Moving from Value Segments to Predicted CLV for Retail: Unlocking the Power of Customer Lifetime Value
Harness Data-driven Insights to Optimize Retail Strategies and Enhance Customer Lifetime Value.
In the dynamic landscape of retail, businesses are competing aggressively for customers, and it is no longer enough to simply segment clients based on value.
In the ever-evolving realm of retail, where data reigns supreme, there emerges a paradigm shift towards harnessing the power of predictive CLV models. By delving deeper into customer value and refining marketing strategies, businesses can now anticipate the departure of patrons, affording them precious moments to intervene and provide bespoke encounters that bolster customer retention rates.
In this article, we will explore the transition from value segments to predicted CLV and how it empowers retailers to make informed decisions, drive customer-centric initiatives, and boost long-term profitability. In addition, we will present a case study on a CLV approach that we have taken in Beta Labs and touch on some of data science concept.
What are the limitations of Value Segments:
Unfortunately, value segments, hinged upon transactional archives and spending patterns, offer a rudimentary comprehension of customer value.
These segments, regrettably, turn a blind eye to the profound influence of customer retention, recurring purchases, and the art of cross-selling.
Our reliance on these segments obstructs our ability to discern the forthcoming high-value clientele and allocate resources judiciously to amplify the customer lifetime value.
Predicted CLV: A Holistic Approach
Predicted CLV, on the other hand, takes a more comprehensive approach by leveraging advanced data analytics and predictive modelling techniques. It amalgamates the transactional archives, customer demographics, behavioural intricacies, and external influences to cast an accurate forecast of a customer’s forthcoming value.
Delving deep into the realms of customer churn, average order worth, purchase frequency, and customer engagement, predicted CLV unveils an exquisite portrait, a reflection of a customer’s latent lifetime value.
How we use predicted CLV in Beta Labs:
a. Customer Segmentation and Targeting: By identifying high-value customers who are likely to generate significant revenue over time, our marketing and CRM team is able to allocate resources effectively and personalize communications to enhance customer engagement and loyalty.
b. Customer Retention and Loyalty Programs: Predicted CLV is helping us to identify customers at risk of churn and provide insights into how to retain them. By proactively engaging at-risk customers with targeted retention initiatives, such as exclusive offers or loyalty rewards, we are able to minimize churn, extend customer lifetime, and maximize CLV.
c. Resource Allocation and Return on Investment: Predicted CLV aiding us to optimize resource allocation by focusing marketing efforts on high-future value customers with the highest growth potential. This approach ensures that marketing budgets are utilized efficiently, resulting in a better return on investment (ROI) and increased profitability.
Case Study: Luxury Fashion Retailer
Before we go through the steps to calculate the predicted CLV, we need to understand the formula to calculate the predicted CLV
Predicted CLV in next 1 year(of Customer A)=
Predicted Average Order Value(AOV) per transaction(customer A) X Predicted No of transactions in next 1 year(Customer A)
Types of model used for prediction
a. We use XGBoost with Poisson objective function for predicted no of transactions in next 1 year
“Poisson regression is a type of generalized linear model (GLM) that is commonly used to model count data. It is specifically designed for situations where the response variable represents the number of occurrences within a fixed time period”
b. We use XGBoost with Tweedie objective function for predicted Average Order Value(AOV) per transaction as AOV data should theoretically follow gamma distribution
“The Tweedie distribution is parameterized by a power parameter(p) determining the shape of the distribution.
When p = 2, the Tweedie distribution reduces to the gamma distribution, appropriate for modelling continuous positive-valued data with a skewed distribution”
Fig 1.0: Distribution of Fitted Gamma compared to actual data
Key Features
In our model the key features with highest level of importance
A. Frequency of visits
1. No of historical visits
2. Visit interval.
3. Online vs Offline visits
4. Recent visit date
B. AOV
1.Historical Sales Amount
2. Loyalty tier
3.Recent sales amount
4. Online vs Offline sales amount
Model code
v1.train(data= training_data, response=”frequency_of_visits”, xgb_conf={“objective”: “count:poisson”, “colsample_bytree”: 0.3, “learning_rate”: 0.1 “min_child_weight”: 5,”max_depth”: 10, “alpha”: 5, “n_estimators”: 170, “seed”: 42})
v1.train(data=training_data, response=”AOV”, xgb_conf={“objective”:”reg:tweedie”, “n_estimators”: 150, “max_depth”: 6, “colsample_bytree”: 0.6, “alpha”: 0, “lambda”: 6})
Final output visualization
The below visualization shows the model performs with respect to actual data.
Fig 2.0: Actual vs Predicted output
As you can see the prediction is overpredicting those customers that have low CLV.
In the below section, we have suggested a few tips to solve this problem.
Tips on solving some issue
1. Remove outliers from your model.
2. Do not run the model on the whole base at the same time. Divide the base into loyalty tiers or value segments and run the model by tiers or segments. It will be more accurate.
3. If your company operates in different region of the world or even different region in the same country run the model separately for each region.
4. If your data is highly skewed by churn customers, it will impact on gamma regression. You can down-sample churn customers to make the overall distribution closer to gamma.
Using the predictive CLV model in business
Based on the following output , we are using the results of predicted CLV to drive the below mentioned campaigns for the CRM team.
Fig 3.0: Results of CLV Model
A. Reactivation Campaigns: We are using predicted CLV to send personalized loyalty points based on predicted CLV value depending on customer interaction with our reactivation campaigns, hereby reducing our cost.
Fig 4.0: Decision to provide benefit to reactivated customers based on predicted CLV
B. Cross-Selling and Upselling: We are using Predicted CLV to identify customers who are not performing well compared to their capability and hereby sending cross selling campaign to them.
Fig 5.0: Gap analysis for cross selling/upselling
Conclusion:
Moving from value segments to predicted CLV represents an evolution in customer-centric retail strategies. By leveraging the power of predictive analytics, we are able to gain a deeper understanding of customer value, optimize resource allocation, and create personalized experiences that drive long-term customer loyalty and profitability.
