Time Series Talk : ARIMA Model

[@TheLionSaidMeow]

I never thought I would be able to learn ARIMA so easily off of one side of a single sheet of paper. This was the most lucid explanation I've stumbled across. Subscribed!

[@Stefan-hl8fe]

Anchors...used to keep things stationary. I caught that pun.

[@ritvikmath]

Hahahaha, I didn't even intend that :) My viewers are clearly more clever than me

[@TheJuwailes]

@Castiel Lewis wow you managed to come off as a creep and an idiot in less than 25 words

[@troykhalil4270]

i guess Im randomly asking but does someone know a way to log back into an Instagram account? I was stupid forgot my password. I appreciate any tips you can offer me.

[@Eizengoldt]

@@huxleyrodney3733this is a clever scam

[@giiigachadsr9960]

@@troykhalil4270how did it go?

[@hameddadgour]

At 45 years of age, I finally understood what the ARIMA model does. Thank you!

[@bestbest-qe3pw]

Thanks a bunch. You've done what my professor failed to do for a straight month in 9 minutes. Cheers to you

[@benoitl.8101]

Really simple and clear explanation of what I've been struggling to comprehend in the past few weeks. Many thanks from France

[@ritvikmath]

Glad it helped!

[@akashjain2694]

Probably the most clean video that explains ARIMA

[@hihi7896]

watch this man before every lecture to make sure I understand what's going on

[@milo1226]

This is exactly was I was looking for and was explained succinctly. Thanks for posting!

[@c4lb333]

I have an interview tomorrow that might involve time series knowledge, and your ARIMA, ARMA, ARCH, and GARCH series are really a life saver! They're explained very concise and clearly and saves me a lot of time looking through slides. Wish me luck LOL

[@laminann8061]

How was your interview? I hope it went well 😊

[@kachappillyjean]

This is what happens when people with the kanck of teaching gets their act together ! I have been banging my head after attending my Masters class that explained ARIMA. I really do not understand why these profs have to write a whole lot of math equations and read through it when all they have to do is to explain the concept just the way you did. This is the way to teach. Thanks for making my life a lot easier !

[@joaocamelier1381]

Congratualions for the quality of your content, it helped me a lot! You have gained one more subscriber.

[@castro_hassler]

Nice vid, I've seen every time series vid, I got so much intuition , thanks

[@willbutplural]

Loved the analogy with the anchor and clear breakdown of the equation! Subbed!

[@m.raedallulu4166]

Thank you so much, sir. I wish I found your channel long time ago.

[@AK-tj4ot]

You explained this so simply. Thank you so much.

[@ritvikmath]

Glad it was helpful!

[@high_fly_bird]

It's really great! You use only one paper sheet, and I basecally understood everything!

[@Blue17918]

You are much better for lecturing TS than my professor.

[@fpodunedin3676]

Note for self: an ARIMA model is the same as an ARMA model except that it will 'de-trend' data. This is through taking the difference of some a_t and a_(t-1) and then letting that be equal to your ARMA model.

[@hbeing3]

Thanks! The second time I watched this video just to revise. A question regarding the final a_k value. 07:38 Is a_k= the sum of all delta + the inital known value instead of the last known value you show here? i.e. a_l should be a_(k-l), or a_0?

[@aanilpala]

I got confused at the same point as well. I think it should be a_0.

[@haow9020]

No, it should not. (k, a_k) is to the right of the last data point, i.e., (l, a_l); assume l=k+1 and you'll see.

[@pedrocamunas5625]

Very clear and direct to the point, it helped me a lot, thanks

[@bugravardal6432]

Excellent clear explanation, thank you very much. I think you have clarified what was a question mark in my head the last few days, that is whether the additional inverse transform would still be needed when the differencing was performed by arima itself. Could be obvious to some but wasn’t to me…cheers

[@gustavosantanavelazquez7205]

You make it so easy to understand! Thank you!

[@yuthpatirathi2719]

Amazing explanation Ritvik!

[@rockyjagtiani]

Great work. Your videos are great contribution to Students and Teachers , during this Lockdown period. Thanks.

[@aryashahdi2790]

This guy is so damn good!!

[@ritvikmath]

this guy thanks you :)

[@benoitconley1126]

Thanks, super clear ! Merci from France !

[@ritvikmath]

You're welcome!

[@cobbdouglas690]

Fantastic and intuitive explanation. Thanks!

[@prevail8380]

At 5:49, is the order of I equal to 1? If so, how would the equation change if the order of I was 2 while the AR and MA orders remained 1?

[@TheEngVibe]

Takeaway for myself: ARIMA is the model applied for the time series data, where there is time dependence. It has a more step if transforming from crrelation of x and time to the correlation of x and x(t-1) (it's precedence). And from the formular of linear regressiin, the diff of x and x(t-1) is const (slope). So it doesn't depend on time. The 3 critiera for a series that can be applied ARMA (stationary): constant mean, constant variance, no seasonality.

