Asynchronous correlations can only be seen via a longterm view; data gets more precise overtime but sacrifices short term accuracy from less data points. The inter day price fluctuations in markets do not reflect long term correlations. Most returns are calculated and compounded weekly/quarterly, so it's best to backtest against long term extant data than sets from the brevity of a few prior days, though it can be argued via a time series that previous events have serial correlations. This is still really tragic because while people (and a multitude of startups) claim that they can fix this with machine learning, they can't because the resolution of data is not that great, and while it seems like a lot, we only have historical data from a few cycles. However, this is not to say machine learning is a of nugatory applications. Tune in to Bridgewater Associates' Co-investment officer from MIT, Greg Jenson as he delineates five new trends where machine learning does apply in their investments.