Dude, you are carrying me through my data science MSc. Sincere gratitude for all of your sublime teaching resources. If it was common practice for schools and universities to train their staff to teach science and mathematics the way you teach it, quite literally the world would be a better place!
almost wasted so many months to find this master piece tutorials for time series. Instead of telling its difficult / complex you gave a clear idea why its important to learn timeseries. Thank you for the complete playlist.
You are the best! Thank you for making these videos and giving us such wonderful explanations. You deserve one of those RU-vid awards if you haven't already gotten one.
Holy shit dude, I've been studying simulating paths for stocks and stuff for the past month and I've been struggling to understand the point. The prediction intervals and accumulating uncertainty explanation clears up so much! I am almost in tears! thank you
Thanks to bard AI, it provided a link to your website which led me to watch your videos. Your teaching is not only has depth but also easy to understand. Such a rare combination. Please keep doing this and thank you.
Just watched your VAR video and this is the second video of yours I've watched. I gave you a like when I was just 7 secs in. This does not disappoint. Your explanation is soooooo good. Please don't stop making more videos. I just subscribed to your channel.
What an amazing way to introduce Time Series, mate! Can't wait to see the other videos in your playlist about this subject. Many thanks for that and greetings from Brazil! ;)
Thank you so much! Great video for me who has no math and data science knowledge. I finally got clear about the differences of regression and time series!
It's actually nuts how nonsensical our lecturer is trying to explain all these concepts. Been binging this channel for a bit over the past few days and everything makes so much more sense now. It's crazy how bad some people are at teaching, yet take on the job of teaching. Obviously it's nothing personal, but my god if you're gonna apply as a teacher, lecturer, whatever, at least have some basics in teaching :
It's crucial to be clear about the terminology, as the terms "regression" and "interpolation" have specific meanings and uses in statistical and mathematical contexts. If someone refers to regression as a form of interpolation, they may be emphasizing the predictive aspect of regression models within the observed data range. However, it's not a universally accepted terminology in formal statistical discussions.
Your prediction interval on the demand vs temperature graph also grew at the highs and lows for exactly the same reason - your basis for the extreme temperature demands is based on less data than near the average, say, monthly temperatures where you have tons of data.
Notes for future - TS is an extrapolation problem and error keeps on increasing as we move away from known data - Reg is an interpolation problem and error is more or less same, since prediction is usually made in the range of available data.
The Non-time series way where we find out the relation between x_i and y, the value can be interplated or extrapolated based on x_i. It need not be interpolation always. Say we have linear regression y=3x+2. We can find y for any points outside of what values we have. So it can be interpolation or extrapolation. But for time series X_i which is time now will never occur again so it will always be extrapolation.
Excellent explanation. it helped a lot. i have watched most of the videos and it is so good in theory. i would like to see if you can put some practical examples on VAR, ARCH, GARCH, of course with the help of packages like Eviews.
During muy Msc, I had one full course on financial time series... Even though they were toy examples, predictions were awful... Don't know why many grad programmes don't focus on incorporating additional predictors (within the ts domain) rather than just the lagged values themselves and their variants. For prediction purposes, I have not found ts really useful. Interesting explanation and your videos are great btw.
Excellent work.................. Request to make separate and easy videos for Machine Learning, especially real-world data, energy, water, or climate change.
This is very well done! Had never heard this interpolation vs. extrapolation distinction being articulated before. One question: for multivariant time series forecasting, if we have a data volume that is large enough for RNN, how should we think of RNN under this framework? Strictly speaking, it is an ML approach so it should be interpolating, but it is also used (when conditions allow) in some forecasting tasks. What could the best way to categorize RNN? Thank you :)
Thank god we are not yet travelling back in time, like the move ' The Adam Project'. Otherwise the time series would become an interpolation.. lol !!! Awesome video by the way
Your videos seem great, do you mind giving some insight into the ARIMAX/SARIMAX models, specially how the external regressors can influence your dependent variable of the series in question.
Good talk! So in this case, deep learning based models such as LSTM should also suffer from this extrapolation behavior. The more it predict into the future, the bigger error should be witnessed ?
That makes intuitive sense. My question is aren’t the data points on the temperature-sales graph that haven’t appeared yet also “future” data points, so they also carry the uncertainty in the time component with them? So we are still extrapolating but unlike time-series data we are not extrapolating based primarily on the time-related/varying features? But only extrapolating based on features that we may assume time doesn’t have an effect on? To me it sounded like they both have components of interpolation and extrapolation? Am I misunderstanding some concepts?
Great analysis! Yes, even a non-time-series trend has interpolation and extrapolation components. The big difference though is how often we need to perform these operations for time-series vs non-time-series data. If we think about non-time-series data, the more data points we collect, the more and more likely any future prediction will be inside the range of observed data points making it usually interpolation. This is not true for time-series data since we're mainly making predictions about the future meaning we're usually doing extrapolation.
@@ritvikmath thanks for explaining! That makes sense. I guess when we say we are usually doing interpolation on non-time-series data assumes that the variables measured aren't significantly impacted by time, or maybe time impact them equally (same magnitude, same direction) and it all cancels out -- because when we are not accounting for time as a variable when we're doing non-time-series data, the effect of time is implicit/contained in the data/observation, and we are assuming it has no effect on the (placement of the) data, when we are doing interpolation. And we know for sure for time-series data, we are definitely extrapolating because it is always beyond the data points we have because we are mapping things exactly against time. Maybe I am reading too much into this!
Could you explain stability in times series? I do understand that stationary times series have finite mean and variance and their covariance or correlation is not time dependent and as two points are far off, the covariance turns to 0. Is stability in TS actually different from stationary property?
Our project was given 5 years of annual data and we're asked to predict up to 10 years. Seeing that last part now makes me scared on why the project was approved at all.
Question: If we are forecasting tomorrow's ice cream sales using temperature as the only input, how are we not extrapolating? What's your exact definition of extrapolation here?
Hi! immensely helpful was searching for your email to put in a request. please to make a few videos explaining with practical examples of research work how the var, arima, cointegrtion, garch arch models are being used. and please put up some videos on cointegration too
Hi, Can u pls share the sequence in which we should watch these videos? It's very confusing otherwise becz we seem to be jumping between the topics in unexpected manner.
If you can give some materials of the videos, it will be better. Only videos isn't friendly to review and learning, Especially when the words of videos can't be showed and a viewer isn't native English speaker.
