Predict Cash

Despite the growing push in the financial industry to provide almost fully digital platforms, there remains a tremendous need for physical locations to provide full-service assistance and great customer service. At Teachers Credit Union (TCU), they strive to make every interaction between members and team members as seamless as possible. One tremendous consideration in financial member service is having enough cash on hand at a branch to serve the member’s needs.

 
The problem becomes complex as the organization adds new locations with unique needs, new members, different cash-on-hand limits, and ever-changing regulations with regards to banking and cash management. Teachers Credit Union (TCU) is committed to being the best in-store member service experience it can be, and it starts with the logistics of providing enough cash. They operate 60 branches across Indiana and Michigan, with over 300,000 members interacting both digitally and in-person.

 
TCU’s challenge was to build a model that more accurately forecasts daily branch cash usage. Too much cash-on-hand will put the branch over their cash limit, but not enough cash will cause an expensive “special order” between normal deliveries. 

Prophet & Portal

Knowing that it was important for TCU to have a solution that could be easily reproduced, we chose to use a time series forecasting package in R called Prophet, which is a free, open-source procedure developed by Facebook. It is capable of generating robust models that handle outliers well, but is easy to understand and implement for even the most novice data analyst.

1. We generated a model using the information about holidays, Saturdays, and Sundays as regressors.

2. Then, we created a grid to identify the parameters that could be tuned to improve the accuracy of the model.

3. Lastly, we created a loop to generate the optimal combination of parameters for each branch to generate the most accurate predictions. Meaning, we developed 49 different outputs, one for each branch.

 The evaluation metric for this competition used a measure called SMAPE, or symmetric mean absolute percentage error, which, put simply, measures how closely our predictions matched the actual values. Our team achieved a SMAPE value of 78.5, which was the third most accurate achieved, allowing us to move onto the 2nd round.

For the 2nd round, we needed to create a 5-minute video and present our results to each school to be judged.  The top team from each school would be chosen to move to the 3rd and final round.  We used Data Storytelling techniques, presented our findings, showed our Tableau dashboard, and concluded our story. 

After finding out we won the 2nd round for Notre Dame, we put together a plan to incorporate feedback from the judges into our final presentation to TCU. 

During our final round preparation, we developed a tool, using R shiny app.  This allowed TCU to execute predictions with a click of a button.

Success

Below is the feedback we received from TCU on our submission.

“ MEET BILL. Excellent! We all loved that and it was highly effective”

“ You built a nice web interface. The web app seemed robust for the short time of the project. This was a differentiator. This was the best beginning to end project in the ones reviewed.”

“ Good roadmap and use of animation to guide the audience’s attention. It helps us follow and listen to your presentation instead of reading. Excellent!”

“ Using R and explaining the prophet package along with the help website to view some FAQ and get documentation. Essential in reproducibility.”

“ Good job explaining SMAPE in a way that our executives could understand.”

“ You incorporated TCU into the presentation as well as the final web product with the Core Values and the community partners.”