[@abhishekv7171]

Well Explained Ritvik...Keep spreading knowledge!!

[@MrJatind3r29]

You explained it so easily! Great Job!

[@xwcao1991]

Man, you deserve a Prof. title

[@rezajavadzadeh5597]

You're awesome, thank you so much for making these

[@lavidrori7518]

You are the best I ever saw!

[@sanjukumari6453]

Thanks for explanation of mathmetical equations of ARIMA model

[@ritvikmath]

Most welcome!

[@nitsuanew]

This is an awesome video for ARIMA model.

[@dipeshkhati4895]

Saved the day for me! Thank you

[@tejaljadhav1275]

You explained it so easily!

[@sannederoever1320]

Writing out the equation for a_k, the logical conclusion seems to be that the equation ends with a_0 instead of a_l. Isn't a_l = a_{k-1}?

[@mmczhang]

that is what I thought as well

[@muhammadghazy9941]

@@mmczhang yep me too

[@joaojulio435]

I think it is, and the upper limit of the summation is k and not k-l (In my opinion). It makes more sense now, thank you for spotting this!

[@lanjiang5564]

Thank you so much for such a clear explanation!

[@manglem10]

Very different from others !! All the basics covered

[@kaiyanzhu3075]

I have a question, so in this video, the ARIMA is Stationary or non-stationary? or if it was transferred to the differences between a(t)-a(t-1)it will be stationary? Thank you

[@abrahamraja2088]

This helped me a lot, thanks

[@ritvikmath]

Glad it helped!

[@swastikkhadka6954]

Such a nice way to teach Thank you

[@pianoista6464]

Thanks for the clear explanation. One questions though, in estimating ak where you need to find summation of Zk-i where i=1 to k-l, but how do we estimate Zl+1to Zk-1, as how do you know errorl+1 to errork-1?

[@andreluisal]

Excellent!!! Congratulations!!!

[@JJ-ox2mp]

You're an awesome teacher!

[@kunalkiran3318]

When we had data till t=l, and we were trying to find the value for t=k, we need to a calculate a few Z (the summation of different Z). But for calculation of Z, we need the previous error. Since we do not have values after t=l, how do we calculate say Z at t=k of k-1?

[@gigi-oc8gn]

very well explained

[@HimanshuGupta-gl4ei]

Thanks, your videos are a great help.

[@듈이-k2b]

In the bottom of your sheet, with sigma z(k-i), wouldn't the last component be z(l) which is a(l+1)-a(l) ? But I thought a(l+1) is a future value.. Did I miss sth ? Thank you so much for the videos, I'm going through all of them!!!

[@jorangeeraerts3047]

Excellent video, thanks!

[@dominikc2559]

Hey there! I've got a question to your z_t graph, i get the part, that the average of z_t should be positive, since we got a positive linear function. But if we compare the next value with the previous value, we should also get negative values within that graph? If we only get positive values, the initial graph should be monotone rising, but in your example its a noisy rising graph or am i getting something wrong? Best Regards

[@user-cc8kb]

Very well explained! Thank you!

[@wissales-safi4938]

Thank u so much .. I rly love u man!

[@phuonghanguyen7406]

thanks, It helps me very much

[@ritvikmath]

Glad to hear that!

[@tracyyang1832]

Thanks for the great video. Very clear. One quick question, do we have to make sure the data to have no seasonality and constant variance to apply ARIMA model? Differencing, the I part, is to de-trend the data.

[@sangaviloganathan5194]

I am a beginner. Correct me if I am wrong. For example if the pacf plot shows lag 2,4 and 6 as significant, will the AR model be of the order 6? if so, how does the insignificance of lag 5 get factored into the model

[@ritvikmath]

Thanks for the question! Indeed PACF showing 2,4,6 means you should include those lags in the AR model. By not including lag 5, we are saying that it is not important in "directly" predicting the current value

[@animeshtimsina3660]

@@ritvikmath If we use the order 6 then doesn't the model automatically include lags 1,2,3,4,5 and 6 in it? If this is true then how can we tell the model that lag-5 is insignificant but lags: 1 to 4 and 6 are?...PS. I am a beginner!

[@sohailhosseini2266]

Thanks for the video!

[@alteshaus3149]

Super video man!

[@AlankritIndia]

shouldnt we add a constant term like phi(0) in Z(t) eqn..like we had in previous model for ARMA?

[@ahmadabdallah2896]

i thought the same thing

[@mengnixu7247]

thanks ! U explained clearly

[@ritvikmath]

thanks!

[@xuechen-m9g]

beautiful model

[@guilhermecoelho2354]

The "I" part is to be equal to 1 when we have a unit root on the time-series. Not when there is a trend !!

[@vijayantmehla7776]

Very well explained.. Thank you !

