The challenge
Audience measurement tells broadcasters, platforms and advertisers how many people watch what. Viewing has spread from cable TV to streaming services and YouTube, and no single source sees all of it. The products needed reliable forecasts of viewing and watch time across those platforms, built from sources that each cover only part of the picture, with methods that keep working when the product launches in a new country.
Kristof did this work over two years as a data scientist in the measurement company's own team, before our studio started.
What we built
- Forecasting models for viewing and watch time across cable, streaming and YouTube.
- Data fusion: combining sources that measure different things, for different groups of people, into one unified dataset.
- Population projection: scaling what is measured in a sample up to the whole population, with statistical methods that hold up under scrutiny.
- New modelling approaches for existing products, and the methods adapted so they work in new international markets.
Architecture
- Models and statistics in Python
- Production pipelines in Docker on AWS
- Recurring runs automated with GitHub Actions, maintained in an agile team
Outcome
More accurate predictions for existing products, a methodology ready for new markets, and production runs that happen automatically instead of by hand.
What carries over
Most forecasting problems look like this one: the data you have covers part of what you want to know, the sources disagree, and the model has to keep running long after the first version. Combining sources honestly, checking forecasts against reality, and automating the boring parts is how we approach forecasting for clients today.
Stack
Python · pandas · scikit-learn · statistical modelling · Docker · AWS · GitHub Actions