The gap is messing it up. I think the MA with the longest period (the one with the longest tail) hasn't completely compensated for the gap by the time price begins to catch up to it again.
I was developing an MACD mean reversion system that looks pretty good but some of the individual drawdowns are pretty nasty. While looking at those drawdowns, I saw it was very fond of trading on these gaps. My averages weren't compensating for gaps either. So imagine the price takes a sudden and drastic drop, the averages will move down. Right about the time they hit where the price is, they'll seem to pick up again and give a false signal that it would be reverting. This isn't valid this; when something drops like that it often means a change in information in some way--the situation is different.
The odd thing was how despite this, it was profitable. I considered it less a mean reversion strategy and more a dead cat bounce system of some kind. In other words I should better understand what is happening there specifically and write a strategy for that rather than rely on this MACD setup.
I eventually engineered a solution for the gaps that made the moving averages re-prime themselves if there is a gap of a certain percentage size. I write "engineered" not because it was technically difficult or anything, but because it was tied to what is considered a gap, and that could be adjusted. Anyways, if it noticed what it considered to be a gap in a moving average of period N, it would wait N samples before reporting anything again while it basically resets itself. Anything using that number (like MACD) would chew up the NaN's and themselves report nothing until things were properly in perspective again.
