The workflow before#
PENN Entertainment runs casinos and sports betting across North America, and it had grown fast. Its data grew with it, across on-premise systems and the cloud, without a plan for how the pieces fit together. Departments used different reporting tools. Marketing couldn't easily share campaign results with analytics, and the interactive gaming division ran apart from the core business intelligence setup. Analysts spent hours assembling reports that described last month. Nobody could forecast which guests were about to drift away or put a reliable lifetime value on a customer, so retention and marketing spend were decided on hindsight. In a regulated industry, governance across all those copies of data was its own headache.
What we built#
We built one data layer across PENN's on-premise and cloud environments, so every team works from the same source of truth wherever the data physically lives. Analysts can query across both without copying data around. On top of that layer we added machine learning that predicts which guests are likely to churn and what each guest is worth over time, so marketing can act before spend drops.
Querying across on-premise and cloud
Microsoft PolyBase lets analysts query on-premise and cloud data together, without building an ETL job for every question. We chose that over a wholesale migration because PENN's operations couldn't pause while data moved, and much of it had good reasons to stay where it was.
A secure bridge to interactive gaming
AWS VPC peering connects PENN's core operations with its interactive gaming division over private, high-throughput links. Online and on-property behavior now sit in the same view. We kept the architecture cloud-agnostic across AWS and Azure so PENN isn't locked into one vendor.
Churn and lifetime value models
Models built with PySpark on Amazon EMR and deployed through Azure Machine Learning score guests on churn risk and lifetime value. Marketing uses the scores to aim retention offers at guests who are drifting and to put budget where customer value is highest. The models are 40% better at identifying likely churners.
How we delivered it
We started with six stakeholder workshops to map the existing data and agree on where integration mattered most. Rollout was phased: the data science and analytics teams went first as a proof of concept, then research, then the wider enterprise. We trained more than 50 data users and set up new governance for data sharing and security along the way.
Before and after#
| Measure | Before | After |
|---|---|---|
| Identifying likely churners | Backward-looking reports | 40% more accurate with predictive models |
| Report generation | Hours | Minutes |
| Data access across teams | Separate tools and silos | One data layer across on-premise and cloud |
| Data users trained | None | 50+ |
What's next#
With guest data unified and churn and lifetime value scores in place, PENN can move from scoring guests to acting on them automatically. Next steps include personalized offers triggered by a change in a guest's churn score, and extending the models to new markets as PENN grows, using the same governance framework.
