TL;DR : Given a set of time series as input, this R script groups timeseries based on shared features and generates forecasts using Facebook's Prophet forecasting package.
- Remove activity from services that have not incurred a cost in the last 30 days.
- Create features for each timeseries:
- length of available time series data (short vs long)
- avg daily cost for last 30 days (low vs high)
- Group timeseries data based on feature groupings:
- high costs and long time series are forecasted individually.
- high costs and short time series are forecasted together to improve accuracy.
- low costs or short time series are forecasted together to improve accuracy.
- Anomaly Detection and Removal
- Generate forecasts for each forecasting group.
- Generate forecast accuracy scores using cross validation (Not implemented.)