Inspiration
Scotia owns young Canadians' chequing accounts but loses them to Wealthsimple the moment they're ready to invest. We saw a timing problem, not a product gap.
What it does
Reads 65 behavioral signals per Scotia depositor, predicts who's ready to invest in the next 90 days, picks the right product, and delivers the nudge through the right channel.
How we built it
Generated 250,000 synthetic customer profiles, trained three classifiers in scikit-learn, and built a working demo that fires real personalized emails and in-app cards.
Challenges we ran into
Hyperparameter tuning on 200K rows was painfully slow. Balancing realism in synthetic data.
Accomplishments we're proud of
A working end-to-end prototype: real ML predictions feeding into a live email send and a Scotia-styled app demo.
What we learned
Picking the simplest model that works beats forcing the fanciest one. Logistic regression won readiness scoring, interpretable, fast, and good enough.
What's next
Deploy the saved models behind a real API, A/B test on a pilot segment, and expand triggers beyond tax refunds and salary bumps.
Built With
- jupyter
- python
- typescript
Log in or sign up for Devpost to join the conversation.