This is a project I did for online in-image ad vendor GumGum.
GumGum aims to forecast ad impression numbers on its affiliated websites in the next 30 days based on the past 365 days data.
Five predictive models are created and tested:
- Seasonal naive model
- Linear regression model with trend and seasonal dummy variables (day of week)
- SVM regression with trend and seasonal dummy variables (day of week)
- STL + ETS/Arima model
- ARIMA model
The final model is an ensemble of:
- Linear model w/ quadratic trend
- Linear model w/ cubic trend.
- SVM regression with trend and seasonal dummy variables
- ARIMA model
The model's performance is measured by RMSPE (Root Mean Squared Percentage Error).
RMSPE of the final ensemble model is 5.65%.
Here's the link to my code and visulization: ad-impression-forecast-web-view