database Marketing & Statistical Modeling Douglas Cohen, Director of Business Analytics @ Beachmint
Online Consumer Market • Why do companies bother with database marketing? • Margin Players • Online Gaming, e-Commerce, Lead Generation • Buy low, sell high • Cost To Acquire a Customer < Customer Lifetime Value • Big Budgets • Zyngaspent over $40 million in 2011 Q1 • Acquisition spend rising in many industries • Competitive landscape • Companies are competing for the same customers • Cost to Acquire a Customer is rising • Marketing Analytics • Companies working to understand their target audience • Which customers have highest Lifetime Value, LTV?
Customer Database • Data store used to record all customer information • Attributes • Name, Address, Demographics, Marketing Attribution • Transactions • Internal Sales, Content Delivery • Behavior • Click stream, Visits, Feature Usage • Drives personalized communication • Target customers for products / services • New home owner, recently married, birthday • Customer Lifecycle based promotion • Versus traditional business centric promotion • Importance of Data Warehousing • High attention to data driven discovery • Allows companies to understand their target audience
Your Average CustomerEach individual customer contributes to the understanding of the customer as a whole.
Database Marketing CycleDatabase is the center of the marketing cycle.
Data Mining • Marketing Campaign Assessment • Analysis shows whether campaigns were effective • Identify which customer segments responded well • Visualization Tools • Excel, Tableau, Pentaho, D3 • Statistical Models • Great when number of segments is large • R, Mahout, Weka, Orange
Statistical Model ExampleDecision tree used to find segments with high response rates.
Statistical Modeling • Model customer behavior using statistical techniques • Campaign Management & LTV Prediction • Campaign managers need accurate forecasts of LTV • Buy Till You Die Model • Customer Retention & Survival Analysis • Understand how to improve customer loyalty & reduce churn • Proportional Hazards Regression • Calculate variation in hazard rates among customer segments • General Profit Maximization • Product Recommendations • Increase probability of purchase versus size of purchase • Response Rate Modeling • Optimize response from customer communication efforts • Price Discrimination • Dynamically assign pricing based on customer income levels
Buy Till You Die ModelMost firms lump customers into segments & predict LTV per segment
Buy Till You Die Model • Increase accuracy by looking at customer level data • Transaction Process (“Buy”) • While active, the number of transactions made by a customer follows a Poisson Process with a transaction rate • Transactions rates are distributed gamma across the population • Dropout Process (“Die”) • Each customer has an unobserved lifetime length, which is distributed exponential with a dropout rate • Dropout rates are distributed gamma across a population • Approximates complexity in customer behavior • Simpler to implement than a psychographic model • Astonishingly good fit & predictive performance
Buy Till You Die Model • Poisson, Exponential & Gamma Distributions • Fit the appropriate curve to each customer segment • Coefficients have direct interpretation • Transaction, Dropout Rates are lambda • Gamma distribution describes heterogeneity • Store coefficients in data warehouse & feed into reports
Buy Till You Die Model • Implementation • Customers subscribing 2011, predict behavior in 2012 • Fit in calibration period was great. • Fit in holdout period was … horrible. • Why? • BeachMint made significant changes in discounting 2012. • Behavior did not transpose correctly for 2011 customers. • Solution: Segmentation • Customers starting with no discount should be less prone to change • Segment Customers by starting discount amount • Split into 2 similar sized groups • Start Discount = 0 % • Start Discount = 50 %
Buy Till You Die Model • Goodness of fit within calibration. • More repeat transactions from 0% Start Discount
Buy Till You Die Model • Goodness of fit within hold-out period. • Customers binned based on calibration period transactions.
Buy Till You Die Model • Actual vs. expected incremental purchasing behavior. • Monthly periodicity from subscription model.
Buy Till You Die Model • Actual vs. expected cumulative purchasing behavior. • Irregularities in the holiday period not captured.
Buy Till You Die Model • Transaction Rate Heterogeneity • Distribution of Customers’ Propensities to Purchase
Buy Till You Die Model • Dropout Rate Heterogeneity • Distribution of Customers’ Propensities to Drop Out
Buy Till You Die Model • Discounted Expected Residual Transactions • Given Behavior during Calibration Period
Buy Till You Die Model • Discounted Expected Residual Transactions • Higher Frequency & Recency has more impact for Discounters.
Proportional Hazards ModelExplain what factors contribute to survival over time. • Explain hazards of various conditions / customer variables • Commonly used in medical industry to compare risks of treatment groups • Hazard Ratios • Simple, easy to interpret • Relative risk ratios • Example 2X increase • Weibull versus Gamma distribution • Better curve fitting