Hi Marcin
I will prepare a quick brief of what I am attempting to emulate plus some real data. The overall exercise is slightly more complicated than fitting a simple Bayesian trend as will become clear when I explain the methodology and its application.. You can then decide  a) whether what they are doing can be replicated without knowledge of the prior they use; or b) whether what they are doing is really so full of statistical problems that the approach is essentially so theoretically weak that  the context of using it for projection purposes is invalid/meaningless -which is my view -but I would appreciate that of Bayesian experts. Essentially in frequentist terms it is so oversimplistic as to be meaningless.
Brian

Brian J Revell
Professor Emeritus (Agricultural Economics)
Harper Adams University UK
Current Chair of Defra Economic ADvisory Panel
Former President Agricultural Economics Society

Tel: home 01952 728153  Mobile 07976 538712
Address: Orchard Croft, Vineyard Rd, Homer, Much Wenlock TF!3 6NG
Alt. Email bjrevell@harper-adams.ac.uk


On Tue, 15 Aug 2023 at 17:52, Marcin Błażejowski <marcin@wrzosy.nsb.pl> wrote:
On 15.08.2023 16:55, Brian Revell wrote:
> Hi Luca,Sven
>
> I would be happy to give Luca's BayTool a go though at my advancing
> years, Scripting is not a skill I want nor need to develop any
> expertise in. KIS is my motto.
> Fair question to ask why would one want to use Bayesian regression?.
> The answer is that scientists now seem to have abandoned frequentist
> statistics. As my personal research is challenging work that fisheries
> scientists are doing on salmon stock estimation who do not publish
> their underlying coefficient estimates, they just publish graphs of
> their 10 year Bayesian trends in official statistics with 5 year
> projections and very wide unsurprisingly (in)credible intervals. I am
> keen to replicate their results and examine their robustness. I have
> the same data from official published statistics and know that most of
> the trends they fit to the 10 year river data sets are not
> statistically significant from zero from frequentist standard log
> linear regression.
> In summary, it is about identifying your opponent's weaknesses from
> the inside

Well,

since prior plays a non-negligible role, so as Sven said: do you have a
"true" prior or you just want to use some data-driven (like Zellner's
g)? It is important, because you deal with non-stationary series with
possible AR component. If so, using "standard" Bayesian set-up and tools
(like g-prior with with analytical solution for posterior moments) may
be misleading.

So, I would advise to follow:
1. Approach developed in late 90-ties in: Koop, Gary & Ley, Eduardo &
Osiewalski, Jacek & Steel, Mark F. J., 1997. "Bayesian analysis of long
memory and persistence using ARFIMA models," Journal of Econometrics,
Elsevier, vol. 76(1-2), pages 149-169.
2. Setup based on g-prior proposed by us in: Błażejowski M, Kwiatkowski
J, Kufel P. BACE and BMA Variable Selection and Forecasting for UK Money
Demand and Inflation with Gretl. Econometrics. 2020; 8(2):21.
https://doi.org/10.3390/econometrics8020021

Unfortunately, in both cases using GUI-based tool is not possible. But
we (I mean Jacek Kwiatkowski and I) would help, if you're interested.

Marcin

--
Marcin Błażejowski
_______________________________________________
Gretl-users mailing list -- gretl-users@gretlml.univpm.it
To unsubscribe send an email to gretl-users-leave@gretlml.univpm.it
Website: https://gretlml.univpm.it/postorius/lists/gretl-users.gretlml.univpm.it/