TS-1: Curve fitting is (almost) all you need

[@abhishekkrthakur]

Notebook: www.kaggle.com/konradb/ts-1-curves

[@HoneyofKnowledge]

Timestamps Start : 2:29 Importing Libraries : 3:45 Defn. Of The Series : 5:29 Example Time Series : 10:24 Plot : 13:12 Seasonal Decomposition : 17:05 About Residuals : 18:36 Probably Multiplicative Series : 19:11 Multiplicative Decomposition : 20:49 ACF & PACF : 27:29 Exponential Smoothing : 32:41 Next Example of Time Series : 43:06 Prophet Starts : 43:46 About the Data : 44:32 Prophet Forecasts : 56:56 Covariates : 1:20:47 Neural Prophet 1:26:04

[@shivamkaushik6637]

Konrad is the kind of professor we need, not the one we deserve!. Truly enjoyed this. Thanks Abhishek and Konrad!

[@lalitmrinki]

Very Informative QA. Notebook demo was also excellent.

[@ranjansingh9972]

This was fantastic. Great learning from a master of the topic. Time Series is such a huge and complicated topic it’s extremely rare to get such a fantastic survey across the field. Can’t wait for the next session.

[@TusharKale9]

This is very useful video to learn the evergreen Time Series Analysis. Thank you for sharing

[@channelname9332]

thanks. it was great session. looking forward for next. if its possible please add fourier transformation concept. it is bit confusing how to use and interpret it.

[@virajkadam3017]

Thanks a lot Abhishek and Conrad, this series is really fun and useful!

[@richardpinter9218]

Great session, thank you both.

[@vermichel5021]

Please tell how you choose a suitable transformer for NLP tasks. I know that e.g. for text generation we use decoders but there are many and how to choose decoder that suits the best?

[@harisumanth]

Thanks for the session.

[@HoneyofKnowledge]

At 7:52 Konrad Says, Deterministic Trend. Trend can also be stochastic in nature

[@luffyd2126]

Really good stuff, thanks so much

[@ВадимШатов-з2й]

It is very useful. Abhishek, please tell me, do you have a video on the analysis of an anonymized dataset? There is nothing useful on the Internet and on KZbin about this topic.

[@simymathew7887]

Thanks for the video.. Can you please explain how to handle multiple time series data with exponential smoothing at once.

[@rohitkrsingh]

Use freq instead of period to run seasonal_decompose function.

[@the_cosmic_phakeer]

16:58 probably the data is for every 30 days more exactly, that is why it is 1973-01-31 & 1973-03-02 etc

[@bhavinmoriya9216]

Hi, Thank you very much for a wonderful lecture. You selected period=12, because you thought every 12 months data would repeat itself?

[@konradbanachewicz8641]

Yes - if we postulate annual seasonal pattern, that's what we expect.

[@bhavinmoriya9216]

@@konradbanachewicz8641 Thanks very much:)

[@RaDha0102]

Hi Abhishek & Konrad: I deal with lot of Time series data in my day-day job. My challenge would be like, say for example i have to forecast the all the listed stocks in a exchange. Please say how as a Data Scientist am i supposed to approach the solution to this problem? (I usually treat it as a regression problem with snapshots of data with different lags, also add external variables to predict my future value(s)). Should i build "n" models, one for each stock (where model maintenance is a problem) OR one model for the whole data (afraid of each series generalization here)? is there a correct way to approach this problem Or any alternative methods that are followed across industry??

[@konradbanachewicz8641]

Hi Rajiv 1. what do you want to forecast: the raw prices or returns 2. how long ahead into the future 3. at what frequency

[@RaDha0102]

@@konradbanachewicz8641 1. Raw Prices (closing price) 2. assume 5 days ahead. 3. One prediction per day per stock. (Note: I've mentioned stocks as example because its easier to explain multiple time series forecasts problem. But stocks are not exactly i deal with.)

[@Kaushikcolab]

Hey Did you figure out the approach to deal with multiple series?

[@econhelp583]

Time series data should be handled with caution and by a knowledgeable practitioner. How well a time series model performs and how risky it is to use time series methods will depend on the field of study. A TS model may produce very good results in a natural science problem but disastrous results in a finance problem. It all boils down to a basic understanding of statistics. If your data is not being generated by a controlled experiment then be highly cautious in terms of causality and forecasting.

[@ManojKumarmamtha]

Thanks sir

[@bhavinmoriya9216]

How do you decide mcmc_samples = 10 in Prophet?

[@konradbanachewicz8641]

Just a small number to demonstrate how the functionality works, but not spend too much time on it during the live session.

[@bhavinmoriya9216]

@@konradbanachewicz8641 Thank you very much for prompt reply. So it is more of a trial and error? What is the purpose of this parameter?

[@konradbanachewicz8641]

@@bhavinmoriya9216 the more MCMC samples, the more reliable your estimate of the uncertainty band around your model components.

[@bhavinmoriya9216]

@@konradbanachewicz8641 Só overdoing it might overfit and hence to get the perfect balance we have to tradeoff performance vs mcmc? Am I getting it right?

[@konradbanachewicz8641]

@@bhavinmoriya9216 no, using too many samples won't overfit - it will just take a really long time and above a certain level won't bring any meaningful improvement. sort of like number of trees in a random forest.

[@TheEngomusic]

Where have you been all my life guys