congrats :tada: Nixtla team. Massive win.
# general
v
congrats 🎉 Nixtla team. Massive win.
❤️ 4
🥳 2
m
Thanks @Valeriy!!!!
🙌 1
m
Hi, this voting introduced nixtla to me so thanks for that. Really appreciate the nice guides on the website which are very helpful for beginners as me. Just wanted to point out that in the statsforecast readme the "Follow this end-to-end walkthrough for best practices." link is broken and leads to 404 file not found. In my work I should predict daily ecom website revenues for multiple markets which has a clear yearly and weekly seasonality as well as heavily dependent on exogenous variables such promotions as well as recent trend. Would you recommend me to start with deseasonalizing and detrending the target variable and try ARIMA first or would it be better to go to mlforecast directly and focus on the features?
m
Pinging @José Morales for the 404. @Miro Lavi : maybe You should try using an MSTL model with exogenous variables.
👍 1
t
Hello @Max (Nixtla) . Does MSTL support exogeneous variable? So far I thought that only Arima is supporting it in statforecast? Am I wrong? I notice a demand for evolution to include it for CES in your github. Thank you for the clarification .
m
Hi, I tried with MSTL having seasonality 364, 30, and 7, which captured the overall seasonality quite well, but performed very poorly with promotion shifts. For example, this year black friday promo started two days earlier leading to 20x sales vs non promo day last year. Even when adding hotcoded exogenous variable indicating promotion type, the model more or less predicted similar pattern as last year. Maybe I should explore mlforecast lightgbm etc. if it could capture the promotion impact better. Also, if someone has good tips for beginner would be more than glad to hear :)