Hello,

I am using 5 minute Bitcoin/USD logreturns since 2017 (n=256,705) and am trying to estimate a GARCH(1,1) model.

The upcoming error " r(430) flat log likelihood encountered, cannot find uphill direction " seems to be an error that is frequently occurring in this context.
The solutions I found using this Forum and the Stata guides can be summarized as follows:

1) eliminate outliers
2) set initial values
3) change convergence criteria

1) I winsorized the data to eliminate the possibility that outliers are the reason for non-convergence
2) I used the following code: arch logreturn_w, arch(1) garch (1) from(0.008 0.9 0.1 0.001, copy)
3) I used the following code: arch logreturn_w, arch(1) garch (1) tolerance(0.1) nrtolerance(0.1) ltolerance(0.1)

I also tried several different starting values and tolerances, all of them without successful conversion.
Could you please check, if there is a problem with the codes (since I am not allowed to use another way to eliminate outliers)?
Stata runs the maximum-likelihood estimation so I assume there is no syntax error etc, but maybe the values are unreasonable?

In addition, there might be other ways to handle the upcoming problem that I have not found or read about, yet?


Thanks a lot for any help in advance.
Mathis