The workflow before#
Our client is a major professional sports organization competing for fans against an at-home viewing experience that keeps getting better. Attendance swung with the team's record, and younger fans expected a game day that felt personal and worked from their phones. The organization knew a lot about its fans, but only in pieces. Ticketing, concessions, merchandise, parking, the mobile app and social media each kept their own records, and nobody could see one fan across all of them. Marketing fell back on mass campaigns that sent the same offer to everyone. Concession stands were staffed by habit rather than forecast, and lines got long at the worst moments.
What we built#
We built a fan data and personalization layer on Databricks that pulls every source into one view of each fan, then puts that view to work on game day. Recommendation models suggest the food, offers and experiences a particular fan is likely to want, and those suggestions reach the mobile app and stadium displays as the game happens. Demand forecasts help operations staff the stands.
One view of every fan
AutoLoader ingests more than 50 million data points a day from ticketing, mobile apps, concession registers, parking and social media into Delta Lake. The history covers more than 2 million fans across 5 seasons. We built ingestion to scale up on game days, when data volume runs about 10 times a normal day, because that's exactly when the data is most useful.
Recommendations that know the context
Collaborative filtering models learn from purchase history and demographics, then factor in the weather, the opponent and the time of the game. A family at an afternoon game sees a kid-friendly meal deal, while a group of young professionals at a night game sees a craft beer tasting. The models predict individual preferences with about 85% accuracy.
Offers delivered where fans already are
Custom APIs serve recommendations to the mobile app and stadium displays during the game. We sent offers to the screens fans already use in the building instead of to an email the next day, because a promotion is worth the most while the fan is still in their seat. Targeted offers converted 35% better than the old mass campaigns.
Forecasts for crowds and concession lines
Demand forecasting tells operations how busy each stand will be, so managers can staff ahead of the rush. Concession wait times fell 22%. Attendance forecasts, accurate to about 92%, give the ticketing team a basis for dynamic pricing and earlier marketing pushes.
How we delivered it
We delivered in four phases over 16 weeks: data integration and quality checks, the analytics layer and first models, the recommendation engine and operational dashboards, and a pilot across 10 games to tune everything against live results. Weekly workshops kept ticketing, marketing, operations and IT aligned. We trained more than 200 staff on the new dashboards and recommendations so the work changed what happened on the concourse.
Before and after#
| Measure | Before | After |
|---|---|---|
| Spending per fan on game days | Baseline | 27% higher |
| Promotional offer conversion | Mass campaigns, same offer for everyone | 35% higher with targeted offers |
| Concession wait times | Baseline | 22% shorter |
| Fan satisfaction scores | Baseline | 18% higher in post-game surveys |
| Fan data | Split across ticketing, concessions, merchandise and social | One view per fan |
What's next#
The same fan view now supports premium experience packages built for specific segments, which have already lifted premium sales 40%. The pattern extends to more venues and to other places where guests spend money on site, like casinos, cruise lines and large retail, wherever a visit can be shaped by knowing who's there.
