How To Use Machine Learning In Predictive CLV Models (2 of 2)

Steps in building an ML model for predicting CLV

The following steps are involved in the building of a machine learning model to predict customer lifetime value (CLV):
1. Clean And Prepare Data:

• The first step is to prepare the data set and select the variables to be used as features for the training of the model. Before performing any processing or analysis on the data, some basic data cleaning operations need to be performed on the data set. These include:

• not using columns with low variance
• filling or eliminating rows containing null values
• getting rid of duplicate records to avoid redundancy

• For the purposes of this model, we are using a sample data set consisting of standard transactional data of a company, where each row corresponds to a purchase made by a customer. The data set consists of fields like the unique customer ID, date of transaction, quantity of products and total price of the purchase.

• Here is a sample of the initial raw data set we will be starting our CLTV model building with:

• These are raw features for a basic model, but one can incorporate additional features like browsing data that may include- browsing time, addition to carts, product searches. Also, email response data from marketing campaigns or promotions can be added to the set of features. Some metadata about the customer can sometimes be useful for predicting the lifetime value as well.
• We split the data into training and test sets using a threshold date (here taken as 2019-08-09). A training period of 1681 days is considered between 2015-01-01 and 2019-08-09. The test period will be the next one year till 2020-08-09, which is the time period for which we will predict the lifetime value of a customer.
• We just have 2 variables right not corresponding to each transaction, which won’t be enough for predicting customer lifetime value. So, we apply some feature engineering on the raw data to evaluate some insightful features for this purpose.

• First, we group the rows on customer ID and transaction day and aggregate using sum() to evaluate the transactions made by each customer daily.
• Now, for each customer ID we calculate the following values:

• Recency: The active duration during which a customer made purchases in that time period
• Monetary_value: Total amount of all purchases made by the customer in the given time period
• Frequency: Number of purchases made by a customer
• T: The number of days between the first transaction date of a customer and the last day of the training period (in this case 2019-08-09).
• time_between: The average number of days between successive transactions of a customer
• avg_basket_value: The average cost of purchases made by the customer
• avg_basket_size: The average quantity of products purchased
• 
• Target_monetary: This is the target variable which is calculated as the monetary value of the customer calculated in the test period of 1 year.
• This is how the final data set looks like:
• 

2. Pre-processing:

As we can see from the description above of the input variables, the difference in the min and max values is considerably high.

The values under some variables like monetary_value, average_basket_value are quite skewed which would affect the model training process. We have applied some pre-processing on the data to normalize.

We initially apply a log function on some columns with high range values like monetary_value, time_between, avg_basket_value & avg_basket_size to bring down the values.

Then subtracted all the row values by their mean and divided by the standard deviation, for normalizing the data for all the columns.

The normalization of data reduces the range of values for the model to train thus speeding up the training process and optimizing its performance.

3. The Two-Stage Pipeline Approach:
Most CLTV models use the conventional approach of predicting the CLTV value for a customer using the RFM variables.

But there is a possibility that a customer might have churned (stopped doing business with the company) within the training period and despite that, a finite CLTV could be forecasted for that customer in the test period.

A drawback of this approach is that it won’t be able to take customer churn into consideration before training. Hence, in this model, we are going to follow a 2 stage pipeline approach which includes:

1. Classification model for predicting customer churn:

The customers who have a monetary value <=0 in the test period are supposed to be the customers that churned in the training period.

So we label the customers that churned as False and the other as “True”. We train a classification model on 80% of the data as a training set using deep learning with the features described above for predicting whether the customer churned in that time period or not.

We trained a classification model with ~83% accuracy and auc of 0.737.

2. Regression model for forecasting lifetime value:

We train a regression model using deep learning on the customers who did not churn from the training set to predict the lifetime values for next year.

Using this model, we predict target_monetary value for customers in the test set. The following parameters were used for training the regression model.

We then combine the two model outputs using probability from the classification model and the monetary value obtained from the regression model.

The final_amt is calculated as the product of predicted probability and target_monetary value which will be considered as the lifetime value of that customer for next year.

The mean absolute error of this model is calculated at about $103. We also calculated the overall lifetime values across all customers as: