Eleven Strategy · Hackathon winner · 2024
The InvitAItion: choosing the right guests
Winning hackathon model that raised expected gain per luxury event from about €28K to €49K.
- Role
- Project leader, five-day challenge
- Built with
- Python, scikit-learn, pandas
The challenge
How can we generate the best guest list for a luxury event to increase revenues? That was the objective Eleven Strategy, a leading Paris consulting firm, set half of my MSc class. After a random group allocation, we had five days to present the most convincing data-driven consulting proposal. We competed against ten other teams and won.
Context and data
We were given about 55K transactions from more than 13K customers, plus 10K event invitations, some successful and some not. Our solution was tailored to a luxury business that wants to invite the optimal guests: those most likely to buy after attending an event (a classification model) and expected to spend the most (a regression model).
Current figures
Only about 50% of invitations were successful, meaning the client attended the event: the current invitation strategy was no better than flipping a coin. Most clients who accepted an invitation (93%) attended a single event. To get an unbiased estimate of the current average event gain, we excluded clients who attended several events and set a fixed 90-day window to compare sales before and after each event. The mean event gain was about €28K.
Our solution KPIs
- Average event gain (business KPI)
- Invitation success rate (business KPI)
- Model performance (technical KPI): recall for the classification model, MSE for the regression model
Measuring the true impact of an event
First we grouped customers with similar purchasing behaviour before any event, using K-means clustering, to get a baseline. Then, within each cluster, we compared the average change in purchases of clients who attended an event with those who didn’t. The difference between those two changes is the true event gain.
Six clusters
| Cluster | Clients | Sales (€M) | Purchases | Customer life (months) |
|---|---|---|---|---|
| Loyal Shoppers | 297 (2%) | 31 (30%) | 30 | 16 |
| Women’s Style Enthusiasts | 8,719 (67%) | 58 (56%) | 3 | 3 |
| Men’s Fashion Explorers | 2,364 (18%) | 7 (8%) | 2 | 2 |
| Rare Shoppers | 1,286 (10%) | 3 (4%) | 1 | 1 |
| Tiny Trends | 257 (2%) | 0.5 (0.5%) | 2 | 2 |
| Timepieces Lovers | 163 (1%) | 2 (1%) | 2 | 3 |
We clustered on purchasing behaviour: purchase amount, number of purchases, categories bought and customer lifetime. These are strong indicators of future behaviour and should be prioritised; demographic variables are more useful for reaching and finding customers.
A successful event
In both clusters, average spending changed after the event. Take cluster 1: both attendees and non-attendees spent more, but attendees’ spending rose by more. Since both groups belong to the same cluster, they had similar purchasing behaviour before the event, so the difference between the attendees’ gain and the base gain can be attributed to the event itself (highlighted). Repeating this for every cluster and summing the results gives the total true event gain. This event was successful because there is a gain due to the event.
An unsuccessful event
Here both groups in the same cluster increased their spending by the same amount. That increase comes from external factors, such as better economic conditions or seasonal events like Christmas, not from the event.
Scoring the guests
To find the optimal guests, we scored each customer by multiplying two figures: the expected purchase amount after attending an event (regression model) and the likelihood of attending after being invited (classification model). Customers are then ranked by score in descending order.
An automated platform
We prototyped a fully working platform that takes a new event’s characteristics and returns the optimal guest list. The client can choose the type of event, the product to promote, the invitation channel, the number of invitees, the cost per attending invitee and other product-related variables. The list contains the clients with the highest scores, and the total gain is the sum of their scores minus the expected cost of the event (number of invitees × cost per invitee).
As a data-driven consulting firm, we also proposed additional services: weekly training for an easy onboarding, maintenance of the model and dashboard, an improved data collection process, and continuous model retraining with new data.
Results
The initial invitation success rate was about 50%; our model reached 89% accuracy, with 91% recall on successful invitations. Our model predicted an average expected gain per event of about €49K, up from the initial €28K.
The competition wasn’t about a perfect model, as we only had five days. The goal was to craft the most convincing data-consulting proposal, the one a client would choose. In real life, a solid solution for such a project would need months of work and thorough analysis.
Leading the team
The hackathon was demanding and rewarding at the same time, and the random team allocation made it feel even more like a real work environment. I was lucky to work with equally driven and qualified people, without whom winning would have been much harder.
As project leader, I kept everyone motivated, kept us working towards our common goal, and presented our final solution.
Every member of a team needs an environment where they feel valued, listened to and motivated. That’s why I believe great leaders are great environment creators.