18:46 State Space Models 25:50 Structural Time Series 30:46 Kalman Filter 38:40 Implementing Structural Time Series 1:04:00 Hidden Markov Models (HMMs) 1:11:00 Baum-Welch and Viterbi Algorithms 1:20:08 Implementing Gaussian HMM 1:40:00 Machiene Learning for Time Series 1:57:40 Implementation ML for TS 2:44:19 Deep Learning for Time Series 2:47:20 Recurrent Neural Networks (RNNs) 2:51:33 Convolutional Neural Networks (CNNs) 2:56:35 Implementing Deep Learning
Ahhh such a polite teacher and the way she talk abd explain. OMG she and people like her are really a gift to our society. Stay safe, keep teaching and keep smiling. thank you
Have to agree with everyone on here. Excellent lecture - a great mix of detail and higher-level overview. It sounds like this isn't even her full-time gig. Impressive. My new learning strategy - watch every one of her you-tube tutorials.
Wow..took sometime to complete it...but this is best explanation for time series so far..although it tells me to learn more about these things ....one should be very much familiar with the numpy to code these things
This is an amazing presentation on many levels! Perhaps a very odd question, but would anyone be able to explain how to establish a presentation setup as shown here with the the speaker on camera and the code window in full display?
Great talk. What is being forecasted at 2':15" using XGBoost? It seems like she is not using the time series values at all for regression. What is the target value in the training?
I am a little confused about the feature generation in the ML forecasting part. It seems like we're spending a lot of effort to create features that end up not being very predictive of the target. Couldn't we use use the lag values themselves as features in the model? Xgboost (or even a simple linear regression) should be able to detect the correlations and provide a decent prediction.
Hello, everybody I type gcag_mod=sm.tsa.UnobservedComponents(train['GCAG'], **model) gcag_res=gcag_mod.fit() then I got name 'train' is not defined. Could anybody help me?
NOTE: 1:03:36 MAE calculation should be a subtraction of the fitted and the training data, not a concatenation with a comma! Ends up being like 0.072191.....
I'm using RNN for my PG thesis work. I've a query. Do we have to run stationarity test for our time series data before feeding it in the neural network model... or this step is only required in traditional time series models like ARIMA?
At 1:03:21 when Aileen speaks about the mean absolute error, the code in Cell 47 is wrong: instead of a negative sign, there is a comma, and this is still present in the github repo as of this writing.
@@joaoantonio9337 Hi, you can read it behind her on the white board, but she also mentions it at 37:42. More specifically, here it is: github.com/theJollySin/scipy_con_2019/tree/master/modern_time_series_analysis
very god tutorial - How to start learning time series from scratch and fourier analysis for stock market time cycles? Or any good books Or Courses to study?
Surely an LSTM (or any recurrent neural net) is an machine learning model setup/designed for time series? Also random forests do give feature importances (like XGBoost). Still enjoyed the talk :)
Great video with a lot of depth. Knowledgeable speaker. Thanks for creating and sharing. I have a couple of "concerns": * Why does none of the data have measurement units? It is almost as if statisticians do not care what they analyze (it is just numbers). Look at the charts, there is no measurement unit on the x-axis (looks like it is mostly months) and no measurement unit on the Y-axis (what does it mean that the global temperature is varying between -0.75 and 1.25 but what is it? Apples? Oranges? Degree F? Degree C? Kelvin? Is it absolute? Is it a delta from a base measurement (relative)? Where was the measurements taken? I am concerned about this, as a measurement unit is one of the most basic contextualization elements. My middle school math teacher would mark answers as wrong if the measurement unit was missing. * My not so humble opinion is that dynamic time warping is bullshit. There are so many issues with the approach. The presenter is taking two sinewaves and merging them together to get a correlation and then use the result to show that there is correlation. This is the definition of circular argumentation. Another issue is that there is an assumption that the correlation is positive, what if a lower value in variable 1 caused a higher value in variable 2, then the whole error function would fail. An no point does the presenter show the warped result. This completely messes with the notion that time series generally deals with ordered continous data. It would be much better to take a fourier transform and look at the harmonics of the frequencies.
she did say that at the beginning of the talk - but what is old may still be some of the best tools for the job today - her words again... because changes in infra, data and tools have allowed better results from long standing concepts and approaches...
