Feature Engineering for Time Series Forecasting - Kishan Manani

[@iftikhar58]

It was a great talk about data. Thank you so much. I hope you can share similar talks on the future as well

[@ivanliu1173]

Thanks for this informative video! 👏👏👏

[@jossec1344]

Magnificent work Bravo!!!

[@dariozoric7181]

Great talk!

[@distinctga5811]

What if you have 2 products that have the same target encoding value but different seasonal patterns? The algorithm encodes them to the same class due to similar averages without considering variance differences. Would adding another target encoded feature using average absolute STD work or should the variance be captured with lagged values?

[@offeibekoe452]

I also had the same question

[@MinhVu-ym4tk]

good to know :D I am working on RUL estimating and prognosis using time series data.

[@onurkoc6869]

You are Hero thanks:))

[@piotrbjastrzebski]

It is great but something is wrong with time_col in definition of the procedure. It seems to work if that column is an index and not mentioned in a function call.

[@AhmedThahir2002]

Hi, does anyone know how to implement the recursive forecasting that he did in Darts using sktime. I couldn't really find an intuitive explanation online.

[@gurjinderkaur5007]

In target encoding section, when product ID is encoded dynamically, how will the model distinguish between the data points belonging to same time series or different time series?

[@b1ueocean]

What tools are folks using to expose/extract/generate features? Tsfresh? getML? I work in Java for my ML tasks but will happily integreate Python or C/C++ based tools into the pipeline. I'm not a statistics guy so I can't write these feature generation algos myself.

[@jacobschultz3168]

Great presentation. To clarify, is overfitting always an issue? I'm assuming it always is. In the scenario where you compute the window values, ensuring you're only using the available data... there will be no leakage at a row-level. But when you consider all training values.. for example at Time = 1 vs Time = 8, the relationships being built by the Forecasting algorithm when predicting Time = 1 will still use Time = 8 values.

[@anoubhav]

For two different time series, does it make sense to build two separate models instead of having the targets of both the series in the single model (as shown at 24:40)?

[@TraininData]

You can. It depends on the data and the situation. When data is small, forecasting multiple time series with a single model, allows the model to learn and use patters shared across time series, with usually improves performance. Deploying and maintaining a model is also easier than deploying and maintaining 2, or more likely several. But if time series patters are very different, it might be better to have a model per time series. Hope this helps.

[@pranavkhatri9564]

can you explain something about stock prediction?

[@RDarrylR]

What is the name/link of the “chunky” review paper you mentioned at the end of the presentation?

[@kishanmanani1466]

The paper was indeed in the references slide. It is: Petropoulos, Fotios, Daniele Apiletti, Vassilios Assimakopoulos, Mohamed Zied Babai, Devon K. Barrow, Souhaib Ben Taieb, Christoph Bergmeir et al. "Forecasting: theory and practice." International Journal of Forecasting (2022). It's also free to access online.

[@RDarrylR]

@@kishanmanani1466 Thanks! I must have been looking in the wrong place!

[@oneforallah]

@@kishanmanani1466 Thanks !

[@AhmedThahir2002]

Hi @@kishanmanani1466 , it was a lovely talk. I was wondering if you could point me in the direction of how to implement the recursive forecasting that you in Darts using sktime. I couldn't really find an intuitive explanation online.

[@mamyrak1114]

can someone help me to deal with categorical features for forecasting time series in ML