clemvst/timeseriesforecasting

Time Series Forecasting

Kaggle Project for Machine learning course at Mines Nancy school of engineering.

Predict Future Sales Final project for "How to win a data science competition" Coursera course

Training of the model

To install the prerequisites into the pytorch_lab conda environment, run

conda pytorch_lab_env create -f pytorch_lab.yml

The main code is in the Jupyter Notebook but also converted in PY files if needed.

Analysis

Of course my model is not really representing the reality and is not complete, I would normally need to create a model by item-shop duo and then connect them all together for example. To work with a bigger dataset and add additional values I could also use a k-fold technique. Maybe convolutional layers are also not the best way to predict this kind of information. Overall, the Loss looks good but I'm not sure it would with a different dataset of values.

If I had more time here's what I wouldve done:

• Build a model for each item-shop duo
• Connect them all with a fully connected layer
• Build a model by shop (for all items)
• Build a model by item (for all shops)

I could've compare all those models and select my favorite one or analyse them too.