I love how you revisit and review things from the previous videos of the series, it helps me further deepen my understanding on the concepts! Great work! Keep up!
Thank you for the explanation! In the video you mentioned that the ARMA(4,1) model shouldn't contain the constant due to it being p < 0.05. Does the model.predict() automatically take this into account or do we have to specify this before using the .predict() function?
Just minus the constant coef value (37.3039) from the prediction series that'll do. Notice that won't change the result much since the constant is fairly small to start with.
This is great, I had a lot of suspicions about the statsmodel coefficient table. This is very useful and similar to my other regression model process whereby I assess multicollinearity using an ols statsmodel. Just a question of curiosity, can we include all lags to see the coefficient factor for the entire time series? What will that tell us about the model prior to building in seasonality? PS. I'm relatively new to time series and your videos are helping me massively. Thanks for that, the maths are also very clear in the other videos and easy to understand.
We have predicted the sales at the first difference level. How do we relate it to the actual sales of the product? What is the relevance of the ARMA if we just predict the first order difference?
This actually frustrated me too. If the first difference is centered around 0 that means that the original series didn't actually have an upward trend, so why can't we just regress on the original data? Am I missing something?
@@ethanspitips5903 , Yes, Is this not ARIMA because he has taken the first difference to model ?? He has explained the same thing in ARIMA video and how to get back original series from first difference
I think I got the difference...Basically here in ARMA, you take the first difference explicitly and fit ARMA on it. In ARIMA, we let the model do the differencing itself by just passing the d parameter...
Thank you for making these videos, they are great! Just a quick question- if you, - or anyone else knows the answer. Which statistical forecasting technique is typically used to find the error term for the MA part of the model (what does Python do...)? That is, if you are to find the error terms - I guess you first have to make predictions, then calculate the difference between the real observations and the predicted values. Previos legs (AR-model) already exists so that part is fine, but previos error terms (MA-model) do not. So how do we initially make predictions for the MA-model? Thanks so much for any help or comments in advance :)
Hello, in your model 15(Time Series Talk : ARMA Model), you mentioned better orders of ARMA were the the total number of strong lags in either FAC or PFAC plots. However, in this model, you only picked the first strong lag as ARMA orders. Could you please explain why you only chose the first strong instead of all of them just as in model 15? I have thought I was clear how to choose orders after watching model 15 but I am confused now after I see model 16. Thank you!
How to build ARIMA models in Python without dates? If I'm estimating a target boats sinusoidal position in the ocean, do I wanna map milliseconds as dates 🤔, nah
Hi, you started with MA(1) process, but you didn't make sure that the autocorrelation is statistically significant. As a matter of fact, there are no significant autocorrelations except for lag = 12. This appears to be clearly when you use ploc_acf function from statsmodels.
How should we explain this error: "The computed initial AR coefficients are not stationary You should induce stationarity, choose a different model order, or you can pass your own start_params." What does it mean by coefficients are not stationary?
I think I got the difference...Basically here in ARMA, you take the first difference explicitly and fit ARMA on it. In ARIMA, we let the model do the differencing itself by just passing the d parameter...right ?
How can the model have access to the test data when you didn't pass it as an argument, or are you training and testing in the same dataset (and if so why??) ?
In this video you rejected the constant saying p-value is high. But while predicting you go ahead with the same model fit. Does the model fit by default reject all those variables whose p-value is more than 0.05? If not, I think we should somehow re-run using just those variable which passes p-value test. It would change the coefficients as well. Though, I am not sure how to do in Time series. Doing so in regression is very easy.
This is what confuses me a lot too. There is a lot of videos where it is said not to use the insignificant coefficients but no one explains how they are excluded from the prediction calculations. Have you managed to get to the bottom of this?
DUDE! How does this video only have 5k views... should be 500k at least!
Great explanation. Clear and Concise. 10/10.
I love how you revisit and review things from the previous videos of the series, it helps me further deepen my understanding on the concepts! Great work! Keep up!
Thank you for the explanation! In the video you mentioned that the ARMA(4,1) model shouldn't contain the constant due to it being p < 0.05. Does the model.predict() automatically take this into account or do we have to specify this before using the .predict() function?
Did you manage to find an answer to this?
Just minus the constant coef value (37.3039) from the prediction series that'll do. Notice that won't change the result much since the constant is fairly small to start with.