I would like to know if there is a better talk/repo/book any kind of resource you can suggest. (not trying to defend the video here, I want to look into the new stuff)
18:46 State Space Models
25:50 Structural Time Series
30:46 Kalman Filter
38:40 Implementing Structural Time Series
1:04:00 Hidden Markov Models (HMMs)
1:11:00 Baum-Welch and Viterbi Algorithms
1:20:08 Implementing Gaussian HMM
1:40:00 Machiene Learning for Time Series
1:57:40 Implementation ML for TS
2:44:19 Deep Learning for Time Series
2:47:20 Recurrent Neural Networks (RNNs)
2:51:33 Convolutional Neural Networks (CNNs)
2:56:35 Implementing Deep Learning
1:05:11 hidden markov model
1:40:00 machine learning for time series
2:44:00 deep learning for time series
This comment was really helpful 🙏🏼
a THREE HOUR lecture on Time Series Analysis. What a gift!
Very VERY good explanation of the different approaches to time series analysis. Thanks a lot!
An absolute delight of tutorial. Many thanks for preparing it and communicating it so well!
Best explanation of time series analysis I've ever seen. Very good mix of intro to the models, examples, and links to more in-depth information.
Very well organized, informative, thorough, and polished. All-around impressed with Ms. Nielsen.
Such an excellent video. Took me ages to finish but still, wish it was longer.
Ahhh such a polite teacher and the way she talk abd explain. OMG she and people like her are really a gift to our society. Stay safe, keep teaching and keep smiling. thank you
Best time series talk i have ever watched.
Super cool presentation ! Thanks a lot
Have to agree with everyone on here. Excellent lecture - a great mix of detail and higher-level overview. It sounds like this isn't even her full-time gig. Impressive. My new learning strategy - watch every one of her you-tube tutorials.
We think alike
perfect teaching, It was very informative. Thank you
Outstanding! Congratulations and thank you!
Truly phenomenal.
Great tutorial! Thanks!
Wow..took sometime to complete it...but this is best explanation for time series so far..although it tells me to learn more about these things ....one should be very much familiar with the numpy to code these things
Excellent and in detail explanation.
Cogent and useful well done.
This is an excellent tutorial and I like the fact that Aileen didn't skip the math part of the algorithms
Nice talk and topic cover!
Great talk, ty.
should leave the link for the lecture she mentions in the description. great material
This is an amazing presentation on many levels! Perhaps a very odd question, but would anyone be able to explain how to establish a presentation setup as shown here with the the speaker on camera and the code window in full display?
She is outstanding!
Great talk. What is being forecasted at 2':15" using XGBoost? It seems like she is not using the time series values at all for regression. What is the target value in the training?
Very eloquent and dominant speaker
Is there anything about change point detection in the lecture?
Most important video on youtube
I dont see any slides in the mentioned website or link . can some one help me to get the link ?
What to do before applying np.log if our data has zero values? What's the best technique? I added +.000000001 to all values? is that correct?
Well done!
I am a little confused about the feature generation in the ML forecasting part. It seems like we're spending a lot of effort to create features that end up not being very predictive of the target. Couldn't we use use the lag values themselves as features in the model? Xgboost (or even a simple linear regression) should be able to detect the correlations and provide a decent prediction.
Hello, everybody
I type
gcag_mod=sm.tsa.UnobservedComponents(train['GCAG'], **model)
gcag_res=gcag_mod.fit()
then I got
name 'train' is not defined. Could anybody help me?
NOTE: 1:03:36 MAE calculation should be a subtraction of the fitted and the training data, not a concatenation with a comma! Ends up being like 0.072191.....
