Respecting chronological order when splitting training and test data.

Removing the time index before all analysis.

Randomly shuffling future observations into the training set.

Evaluating only with a clustering metric.