Multivariate Time Series Analysis with Deep Learning Solutions
In the end, our best performing submission was BiLSTM with Convolutional Layer with an MSE score of 0.010 in the final phase of the competition. Even though the transformer,that is built from scratch, is a larger model, it didn’t manage to surpass the BiLSTM, getting a score of 0.011 in the final phase.
BiLSTM Model -- Positional Encoding
Seasonality -- Similarity