CUSTOMER GROWTH
Cross-sell prediction: find who will buy the next product
Cross-sell prediction with TIMi estimates, for each customer, the probability of buying a complementary product and shows the factors behind it, so sales effort goes to the customers most likely to say yes.
Purchase propensity
Modeler compares customers who already bought the product with those who did not.
Behavior first
Behavioral variables usually predict better than socio-demographic ones.
Full base, one server
Every customer scored, for every product you evaluate, on a standard server.
HOW IT WORKS
How cross-sell prediction works.
Three steps, in one engine, on a standard server with no cluster.
1. Prepare the data
Anatella gathers purchase history, the product portfolio of each customer and the behavioral or demographic data available. The result is one row per customer, with a target column that says whether the customer bought the product in the following period.
2. Build the model
Modeler compares the customers who bought the complementary product with those who did not, and estimates a probability of purchase. In general there is one model per product or product family, because the drivers differ.
3. Score and explain
Each customer gets a purchase probability for a given product, with the factors behind it, such as products already held or a behavior close to that of customers who already bought.
GETTING IT RIGHT
What makes a cross-sell model accurate.
Three choices matter more than the choice of algorithm.
Define the prediction window
The target says whether the customer bought the product in a given period after the observation date. In telecom, one month is a common window.
Favor behavior over profile
Recent purchases, usage patterns and contacts predict better than socio-demographic data, which is often estimated from a postal address and is less reliable.
Judge the model on the top of the curve
A cross-sell campaign contacts only a small share of the customers, so what matters is how well the model ranks the top of the list. Modeler reports the AUCtop, a measure focused on that part of the lift curve.
EXPLAINABILITY
A probability says who. The drivers say how to sell.
A probability alone tells you that a customer is likely to buy, not why. A white-box model shows what motivates it for this customer, so sales teams adapt the pitch to the actual reason for interest instead of sending a generic offer.
For sales teams
Sales effort concentrates on customers with a high probability, with a pitch built on the factors the model identified.
For transparency requirements
A model that justifies each result variable by variable gives you the documentation to answer transparency expectations such as GDPR and the AI Act. The compliance assessment itself stays with your legal team.
For model owners
Modeler’s built-in Drift Report shows when a model loses accuracy, so you update it when needed instead of on a fixed calendar.
DATA
What data a cross-sell model uses.
Anatella connects to more than 80 sources, so this data can be gathered without prior custom development.
Purchase history and portfolio
What each customer bought and which products they hold today.
Profile data
Demographic or profile data, when available.
Past commercial contacts
Responses to previous campaigns and solicitations.
Product characteristics
The features of the complementary product being targeted.
PROOF
Seen in production.
Results from our published customer stories.
Cross-sell revenue doubled in 15 days
30 product models, a campaign to 2 million customers and an offer allocation engine, built by 2 data miners on one standard desktop PC.
Conversion up from 4% to 34%
Models built per product rather than per product category are now the standard at BNP Paribas.
26% of the top 15% contacted bought a credit
ING targeted the 15% of business clients most likely to need funding, at two key moments of the year, and raised conversion by 20%.
FAQ
Common questions.
What is the difference between cross-sell and upsell?
Cross-sell offers a different, complementary product to what the customer already holds. Upsell offers a higher or more complete version of the same product. Both follow the same modeling logic, with different variables.
Do we need one model per product or a global model?
Generally one model per product or product family, because the drivers of interest vary from one product to another. Modeler makes it practical to generate these models quickly with AutoML.
Does it work for a new product with no sales history?
It is harder with no history at all. A model can rely on the sales of a comparable product, or on a pilot phase that creates a first training dataset.
How does the cross-sell score relate to next best offer?
Cross-sell assesses the appetite for one given product. Next best offer goes further by ranking several products against each other to keep the most relevant one for each customer.
Can relationships between customers improve the prediction?
Yes, in telecom especially. Where a relationship graph exists, such as calls between subscribers, the Social Network Analysis operators in Anatella can compute variables like recent purchases in a customer’s group of friends, which is usually among the most predictive.
RELATED USE CASES
Keep exploring.
Next Best Offer
Rank the offers each customer is most likely to accept, with the reasons why.
Next Best Action
Choose the best next step per customer: an offer, a retention contact or a service action.
Customer Lifetime Value Prediction
Estimate what each customer is worth over time, and what drives that value.
SEE FOR YOURSELF
Test it on your own data.
Run TIMi on your own data with a free trial, or talk to us about your case.
