Hey, I've been following almost all of your videos. Great explanations indeed. Haven't found a better channel on youtube for time series explanations. Could you point me as well to some online pdfs/refs where topics like time series can be read through ?
Hey Ritvik, it really a great videos about time series, I have a question after we do all the transformations to get a stationary series and calculate a PACF & ACF to get ARMA model p and q, then I think we can apply it direct on the original series to fit the model and get the predictions. Thank you so much
If my time series curve is way more complicated than a log curve (like yours), how should I decide on the number 2 (after normalize), which in you case is exponentiate? Edit: how can I use Fourier Transform to figure out the transformations I need to make. My time series curve, for example, doesn't look like any obvious mathematical curve like e^x or log(x) or 1/x, etc.
Thanks for the awesome series, I’m wondering though how you’d reverse a volatility removal, since you potentially don’t have future data you can’t calculate the stdDev. Would you use something like GARCH or assume the variance is kind of the same as in the last period? The example from the last video would be perfect there, like how do you reverse the volatility removal for the ice cream - heater example?
This is perhaps not the best dataset for timeseries forecasting. With the previous examples such as catfish sales or icecream sales, those data points (months) are independent of each other and theoretically "restart" at 0 every month. Views on a youtube video is more cumulative - if you have 1000 on the first day, then 1100 on the next day, it's not a "net new" 1100, it's a net new of only 100. You'd want to predict on the "delta" views (almost like "returns" on a stock) in this case, in my opinion. I get that's why we do the exponentiate and first difference steps but it just seems like the original timeseries is too predictable already and we're making extra work for data that we can just "extend the curve"... but I do appreciate the mathematical concepts at play here. Thank you
wouldn't it be better to just curve_fit to a log with some params? Or is this more of a general demo vs the right thing to do w.r.t. that sort of time-series shape?
Hi Ritvik, thanks for the video! Is it correct that when forecasting on the transformed data, back-transforming might not preserve MSE-optimality? Would you consider making a video explaining why and what we can do about it?
How does the first command at 2:50 in cell [5] work? I assume you use the `mean` method of numpy on a series of datatype pd.Seris() named `ts`; This results in a _scalar_ of datatype: float Floats do not have a .iloc[] operation... What is going on here?
![PURPEECH - นี่ฉันเองคนที่... (It's me) [Official MV]](http://i.ytimg.com/vi/NuhHd0hOOg8/mqdefault.jpg)
Thank You so much , very well explained. Indeed, much needed reverse transformation specially for algorithms like VAR.
You are most welcome!
Hey, I've been following almost all of your videos. Great explanations indeed. Haven't found a better channel on youtube for time series explanations.
Could you point me as well to some online pdfs/refs where topics like time series can be read through ?
Hi! would you please conduct an online course on Time series analysis, model fitting and allied concepts. Would love to attend if done.
Thank you for the detailed example! This is great!
Very well explained. Thanks!
You're welcome!
Thanks a lot, this video is really helpful!
Glad it was helpful!
Hey Ritvik, it really a great videos about time series, I have a question after we do all the transformations to get a stationary series and calculate a PACF & ACF to get ARMA model p and q, then I think we can apply it direct on the original series to fit the model and get the predictions. Thank you so much
absolutely brilliant!
very informative, thank you!
Glad it was helpful!
If my time series curve is way more complicated than a log curve (like yours), how should I decide on the number 2 (after normalize), which in you case is exponentiate?
Edit: how can I use Fourier Transform to figure out the transformations I need to make. My time series curve, for example, doesn't look like any obvious mathematical curve like e^x or log(x) or 1/x, etc.
Thanks for the awesome series, I’m wondering though how you’d reverse a volatility removal, since you potentially don’t have future data you can’t calculate the stdDev. Would you use something like GARCH or assume the variance is kind of the same as in the last period? The example from the last video would be perfect there, like how do you reverse the volatility removal for the ice cream - heater example?
This is perhaps not the best dataset for timeseries forecasting. With the previous examples such as catfish sales or icecream sales, those data points (months) are independent of each other and theoretically "restart" at 0 every month. Views on a youtube video is more cumulative - if you have 1000 on the first day, then 1100 on the next day, it's not a "net new" 1100, it's a net new of only 100. You'd want to predict on the "delta" views (almost like "returns" on a stock) in this case, in my opinion. I get that's why we do the exponentiate and first difference steps but it just seems like the original timeseries is too predictable already and we're making extra work for data that we can just "extend the curve"... but I do appreciate the mathematical concepts at play here. Thank you
wouldn't it be better to just curve_fit to a log with some params? Or is this more of a general demo vs the right thing to do w.r.t. that sort of time-series shape?
Very nice. Thank you.
Hi Ritvik, thanks for the video! Is it correct that when forecasting on the transformed data, back-transforming might not preserve MSE-optimality? Would you consider making a video explaining why and what we can do about it?
Thanks ;)
You're welcome!
Hi thanks for the video, are you planning to make a video about singular spectrum analysis sometime?
Great video. Unfortunately, i cant get to work due to updates to statsmodels
can you please contact me I had some questions about the video you made about the blackjack.
Beautiful
great!
How does the first command at 2:50 in cell [5] work?
I assume you use the `mean` method of numpy on a series of datatype pd.Seris() named `ts`;
This results in a _scalar_ of datatype: float
Floats do not have a .iloc[] operation...
What is going on here?
What to do if no lag in pacf and acf are significant. what is the parameter we fill in ARMA model.
thanks, I wanted to know how did you calculate the lower and upper limit ?
May I have the predicted value?
Hey Ritvik! Can you share the notebook?