Same code+Data+Statsmodels Version -> Different Results

[YueHaiming]

Hello guys!

I was running a SARIMA (2,0,1)x(1,0,2)x24 model using my old laptop. It took around 20 to 30 seconds to finish and did not throw any warning. Python version is 3.9.7 and statsmodels version is 0.14.0.dev535.

I got a new laptop a couple of days ago. I installed Anaconda with default Python = 3.9.13. I used night wheels to install 0.14.0.dev535 statsmodels as well. I tried to install the same Numpy and Scipy, but I was not able to do that because of the Anaconda distribution issue (Numpy and Scipy are newer on the new laptop). I was expecting that the code would run faster because the new laptop has much powerful CPU and more cores. However, it took around 4-5 minutes to finish and throw warning "ConvergenceWarning: Maximum Likelihood optimization failed to converge. Check mle_retvals". I tried different older versions of Scipy, some did converge but still took 4-5 minutes.

For both, I set maxiter at 10,000.

Can anyone help me to clarify the issue? Much appreciated!

[bashtage]

Do you have the same scipy versions? They have been changing optimizer code a lot recently and it isn't always helpful.

…

On Sat, Mar 4, 2023, 00:50 Haiming Yue ***@***.***> wrote: Hello guys! I was running a SARIMA (2,0,1)x(1,0,2)x24 model using my old laptop. It took around 20 to 30 seconds to finish and did not throw any warning. Python version is 3.9.7 and statsmodels version is 0.14.0.dev535. I got a new laptop a couple of days ago. I installed Anaconda with default Python = 3.9.13. I used night wheels to install 0.14.0.dev535 statsmodels as well. I tried to install the same Numpy and Scipy, but I was not able to do that because of the Anaconda distribution issue (Numpy and Scipy are newer on the new laptop). I was expecting that the code would run faster because the new laptop has much powerful CPU and more cores. However, it took around 4-5 minutes to finish and throw warning "ConvergenceWarning: Maximum Likelihood optimization failed to converge. Check mle_retvals". I tried different older versions of Scipy, some did converge but still took 4-5 minutes. For both, I set maxiter at 10,000. Can anyone help me to clarify the issue? Much appreciated! — Reply to this email directly, view it on GitHub <#8717>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABKTSRIEPNDCB7XW5SQVVH3W2KGXBANCNFSM6AAAAAAVPG2ITM> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[YueHaiming]

Do you have the same scipy versions? They have been changing optimizer code a lot recently and it isn't always helpful.
…
On Sat, Mar 4, 2023, 00:50 Haiming Yue @.> wrote: Hello guys! I was running a SARIMA (2,0,1)x(1,0,2)x24 model using my old laptop. It took around 20 to 30 seconds to finish and did not throw any warning. Python version is 3.9.7 and statsmodels version is 0.14.0.dev535. I got a new laptop a couple of days ago. I installed Anaconda with default Python = 3.9.13. I used night wheels to install 0.14.0.dev535 statsmodels as well. I tried to install the same Numpy and Scipy, but I was not able to do that because of the Anaconda distribution issue (Numpy and Scipy are newer on the new laptop). I was expecting that the code would run faster because the new laptop has much powerful CPU and more cores. However, it took around 4-5 minutes to finish and throw warning "ConvergenceWarning: Maximum Likelihood optimization failed to converge. Check mle_retvals". I tried different older versions of Scipy, some did converge but still took 4-5 minutes. For both, I set maxiter at 10,000. Can anyone help me to clarify the issue? Much appreciated! — Reply to this email directly, view it on GitHub <#8717>, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABKTSRIEPNDCB7XW5SQVVH3W2KGXBANCNFSM6AAAAAAVPG2ITM . You are receiving this because you are subscribed to this thread.Message ID: @.>

Thank you for your relpy. Yes, I installed the same Scipy version. It did converge but still took 4-5 minutes compared with 20-30s on my old laptop. I am super confused...

[bashtage]

One difference is probably the linear algebra version. Which CPUs were in the old and new?

[YueHaiming]

One difference is probably the linear algebra version. Which CPUs were in the old and new?

Hi, sorry for my late reply. Old has i7-8650U and New has i9-12900HK. It is a huge upgrade. Do you think linear algebra version (in numpy) can cause so big difference? 20-30 seconds slowed down to 4-5 minutes? Thank you!

[ckwan48]

Do you have the same scipy versions? They have been changing optimizer code a lot recently and it isn't always helpful.
…
On Sat, Mar 4, 2023, 00:50 Haiming Yue @.> wrote: Hello guys! I was running a SARIMA (2,0,1)x(1,0,2)x24 model using my old laptop. It took around 20 to 30 seconds to finish and did not throw any warning. Python version is 3.9.7 and statsmodels version is 0.14.0.dev535. I got a new laptop a couple of days ago. I installed Anaconda with default Python = 3.9.13. I used night wheels to install 0.14.0.dev535 statsmodels as well. I tried to install the same Numpy and Scipy, but I was not able to do that because of the Anaconda distribution issue (Numpy and Scipy are newer on the new laptop). I was expecting that the code would run faster because the new laptop has much powerful CPU and more cores. However, it took around 4-5 minutes to finish and throw warning "ConvergenceWarning: Maximum Likelihood optimization failed to converge. Check mle_retvals". I tried different older versions of Scipy, some did converge but still took 4-5 minutes. For both, I set maxiter at 10,000. Can anyone help me to clarify the issue? Much appreciated! — Reply to this email directly, view it on GitHub <#8717>, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABKTSRIEPNDCB7XW5SQVVH3W2KGXBANCNFSM6AAAAAAVPG2ITM . You are receiving this because you are subscribed to this thread.Message ID: @.>

I've been having similar issues with SARIMAX and it uses the numerical solver lbgfs, do you think scipy has any impact on this, too?