I played around a bit in the Grupo Bimbo Inventory Demand kaggle, just to get experience working with a dataset that was 3.2 GB and 74 million rows, and to submit to my first real Kaggle competition. Although I didn't rank very high, I did one investigation that took a different angle from most people: looking for evidence of a causal effect of one variable on another with a time lag, in this case whether the number of returns for a product predicted whether the client would buy more or not.
It is contained in this
To rerun it requires that the train.csv data file for the competition is in the directory '../input' relative to the notebook.