Currently when a summarizer performs the summarization, we take the last column as the targets column. This is a somewhat obscure expectation and a point of brittleness. We could simply document it better, but slightly changing the design to return it as a discrete data structure makes the intent clearer and the code more resilient.
https://github.com/MattRWallace/NHLPredictor/blob/fcdf75edc77f19c4d357d1bee3d9fb8c324597c6/src/nhl_predictor/trainer/linear_regression.py#L47C1-L52C59