ARIMA? IDK... my bot core dumped and Overight convinced me to gig w/ uber instead. Edit: Becky, becky, becky ;-)
] There are two goals here: Accuracy and Precision. The current "method" describes gives you precision. That is all fine and can be any particular formula. The search for accuracy means a few more things are needed. 1) It has to be context based. So, e.g. the arbitrary "days back" needs to be addressed. There are a few ways to deal this. E.g. make sure the arbitrary context window actionable to the strat. Every strat has a time frame that is cares about, does not care about, and a "did not care about" but now I do care about. Others will allude to this. 2) This brings up the "slippery slope". Briefly, as you define what are essentially meta variable measurements on your "precise", measurements, you can accidently create meta "curve fitting". 3) There is more, but this will get you started, maybe . I understand this might appear to be some BS, handwaving, but this is as old a problem as there can be. Basically this is an old conundrum going back to pre mathematics. It has been studies since the pre Greeks. Bottom line: You need additional layers of logic that take your measurement, and makes a determination of its "usefulness and actionable-ness". Try to keep it simple by making the conclusions actionable or not. You do NOT need to create some precision measurements that are better, you need to interpret them better. Hope that helps.