Great, I'll be touch. The datasets are NLP and NLU derived correlation matrices used to cluster stocks in different ways outside of standard time-series
My firm uses a similar system that hedges long and short baskets and can confirm that hedging is really the only way to swing trade profitably in the long term.
That document shows that as algorithms might decay, you can swap out the datasets to out-perform. They also conducted a year long study on the results.