Fyi there’s a tool called the BEAR toolbox in ECB provides all of this including all different kinds of prior distribution which is very easy to implement.
Hi Ritvik, Can you help me to solve the below: I am working with electricity time-series data collected at 15 minutes intervals. I am looking for a procedure/theory to find the pattern/sequence in the time-series data based on given features. As I am working with electricity time-series data and solving the problem of solar PV identification from these data, the given features would be: 1. There is a fall in electricity consumption during 7am-8am as generation from the PV starts. 2. There is a rise in electricity consumption during 5pm-6pm as generation from the PV ends. I have gone through the literature for the same. I got the following: Gaussian prior: This works with considering the prior knowledge and evidence. In this case, the prior knowledge would be above two features, and the evidence would be the time-series data. Cross-correlation: This basically looks for the relation between two patterns. Various ML techniques: The different ML techniques can be applied such as clustering, HMM, DTW etc. I am not looking to solve this problem with option 3. Can anyone guide me with option 1 as it looks more relevant to my problem. I cannot understand how Gaussian prior can fit into the problem. Summarily, I want to utilize the above two features as the prior knowledge and use the given data(electricity time-series) as the evidence to prove whether the solar PV panel is present or not.
If we can estimate the maximum likelihood, then why should we do so much computation to sample from our posterior distribution? Like take for instance time-series prediction, we want point estimates at the end of the day
Hi Ritvik, first of all, thank you so much for producing such explanatory content on Time Series. I'd further like to study about time-series. Can you suggest any good content to refer to?
Amazing video Ritvik! Thank you for posting. Really exciting to see some good alternatives to frequentist time series methods like ARIMA. May I ask, are there any good rules of thumb to follow when predicting percentages in time series? For example, if I wanted to try to build a monthly employee attrition rate forecast for different departments in an organization (where rates can be anywhere from 0%-100%). Would I need to handle my data differently in the preprocessing phase or perform any data transformation? Or would the process be the same as if I were working with natural numbers?
Hi Ritvik, I have a question. Let's say if I want to perform risk control for financial time series so that I need to know the maximum loss and the possibility in tomorrow trading. Can this be achieved by using the Bayesian method, obtaining distribution for different parameters, then applying linear transformation on their 'tails' of distributions?
At 5:22 you are picking phi1_val, phi2_val and sigma_val independently. Shouldn't you pick an index and then pick the three parameter values corresponding to that index?
Ritvik, have had a chance to update this using PyMC v5? and also the new versions of StatsModels.tsa.arima? There are some massive differences and your code no longer works. Thank you.
Very nice presentation. I like how you show at the end what Bayesian methods are adding in term of information.
Dude. This should have been uploaded 2 weeks ago before my assignment submission 🥲
Ritvik, great videos as always!
Thanks Gary!
I love your videos. Tks for teaching so clear with examples! Hugs from Brazil/Rio.
Time Series Talks are the best!
Anything on Bellman equations and optimization in the works?
Thank you for the content you have been producing. A small request: Please make a video on state space modelling if possible. Thank you in advance!
+ one on the state space!
This was epic!!!
Truly amazing series.
nice intro video to forecasting
Thank you!
Great presentation!
thx hope to see more bayes stuff on your channel
Thanks !
Love you
Another great vid on Bayesian methods and pymc3, thanks!
Of course! Thanks for watching!!
excellent tutorial!
Thanks!
Hi Ritvik, Can you please do a video on Bayesian Structural Time Series (referring to Causality)?
ty so much
Fyi there’s a tool called the BEAR toolbox in ECB provides all of this including all different kinds of prior distribution which is very easy to implement.
Great work. Thank you.
perfect videos, thank you so much!!!
Hi Ritvik,
Can you help me to solve the below:
I am working with electricity time-series data collected at 15 minutes intervals. I am looking for a procedure/theory to find the pattern/sequence in the time-series data based on given features. As I am working with electricity time-series data and solving the problem of solar PV identification from these data, the given features would be:
1. There is a fall in electricity consumption during 7am-8am as generation from the PV starts.
2. There is a rise in electricity consumption during 5pm-6pm as generation from the PV ends.
I have gone through the literature for the same. I got the following:
Gaussian prior: This works with considering the prior knowledge and evidence. In this case, the prior knowledge would be above two features, and the evidence would be the time-series data.
Cross-correlation: This basically looks for the relation between two patterns.
Various ML techniques: The different ML techniques can be applied such as clustering, HMM, DTW etc.
I am not looking to solve this problem with option 3. Can anyone guide me with option 1 as it looks more relevant to my problem. I cannot understand how Gaussian prior can fit into the problem. Summarily, I want to utilize the above two features as the prior knowledge and use the given data(electricity time-series) as the evidence to prove whether the solar PV panel is present or not.
If we can estimate the maximum likelihood, then why should we do so much computation to sample from our posterior distribution? Like take for instance time-series prediction, we want point estimates at the end of the day
How do you predict time series with MA terms and seasonal components?
Hi Ritvik, first of all, thank you so much for producing such explanatory content on Time Series. I'd further like to study about time-series. Can you suggest any good content to refer to?
Amazing video Ritvik! Thank you for posting. Really exciting to see some good alternatives to frequentist time series methods like ARIMA.
May I ask, are there any good rules of thumb to follow when predicting percentages in time series? For example, if I wanted to try to build a monthly employee attrition rate forecast for different departments in an organization (where rates can be anywhere from 0%-100%). Would I need to handle my data differently in the preprocessing phase or perform any data transformation? Or would the process be the same as if I were working with natural numbers?
Hi Ritvik, I have a question.
Let's say if I want to perform risk control for financial time series so that I need to know the maximum loss and the possibility in tomorrow trading.
Can this be achieved by using the Bayesian method, obtaining distribution for different parameters,
then applying linear transformation on their 'tails' of distributions?
At 5:22 you are picking phi1_val, phi2_val and sigma_val independently. Shouldn't you pick an index and then pick the three parameter values corresponding to that index?
Great intro, however most of the code is not working with PyMC v5!
Ritvik, have had a chance to update this using PyMC v5? and also the new versions of StatsModels.tsa.arima? There are some massive differences and your code no longer works. Thank you.