Mixed-input network

One LSTM branch reads the last 30 open days and an embedding branch reads each store's fixed traits, such as type, range and competition. On the 2015 test period, R² was above 0.6 in seven of the nine stores and negative in one. Team of 3.

Forecast daily sales for nine Rossmann stores with a network that reads two kinds of input: an LSTM branch for the last 30 open days and an embedding branch for each store's fixed traits, such as type, range and competition. One global model shares what it learns across stores, because some have only about 600 days of history. Validation covers October to December 2014 so it includes Christmas.

I wrote the preprocessing: loading, cleaning, store features, encoding, a time split, a per-store normaliser and the input sequences. On the 2015 test period, R² was above 0.6 in seven of the nine stores and negative in one.