[@terryliu3635]

Again, great explanation! Do you have any videos on multivariate ts analysis or prediction? Thanks

[@randall.chamberlain]

But if I take the original time series and apply a diff1 to make it stationary, couldn' I just apply an simpler ARMA model instead?

[@pallavibothra9671]

Please make video on RNN, LSTM..Eagerly waiting for that :)

[@mrfm10]

Thank you so much for this!

[@LukasHesse-po1ri]

why is a_k further down the x-axis then a_l? shouldnt it be the other way around?

[@aimalrehman3657]

what is epsilon_t-1 in the MA bit of the ARIMA equation?

[@officialmintt]

Thank you so much! May I ask for an example of an application/occasion where we might do the second difference?

[@krzysztofrozanski466]

Hi, sometimes when predicting house price indices, you might need to go with second difference to make them stationary (at least this happened to me once). I would not treat this as a rule for all house price indices in the world, however, as it for sure was "series specific". Hope this helped :)

[@qanhdang4035]

This explanation will be better if the notation used is consistent with the explanation on ARMA model. Also, for ARMA applied on z, likely it lacks the bias phi0 (which is beta0 in your ARMA explanation). Anyway, it's a good explanation of ARIMA.

[@TheTehniga]

Didn't understand how to compute ARIMA(1,1,1), nor how to obtain the predicted value.

[@michaelelkin9542]

Why is the MA part done on a() and not z() shouldn't both parts be on the stationary z() data? Thank you.

[@sfundomabaso3200]

Wonderful videos you make. I'm just curious whether do u do these models on statistical programs such as R or Stata

[@4lex355]

it is not aL in the end but a1.

[@kostyamamuli1999]

Great tutorial man!

[@neilhouse4591]

Great help. Thanks!

[@TheEngVibe]

Many thanks 🎉❤

[@nickcorona3966]

How do you calculate the errors?

[@samk3566]

What is the diff between differencing and removing the trend??? Does stationary simply lack of trend and seasonality??

[@ansylpinto2301]

Not entirely true but presence of trend will violate constant mean and seasonality constant variance. ARIMA models work well with stationary data so it is important the values used to model them do not have trend and seasonality.

[@evrenbingol7785]

What if you want to predict so far into the future that K-i goes out of bound. say L is 100 and K is 1000. (Z sub K - i) would give you out of bound error since.(you are trying to go back to negative Ts, Since you do not have 900 Ts, So the assumption is you can only predict into the future as much as the length of your data? Is that correct.

[@ritvikmath]

Yes that is correct. Intuitively, you likely don't even want to predict out that far since your predictions probably won't be great.

[@Flyer111100]

hi awesome videos, just wanted to know if it is also possible to just multiply my zt value times my a value at t to obtain my future value?

[@longwenzhao9204]

amazing...so clear...

[@alecvan7143]

Great video! :)

[@ritvikmath]

Thank you!

[@areebwadood6273]

Could ARIMA be used if the anchor chart had an exponential trend instead of linear ?

[@mithunim]

My guess is you can use ARIMA but instead of differencing the series once to make it stationary, you might have to difference it at least twice.

[@spytheman]

Why can't we just do an ARMA model where we transform the model into the difference of the anchor? Or by doing so it is a ARIMA model instead?

[@swarnaramakrishnan6614]

At the start, its mentioned ARIMA can be used on models that show a linear upward/downward trend and the only stationarity violation being mean is not constant. In his previous video on ARMA, he would have done the differencing on a non-linear model. But am now wondering why values were not recovered in ARMA sample code.

[@iimram]

What if the time series is exponential? Because calculating Zt also wouldn't help, isn't it? Zt itself will not have constant average.

[@mithunim]

What I think is you can use ARIMA but instead of differencing the series once to make it stationary, you might have to difference it at least twice.

[@krzysztofrozanski466]

If the series is exponential, differencing any number of times would not help. It might mean the series is "inherently" not stationary (you might think of it as a derivative of an exponent is exponent, same function) and instead of "usual" time serie models you need to use some other, nonlinear ones or if you have two non stationary time series, you can check cointegration models. Or simply use log transformation for initial time series instead of differencing, maybe it will help ;)

[@user-or7ji5hv8y]

How about cointegration? Is that useful?

[@amira_369]

Best video!

[@shaswathpatil3439]

Thank you!

[@Ju-dk1eg]

Great teaching

[@ruifernando8066]

how to determine the value of p,q?

[@annasmith7226]

I thought differentiating to fix unit root problem not trend problems

[@davigiordano3288]

Thank you

[@poppyblop484]

If there is still a trend, then doesnt mean that this model is just missing a component that needs to be accounted? Or am i missing something and that was only for the basic OLS model

[@shrikantlandage7305]

Thanks that was too straight forward.Good Work

[@EmouAcademy]

nice. But still couldn't get p, q..

[@sandeepmishra3275]

Amazing