Gift Card Breakage Forecast Approaches

Context

Gift card breakage forecast can be done essentially in two different ways. One being purely based on trend analysis. Another based on a machine learning model. We discuss both.

GC Breakage Forecast based on trend analysis and a statistical model.

We already have a rich body of historical purchase and redemption data. Attributes such as time of year of purchase, GC value, customer profile in some form, GC card category, merchant info all exist.

The approach in this case would boil down to the following:

Run a daily/weekly(or adhoc) batch job which looks at each card from the following perspectives.

1. Historically, what has been the breakage on a card over a long term horizon?
2. What has been the breakage numbers for the card in the last four quarters? Do these numbers corroborate with the long term averages? If so, we probably are fine with saying that the breakages can be predicated based on these averages.
3. Does the short term quarterly breakages on the card show a trend? Has it increased or decreased linearly? Can we do a curve fitting? Panera Bread and Starbucks seem to hold steady probably? But Office Depot and Staples will probably have a seasonal variation since back to school aspect would probably have an interplay here. Some form of statistical model seems to be helpful here.
4. We thus conclude, we need to give due weightage to historical trends, but that's not the whole story. Let's say we give a weight of W1 to historical trends.
5. Next is demographics to which the card was sold. Younger customers may exhibit a different general behavior than middle aged or geriatric customers. What does the long term profiling of such demographics reveal. And in the batch run, we could give a weight W2 to this aspect (essentially customer profile).
6. Next, consumer behavior also exists around the remaining un-redeemed value of a GC. We can give a weight W3 to this aspect.
7. Similarly, a detailed secondary research into GC purchase and redemption sentiments and consumer behavior will most probably help surface out other key driving factors.
8. The actual forecast engine then could factor in each of the above factors and come up with a probability number for breakage. And we can then, based on empirical thoughts, say that a 95% confidence level may safely say that the card has become 'remote' and have near zero chances of redemption.

Conclusion

In essence, we could build a progressively complex model based purely on historical trends and statistical models. The actual control parameters of the fitted curve may be derived by looking at historical data with a good degree of success. But this would take quite some effort and adequate expertise in pure and applied statistics to arrive at a competitive breakage forecast model. The key approach here is purely statistical modeling.

GC Breakage Forecast based on a Machine Learning model

This approach seems to fit for a quick kill, especially if we have data. Interestingly, we do not have a labelling issue at all here. We know historically, which cards had a breakage - so data is essentially already labelled. Training data is thus readily available, thus removing the single biggest risk and expense. This makes this approach attractive right off the bat. Again ML tools have become commodity and sophisticated algorithms are already available.

The Approach

1. Create a fully denormalized dataset to capture as many features(attributes) as needed for each transaction. Include seemingly unrelated features like zip code as well.
2. Next, do an EDA to find out possible correlation between features and target outcome(whether the card sale actually had seen a breakage). Choose the right set of features to include in the model.
3. Apply a variety of algorithms that seem fit for the problem at hand, e.g. is this a clustering problem, categorization problem etc. The data scientist will be well equipped to identify the best algorithms.
4. The data engineer would then need to run multiple training runs for the models to tune hyper-parameters of the model.
5. It appears that, with high quality of training data and readily available test data, a model accuracy of more than 90% could be possible.
6. Create the model, and serve it through a suitable model serving infrastructure. The model essentially will take input parameters like, today's date, card type, customer age, zip code, card value etc (these are the actual features that were chosen for the model) and will answer only two things - a) A boolean value indicating if there will be a breakage on a specific card(plastic). b) The confidence level of this prediction.
7. We would use these inferences that the model returns in our daily batch jobs to calculate predicted breakage and the need to send a notification to the customer.
8. A robust and automated pipeline to support continuous updates to the model in response to transaction data arriving everyday, would make the model just better over time.
9. Currently available services/tools on Amazon's public cloud, could help achieve quick successes.

Conclusion

The Machine Learning approach seems to be very promising at this stage. This approach has already been commercialized by leading vendors in this field. The availability of self-labelled data would make unsupervised machine learning possible. The overall cost to develop a competitive, ready to ship product based on this approach appears to be less than USD 1M over 9 - 12 months.