How to create a up-sell repeat purchase model

What is a up-sell repeat purchase model?

A cross-sell repeat purchase model scores a customer account for it's likelihood to purchase a product that they have bought from you before. This means that in order to score accounts against this model, the account must have purchased the target product of the model at least once before. 

 

What do I need in my CDP to be able to create cross-sell models? 

• You must have imported product purchases and product bundles. 

What is an example use case for a cross-sell first purchase model? 

A cross-sell repeat purchase model's value is really in it's name - you want to use this model to identify accounts that are likely to purchase the same product from you again. You want to use this model in a campaign that has the purpose of driving existing customers to purchase the same product from you again 

How do I build this model in Rev.Up?

Before you can build this type of model, you need to do some prep work. 

Step 1: Identify the campaign you want to run 

Before you build a model you first must decide what the purpose of your campaign is. In the cross-sell first purchase article, we created a model that will help us predict which customers will purchase display for the first time. But now we want to know which customers are likely to purchase displays again.

A Repeat Purchase model is set up exactly the same way as a First Purchase with one small but important difference: You must select the period that represents a lapse. So let’s walk through this.

Let’s say the Hooli Product Manager for Displays thinks that if companies buy our Displays and like them, those companies will buy more again within six months. So the Product Manager wants me to figure out which companies are likely to buy again.

 

Step 2: Build the segment that represents the audience you want to target 

Since our campaign is targeting existing customers to make a purchase of displays again our marketing manager has built the below segment. 

• Customers who have purchased a display product 
• Customers who are located in the United States, Canada, Mexico (Use this and other firmographic factors to limit the your customer base to those in a certain market. This is helpful if you want to run different campaigns in different markets that still target the same product.) 
• Segment Name: NA Customers - Displays Repeat Purchase.

 

Step 3: Use your segment to create a repeat purchase model 

Now it is time to score your segment and build a model. Go to the modeling section within Rev.Up and click on Create Model.

Choose the segment that you built in the previous step. In our case that segment is called NA Customers - Displays Repeat Purchase.

Choose the product that you want to sell to your customers. In our case, this product will be Displays. 

You must also choose the timeframe in which the customers in your segment have not purchase the product. You can use this setting as a way to train the model on how often customers should make a repeat purchase of a product. For example, some products may be purchased monthly, while other products only need to be purchased every 6 months. This setting will vary based on the product. 

In our case, our marketing manager thinks the product should only be purchased every 6 months. 

This screen may seem confusing, but it is actually simple when we get down to it. You have two options on this screen: 

Likely to Buy: Scores your customer accounts using a rating scale of A - D on their likelihood to purchase displays. A is the most likely to buy, D is the least likely to buy. 

Likely Amount of Spend: Measures your customers likely amount of spend on making a purchase of displays. 

Make some selections about how you want to train your model. There are several options on this screen that we will break down. 

• Scoring: Number of accounts in the segment that will be sc
• ored. It’s best practice to create a very specific segment, but the system will filter out any accounts that don’t meet the scoring criteria (fit the segment, have previous transactions, have not purchased the target product), which is why this number could be lower than the number of accounts in your segment.

• Records Selected: 11,913 records will be used for training. 

  • How did this system come up with this number? It’s a little confusing, but to give you a brief overview, the system is looking at accounts in our segment and accounts that have made a Repeat Purchase, and it’s multiplying that by the number of eligible periods.

  • Eligible period: : is dependent on what you selected as your lapse period. The model looks at whether or not a previous purchaser had the purchase lapse specified (so 24 months in this case). If yes, the record will be included in training. Then the model looks ahead to see if the purchaser did make a repeat purchase and tags the account accordingly.

  • The main thing to concern yourself with here: The number of records selected should be at least 5,000.

     

• Repeat Purchases: The number of times customers made a repeat purchase of the target product.  In our case, this means the number of customers who have purchased the product once from me within and then purchased the product again. These accounts become what the model uses to compare against - basically accounts that have made a repeat purchase are good and I want to compare them to other accounts that have not done this yet. 

  Main thing to concern yourself with here: Make sure this number is at least 50. Job will fail otherwise. The optimal model will have at least 200 first purchases. 

  • Let's break this down. Our segment criteria for building the model are: 
    
    • Customers who have  purchased a display product 
    • Customers who are located in the United States
    • This means that there are 914 instances where an account made a repeat purchase of a display product. 

Now that we understand how the universe used for modeling is created, we can talk more about how we can refine the training set used by the model. Adjusting these settings will change the number of records selecting for Scoring, Records Selected and First Purchases. 

 

Only use successes where spend in the month/quarter was: Using this removes purchases from the model training set that do not meet the criteria. You can use this setting to set the total amount of spend that you are looking for when a customer makes a first purchase. You can adjust this to target accounts that are likely to spend more or less in a first purchase of your product. 

Only use successes where quantity bought in the month/quarter was: Using this removes purchases from the model training set that do not meet the criteria. You can use this setting to set the total amount of units purchased that you are looking for when a customer makes a first purchase. You can adjust this to target accounts that are likely to purchase a certain number of units in a first purchase of your product. 

When modeling, only use historical records from the last: Using this forces the model to look only at data with the selected time range. Use this when something significant changed in the past (with Account, Transaction or Product tracking) and you want to keep older data out of the learning.

An example if something significant is if the way the product is packaged and sold was changed. 

Let me model on a similar segment, not exactly the one I want to sell into: This uses a separate segment for model training, but will still score the records in the original segment you created. If you have insufficient data to train the model using your target segment, you can use a similar one that’s larger in its place.

For example, if you wanted to break into the Scandinavian market but don’t have many Scandinavian accounts in your system yet, you could train the model on a segment of Western European accounts but assign scores to the Scandinavian accounts. You’d do that by selecting your Scandinavian segment on the screen that tells you to “Select the segment you want to sell into” and type in your Western European segment here.

Let me model on similar products, not exactly the ones I want to sell: This setting replaces the actual product in your selling situation with a different product for the purpose of training the model. This setting is useful when you have  insufficient data to train using the target product that you want to sell. In this case, you can use a similar one with fuller history in its place.

For example, if we just started selling Displays and don’t have much history of transactions involving Displays in our system, we could select another Bundle that we think is similar and has more history. (i.e. Digital Signage).

Once you have made these selections, you can click Model. The model will take 2 - 4 hours to create. Once it is complete you can come back and review. 

 

How do I know that my cross-sell model is good and I should publish it? 

Unlike account fit models, the platform has checks in place to make sure that the model you build will be good to use once it has been made. The model wizard will not allow you to build a model if there are not enough records available to use to train the model. 

Because of this, you are not required to review the model as closely as account fit models. 

To ensure that you have a good model, you will want to make sure that you follow all of the steps in this article closely. 

 

How do I activate my model and use the output to build a segment? 

Click Activate Scoring on the model dashboard to see what your model predicts for your target Accounts. You can see the Rating buckets, and change the bounds to tune your Ratings.

Click Publish Configuration to create a Scoring Action in your Process & Analyze queue. The next time you run Process & Analyze, your Ratings and Scores will be updated in My Data. You can use them to make new Segments, Rules-Based Ratings, or Plays.

 

 

Click Activate Configuration to create a Scoring Action in your Process & Analyze queue. The next time you run Process & Analyze, your Ratings and Scores will be updated in My Data. You can use them to make new Segments, Rules-Based Ratings, or Plays.