At first glance it makes sense to buy on the expectancy it would rally to close greater than 5%. It would be more prudent to first check the correlation between the few outliers that close down greater than 5% and an intraday value that was below 5%. Obviously all down closes greater than 5% were down that much at some point intraday. The important test would rank those intraday swoons against the ones that still existed at the close, rather than ranking them against the mean which rarely closed down 5%.
That type of design error in determining what to measure something against wasn't included in my mention of bugs. To me a bug is a typo that references the wrong value of a variable. That typically produces unpredictable results going forward, unless the bug itself captures a market characteristic by accident which isn't likely. Bugs are also logic errors that juxtapose the flow of the program in a way that although it's consistent it's still likely to produce unpredictable results.
Design errors where something important wasn't considered fit into some other category than bugs. If I'm determined to do something and a good idea of what it is then I always figure out a way to accomplish it. However the type of correlations you're discussing seem at first glance to be quite different than what I normally do. But prediction based on historical norms and outliers is a cool area to investigate. Good luck with it.
Market Data 
Markets 
AI