I'm using RNN for my PG thesis work. I've a query. Do we have to run stationarity test for our time series data before feeding it in the neural network model... or this step is only required in traditional time series models like ARIMA?
Really informative. Thanks a lot.
Amazing talk! Where could I fin the github?
Amazing
Thank you.
37:50 The python programming starts
Coding begins again for hidden markov models at around 1:20:00
Notebook #3 at 1:58:00
I'm in love
Where can i get the notebook? Can someone link me to it?
Great time series talk! Thanks, the speaker speak really fast :P
Great talk! Keep in mind that many of the things that are said to be computationally taxing are only so if one implements them in Python.
god, you are amazing
is the ppt available for this presentation ?
At 1:03:21 when Aileen speaks about the mean absolute error, the code in Cell 47 is wrong: instead of a negative sign, there is a comma, and this is still present in the github repo as of this writing.
Hello Sam!
Where did you find the github?
@@joaoantonio9337 Hi, you can read it behind her on the white board, but she also mentions it at 37:42. More specifically, here it is: github.com/theJollySin/scipy_con_2019/tree/master/modern_time_series_analysis
@@samm9840 thank you!
@@samm9840 Thanks....I was searching this url too
very god tutorial - How to start learning time series from scratch and fourier analysis for stock market time cycles?
Or any good books Or Courses to study?
Would recommend digital signal processing by Alan, Oppenheimer
Can someone help me to find the notebooks???
Github link - github.com/theJollySin/scipy_con_2019/blob/master/modern_time_series_analysis/README.md
Surely an LSTM (or any recurrent neural net) is an machine learning model setup/designed for time series? Also random forests do give feature importances (like XGBoost).
Still enjoyed the talk :)
fbprophet too
Great video with a lot of depth. Knowledgeable speaker.
Thanks for creating and sharing. I have a couple of "concerns":
* Why does none of the data have measurement units? It is almost as if statisticians do not care what they analyze (it is just numbers). Look at the charts, there is no measurement unit on the x-axis (looks like it is mostly months) and no measurement unit on the Y-axis (what does it mean that the global temperature is varying between -0.75 and 1.25 but what is it? Apples? Oranges? Degree F? Degree C? Kelvin? Is it absolute? Is it a delta from a base measurement (relative)? Where was the measurements taken? I am concerned about this, as a measurement unit is one of the most basic contextualization elements. My middle school math teacher would mark answers as wrong if the measurement unit was missing.
* My not so humble opinion is that dynamic time warping is bullshit. There are so many issues with the approach. The presenter is taking two sinewaves and merging them together to get a correlation and then use the result to show that there is correlation. This is the definition of circular argumentation. Another issue is that there is an assumption that the correlation is positive, what if a lower value in variable 1 caused a higher value in variable 2, then the whole error function would fail. An no point does the presenter show the warped result. This completely messes with the notion that time series generally deals with ordered continous data. It would be much better to take a fourier transform and look at the harmonics of the frequencies.
how can I join the slack channel, please
if this code is already pushed on git..could you provide the github link to your code?
In case you haven't found it already, here's the link - github.com/theJollySin/scipy_con_2019/blob/master/modern_time_series_analysis/README.md
@@gagandeep4850 I struggled to get it from what was written on the white-board behind her, but then you already posted it. Thanks!
@@gagandeep4850 Thanks
@@gagandeep4850 can you send me your mail. Thank you very much
nice. any code available?
When you drink coffee while talking, you get slime in your throat.
1:03
52:27
2:17:20
1:40:00 czcams.com/video/v5ijNXvlC5A/video.html - machine learning for time series
You are just beautiful!
Did she air quotes global warming lol. Why the air quotes hahah
Popo
This is very old stuff...
she did say that at the beginning of the talk - but what is old may still be some of the best tools for the job today - her words again... because changes in infra, data and tools have allowed better results from long standing concepts and approaches...
I would like to know if there is a better talk/repo/book any kind of resource you can suggest. (not trying to defend the video here, I want to look into the new stuff)
the armpits
2:01:17