I can no longer send personal messages on Bluesky because of the idiotic British Online Safety act... I'd be really grateful if you could send me an email on ivan at svetunkov.com.

I'd be curious to know what you think about it.
If I ever get a free copy of this, I'd read it to make my own impression of it. But I don't want to give him my money after all the stuff he told me on social media 😄

Having communicated with the author on LinkedIn over the last couple of years, I'd be curious but cautious about this book. He's been spotted making ludicrous claims and silly mistakes in the area of forecasting without any self reflection. So, I'd be critical when reading this. But this is IMHO.

Thank you for the kind words!

The post currently has 981 likes, 154 comments and 46 reposts. This tells you more than enough about Valeriy and what people really think about him.

#machinelearning #datascience #forecasting

Thanks for promoting my paper, @econmaett.bsky.social !

In the short, we show in the paper that the direct strategy implies shrinkage of parameters of dynamic models (such as ARIMA and ETS), which could be beneficial under some circumstances.

Ah! That one! Yes, that's a good idea. I don't see how this can be done for demand forecasting for companies because the data is typically under NDA, and there's no way to check the results independently. But for the economic forecasting, that's a good idea. 👍

I personally prefer the lines, because I guess I'm more used to that :). But this could work for some, why not?

😆

IMHO, even for prediction stepwise is not the best option...

But practically speaking if you want one thing, you go one way, if you want another, do something else (e.g. decision trees as mentioned by @tiddlydump.bsky.social below). But my original question was more about the "true" model.

I agree on that. And this is a very practical point. Mine was philosophical :)

This comes to a great philosophical question, which I have not personally resolved for myself entirely: should the models for prediction and effects evaluation be fundamentally different or not? I'm more inclined towards "they should be the same", but I don't have strong view on that.

My post was a bit provocative for a reason 😀, just to raise attention of people to the more modern approaches. There's nothing bad about the Holt-Winters method, BTW. It's just it is now encompassed by a bigger and more solid model.

I haven't, but I thought that it has been previously explored in several papers. Giacomo is one of the academics that did that exploration. I'd suggest getting in touch with him :).