This book was in my to-read list, when I encountered this video. It goes right to the heart of a problem I’m trying to solve, and I’ll pick up the book now! Thank you for making these resources available. For the record, the transparency is what helped me decide to make my purchase - thanks for sharing!

what is the coding for the plot for 2:43? I follow the code from the textbook but it is not the same

Note: there seems to be an error on slide 2 here, in the textbook the stochastic trend equation after differencing has y' on the left hand side. Hope this helps anyone confused by this!

the family monster part is very helpful, many thanks!

How to get the CSV used in this chapter? I would like to try out the technique in python.

at 7:20 when you said that the point of forecast if the mean/median of the distribution, as distribution belongs to I so I in univariant or one variable, so yt is I but in future, my question when you are dealing with multi variant analysis, does I also represents the same target feature, or it's some sort of combination between them.

This channel is heavily under-rated

Thanks guys, very useful tutorials.

is it available in python

gotta love your shirt

Wonderful book and I really appreciate the video explanations!

Muchas gracias profesor. Sus explicaciones son maravillosas

thank you.. this is a very useful lecture series

First of all thank you very much for giving all this material out for free, it's a real asset in my day to day. Now the question : at 2:34 you mention that we don't want too many observations in our training set to "have the replication" as we increase the size of the training set. I am not sure i understand that part. The only sens i could make out of it is if the has (for example) a trend change at some time t1 and we are fitting only one trend curve in our model. In this case we would have no issue until our horizon starts encompassing points after t1. First our forecast error would grow, then points after t1 would become part of the training set probably helping a bit but we would be in an underfitting situation. Keeping the same model the only way out of this situation would be to fix the number of points in our training set so that as we shift forward the time of the first point in our training set it would eventually reach t1 and forget about the trend before t1 present in the data. Is this somehow related to what you meant ? is there more to it ? i am all ears for any kind of enlightenment on this. If i wasn't clear enough please let me know and i'll do my best to rephrase. Thank you !

The videos of chapter 3 is so informative. Thank you so much

Thank you for this series, Dr. Hyndman. I didn't see what A, N, and Ad stand for in error, trend, and season functions which are used while fitting models. Could you please explain them?

In 15.56 you mentioned that one should use L-p-q degress of freedom for the Ljung-box test. On the next slide you consider the the ARIMA(3,1,0) model. But for the ljung-box test with lag 10 you use DOF=3. Should not be DOF=10-3=7? Thanks:)

I am receiving the following error when forecasting (step 7): "Error in ets(object, lambda = lambda, biasadj = biasadj, allow.multiplicative.trend = allow.multiplicative.trend, : y should be a univariate time series"

question. Wouldnt the auto arima performed on the full training set have a higher chance of overfitting, compared to when hyperparameter in combination with cross validation (for example expanding window) is performed ?

I'm extremely thankful for these videos. One detail: call me pedantic, but I think I heard you calling GWh gigawatt per hour. That's gigawatt-hour we're talking about here, right? I.e. a quantity of energy produced per quarter? Gigawatt per hour would be a rate of increase of power, which I believe is not what we are dealing with here.

Fan of you.

This content is incredible for time series education. I've always loved the written resource, but videos take it to the next level! Thank you!

This was very helpful, thank you.

please correct me if im wrong : in the grandparents cartoon explanation for autocorrelation, at lag = 2 the reason for the correlation between the grandparent and the monster being negative is being said that they tend to be 'very different'. Actually negative correlation, just means they are related but in a negative sense, if one rises the other falls and vice versa, they are not very different. They are completely different if the correlation is 0.

wouldnt the seasonality be in the Dember momnths as there is wwhen more sales are made and would need more people, and you can see every beggining of a decade pointint at a December month there is the spike

Thank you very much for the incredible material freely available. Can you help me, please. I try to build the multi time series graphs but I can't. I am interested in building the graph at min 5:00. Thank you!

Thanks a lot for this video. This is my very first time seeing the data pipeline operator |>. I tried using that operator (without loading tidyverse and fpp3) in my R but it didn't work. Is this operator only works after we load the package fpp3 or tidyverse?

so grateful for the videos and the book!! they are EXTREAMLY helpful!!

youre a life saver

In additional decomposition why should the seasonal indices add up to 0? And in multiplicative it should add up to m.? In additional the seasonal indices are positive and negative Yt= Tt + St + Rt If sum of St doesn't add up to 0, then St would have a net positive or negative effect in residual. In multiplicative decomposition Yt = Tt*St*Rt Rt = Yt/(Tt*St) Rt = (Yt/Tt)/m This would divide the residuals amongst the season periods according to indices. Is this explanation correct??

Can you share how we can forecast gear box bearing failure

Thank you for that video. Is there a possibility to set the initial value for l[0] manually?

Why you say Moving Average? Isn't it LOESS? Anyway, thanks for your work!

Fantastic course! Thanks for sharing your knowledge, professor!

Thank you!

Thank you so much!

Thank you!

Thank you!

Does anyone have it in digital format that they can share?

Great material!

Please, do more lessons about binary options

Great video! Is there an option to get a report table with coefficients for each part of the piecewise model?

This is great but it is hard to distinguish parts of the code especially the brackets. Using a larger fonts and a drak background might help. PS. - I am using the highest resolution.

Dr. Hyndman - I came across some unusual date format in the sulfur dataset of volcano (tidytuesday data) that are in "decimal years". For example the first entry is 705.958. I take the first part to be the year 705 CE (common era). But it is not clear to me what the 958 is. Do you simply mutate these and round them? In addition in other parts of the data the dates are negative. Again another issue. I know you mostly work with modern day data but some of us (geologists) work with ancient dates. The -1000 I presume refers to 1000 BCE (before common era). Any thoughts or suggestions on how to handle these "digital" years and or negative ones?

Good stuff!

Dr. Hyndman - when dealing with massive datasets it may be difficult to tell what the appropriate time unit is. Have you considered generating a function that can help determine that?

Are your slides Beamer?

Very good explanations. By the way I downloaded the quarto template for Monasch thesis. I would love to adapted to my own department/university. Thanks for contributing that.

I have always wondered: What is ... the point? I mean, let's say you recognize that your trend is caused by changes in q4. Does this realization have any impact on your process of choosing a good forecasting model, like in the case of Arima where you are trying to minimize either the error on your validation set, or some information criteria like AICc? What has plotting and understanding the data actually achieved for you, when the model selection process is seemingly presented as some sort of grid search you do?

God Bless ya prof Rob I'm a huge fan of the book !