There might be a mistake in the lecture where ARMA(5,6) and ARMA(6,1) are said to have all significant coefficients. This however is not true and is even shown in the video. The summaries for those two models show that there is at least on highly insignificant coefficient for them.
I’ve found the exact same summaries in my results for those models. The actually correct models I’ve found to be ARMA(6,2) and ARMA(6,5). Looking at LL and AIC there is no clear winner, however LLR_test is applicable here and shows that the (6,5) model is not significantly better than the (6,2) model. LLR_test also shows that (6,2) is significantly better than (6,6) and (1,1). ACF of Residuals of Returns plots show that (6,5) has all lags (up to 40) insignificantly different from 0 while (6,2) does have some, though not disqualifyingly so. Could someone of the 365 team take a look at this? If necessary I’ll be able to provide my .ipynb file.
Oddly enough the lecturer makes a point about ARMA(4,6) that is spot on with my findings, leading me to believe even more that I have the correct results at least.
Thanks for bringing this to my attention! Just went back and checked and you’re right.
When we initially wrote the code and the text to go along with it, the results matched what we describe in the video (ARMA(5,6) and ARMA(6,1) having all coefficients significant). However, statasmodels had several updates between the time we wrote the code (and text) and the time we eventually recorded the video, so some of the default values of the fit function got changed, hence the differences.
I’ll look into these changes and try and figure out what arguments need to be specified to replicate the original results.