Amazon tasked us with determining the impact that weather has on their ability to deliver packages. This problem was certainly tough for us to grasp, given that we were dealing with time series data, something none of us had done before. We were given weather data and package performance for each zip code on an hourly basis. Our core idea was to train an individual model for each area of the US, but with a twist. Instead of predicting a raw time series, which is very hard, we deseasonalized the data first. By looking at the trend over time, we are able to break the core signal into its trend and seasonality components. Then, we simply trained using the trend, since that is not time dependent. We ended up advancing the final round (top 4)!
Here's the full paper!