You are the best!!
Thank you very very much. You are the best!!
You're welcome!
This is great, I had a lot of suspicions about the statsmodel coefficient table. This is very useful and similar to my other regression model process whereby I assess multicollinearity using an ols statsmodel. Just a question of curiosity, can we include all lags to see the coefficient factor for the entire time series? What will that tell us about the model prior to building in seasonality? PS. I'm relatively new to time series and your videos are helping me massively. Thanks for that, the maths are also very clear in the other videos and easy to understand.
this video is very helpful!!! thanks!! will you do more coding vids?
How do you exclude the constant (with P value of 0.775)? You still used the same model (with the constant) for your predictions.
Did you manage to find an answer to this?
maybe since P is 0.775, it's not statistically significant, so we just set it to the null hypothesis which is 0
We have predicted the sales at the first difference level. How do we relate it to the actual sales of the product? What is the relevance of the ARMA if we just predict the first order difference?
This actually frustrated me too. If the first difference is centered around 0 that means that the original series didn't actually have an upward trend, so why can't we just regress on the original data? Am I missing something?
I was wondering the same thing and am confused about how to relate the first difference back to the data
@@ethanspitips5903 , Yes, Is this not ARIMA because he has taken the first difference to model ?? He has explained the same thing in ARIMA video and how to get back original series from first difference
I think I got the difference...Basically here in ARMA, you take the first difference explicitly and fit ARMA on it. In ARIMA, we let the model do the differencing itself by just passing the d parameter...
lol I was doing a project on some dataset and went through his video to understand it a little better and now it's messed 🥴
👌 👌
Thank you for making these videos, they are great! Just a quick question- if you, - or anyone else knows the answer. Which statistical forecasting technique is typically used to find the error term for the MA part of the model (what does Python do...)? That is, if you are to find the error terms - I guess you first have to make predictions, then calculate the difference between the real observations and the predicted values. Previos legs (AR-model) already exists so that part is fine, but previos error terms (MA-model) do not. So how do we initially make predictions for the MA-model? Thanks so much for any help or comments in advance :)
Hi! By chance, have you got and answer in these tel hesrs? 😅
Hello, in your model 15(Time Series Talk : ARMA Model), you mentioned better orders of ARMA were the the total number of strong lags in either FAC or PFAC plots. However, in this model, you only picked the first strong lag as ARMA orders. Could you please explain why you only chose the first strong instead of all of them just as in model 15? I have thought I was clear how to choose orders after watching model 15 but I am confused now after I see model 16. Thank you!
How to build ARIMA models in Python without dates? If I'm estimating a target boats sinusoidal position in the ocean, do I wanna map milliseconds as dates 🤔, nah
Hi, you started with MA(1) process, but you didn't make sure that the autocorrelation is statistically significant. As a matter of fact, there are no significant autocorrelations except for lag = 12. This appears to be clearly when you use ploc_acf function from statsmodels.
by applying ARMA to first diff ur basically applying ARIMA with d=1 to the time series data right
Where did you discard the constant term before prediction?
Did you manage to find an answer to this?
How should we explain this error: "The computed initial AR coefficients are not stationary
You should induce stationarity, choose a different model order, or you can
pass your own start_params."
What does it mean by coefficients are not stationary?
1:20 where did u define 'num_lags'
Thanks for the video..Is this not ARIMA because you have taken the first difference to model ?
I think I got the difference...Basically here in ARMA, you take the first difference explicitly and fit ARMA on it. In ARIMA, we let the model do the differencing itself by just passing the d parameter...right ?
How can the model have access to the test data when you didn't pass it as an argument, or are you training and testing in the same dataset (and if so why??) ?
In this video you rejected the constant saying p-value is high. But while predicting you go ahead with the same model fit. Does the model fit by default reject all those variables whose p-value is more than 0.05? If not, I think we should somehow re-run using just those variable which passes p-value test. It would change the coefficients as well. Though, I am not sure how to do in Time series. Doing so in regression is very easy.
This is what confuses me a lot too. There is a lot of videos where it is said not to use the insignificant coefficients but no one explains how they are excluded from the prediction calculations. Have you managed to get to the bottom of this?
where can i get the data set??
I notice you have an Australian accent, is that correct?