It is amusing that it took years of academic-use-of-classifiers-on-timelined-datasets to rediscover the old quant trading wisdom: Cross validation on a timeline does not mean anything, roll-forward testing needed.
It should be unsurprising that bad practice with data during ML is *also* bad practice when ML is applied to malware, but hey ...
So torn about that paper. It is a good paper, but the fact that it had to be written to point out the blindingly obvious makes me sad.
