I also read the book and liked it a lot. Based on the extensive focus the book gave on their data collection (e.g. going into local state offices to digitize records), pipeline and academic rigor, i would think it's just looking for correlations and focus on medium term trades (minutes)

It's easy enough to do, but it requires a strong culture of honesty to not trick yourself. For instance you can regress performance based on 32 different lags on another variable and you'll almost certainly find spurious correlations. But in a more comprehensive framework you can adjust your score to account for such a search.

The people in the book state that they see the data no more than numbers and don't even try to come up with a story. And they can only do so because they trust their system to handle spurious correlations