- Newest
- Most votes
- Most comments
In my experience (which is limited, but matches what I've heard from others in the domain), raw weather signals need careful feature engineering to produce usefully-predictive inputs for human behaviour forecasting (like retail sales, etc).
The Amazon Forecast Weather Index uses built-in, proprietary, featurization to tackle this effectively on your behalf - optimized across a range of use cases. As you found, the exact choices are not currently documented.
So my suggestion would be to start out with just the Weather Index enabled and not worry about bringing or choosing custom weather attributes.
If you have weather data available and find reason to think some particular weather feature important to your domain isn't being captured well enough (or just want to try your favourite weather feature out) - then can experiment with adding your feature(s) to explore the impact on model accuracy. The model explainability (feature importance) scores can also help you get an idea of how strongly your proposed features are driving the model.
Relevant content
- Accepted Answerasked 3 years ago
- asked 14 days ago
- asked 2 years ago
- AWS OFFICIALUpdated 3 years ago
- AWS OFFICIALUpdated a year ago
- AWS OFFICIALUpdated 3 months ago
- AWS OFFICIALUpdated 2 years ago