OfferMatch
The right offer.
Right now.
One customer. One moment. Five possible offers. OfferMatch ranks them by how likely this customer is to use each one, and tells you which to send.
Rank the offers.
Describe the customer and the moment. The result updates as you go.
Your recommendation appears here.
Fill in the details on the left. The offers are ranked as soon as the last one is in.
Make this offer stronger
Score a whole customer list.
Upload an Excel or CSV file with one customer per row. Every customer gets an offer, a chance and a message.
| Customer | Best offer | Chance | Decision | Message or reason |
|---|
Missing details are filled from up to 200 real survey respondents. A row is marked "Needs a detail" when a missing detail would change the answer.
Checked against Orange.
| Case | Offer | Orange | App | Result |
|---|
Cases tested this session.
Every run is listed here so new cases can be checked in Orange and any differences recorded in the defect log.
| # | Customer and moment | Terms | Coffee | Takeaway | Budget | Mid-price | Bar | Advice |
|---|
How it decides.
A prediction is only useful when it leads to one clear action.
-
1
It learned from real answers.
567 drivers each judged about 22 offers in different situations: 12,610 answers in all.
-
2
It scores every offer.
For the customer and moment you describe, it estimates the chance of each of the five offers being used.
-
3
It recommends one, or none.
Send the best offer when its chance is high enough. When none is, send nothing and save the message.
The model, in numbers.
A Random Forest of 300 decision trees, built and validated in Orange.
0.762
AUC on 113 people the model never saw. It ranks a user above a non-user about three times in four.
30/30
Test cases where this site gives exactly the same answer as Orange.
49%
Of the model's decision weight sits on the offer type. Expiry and coffee habits come next.
Limitations.
Predictions come from a model trained on a US survey in which drivers said whether they would accept imagined coupons. They are stated intentions, not real redemptions on an Indian app, so treat each percentage as a guide, not a promise. The model covers only the situations in the survey, which is why some combinations cannot be selected. Age, gender and occupation are not used.