live_tv
Livestream Starting Soon
00
Hours
:
00
Minutes
:
00
Seconds
Up next in 10
Did you know apart from enterprise-scale data modeling and data visualization features, Power BI also includes advanced analytics tools such as clustering, forecasting, anomaly detection and Machine Learning? If you already knew that, do you know how to use them correctly and get the most out of them? In this session spanning multiple parts, we will cover these advanced analytics tools in detail to gain thorough understanding of them so you can use them confidently in your reports. We will also cover how to deploy machine learning models in Power BI.
About the Speaker:
Sandeep is a Simulation & Data Analytics engineer. His background was in research in Mechanical engineering and product development and always used Statistics to analyze data. For the last four years, he has been using Data Analytics & Machine Learning to improve manufacturing operations. He has an MS in Mechanical Engineering, Masters in Engineering Management with focus on business analytics, and is a certified Azure Data Scientist.
Show More Show Less View Video Transcript
0:17
Hello everyone on AI42. We have again the Sandeep today. Sandeep, can you introduce yourself
0:40
sure yeah thank you for having me here halkan and gozia my name is sandeep powar i'm the data
0:46
ytics engineer i work at pre-lighting in wisconsin united states um yeah and i this is
0:55
my second or third actually so this is my third session here with ai42 the first one that i did
1:02
was excel for data science second was the first part of this series on power bi doing advanced
1:09
ytics in power bi and now we're here for the second part yes we are exciting to hear what you
1:16
will teach us this time about power bi yeah so this one the last time we looked at uh clustering
1:25
and then key influencer visual for this one i was a couple of people reached out to me
1:35
after the presentation and they wanted to know more about forecasting. So in this one we'll be
1:40
looking at a very deep look at how forecasting in Power BI works, how you can use it, how you should
1:48
not be using it and I think forecasting is something that most businesses use, so I think
1:57
our audience will find it very helpful in their way. Yeah it sounds very practical and very useful
2:03
so we're really looking forward here to learning more about it but I think before that we will just give some more information about AI42
2:11
yes sure so so the motivation for starting AI42
2:32
comes from the recognition that there is no good starting material. And AI42 is a strong team consenting of free Microsoft AI MVPs
2:41
that strive to provide you with a valuable series of lectures that will help you jumpstart your career in data science
2:48
and artificial intelligence. AI42 aims to provide you with the necessary know-how
2:54
that can help you dream job as long as it's related to fields of data science
3:01
or machine learning. The concept is simple and involves professionals from all around the globe
3:06
explaining to you underlying mathematics, statistics, probability calculation, data science, and machine learning techniques
3:14
And we'll guide you through all you have to do is follow our channel
3:19
enjoy the content every second week filled with real-life cases and experts' experience
3:25
Don't you worry, we all started from scratch and we are happy to help you build it up from there
3:30
You can always stop and rewind the videos or ask for clarification in the comment section
3:36
We hope to assist you on this wonderful journey and have you as a speaker one day
3:41
By creating cross-collaboration with other organizations, we can give the best opportunities to broaden your network to the AI and data science communities
3:51
With the combination of our offer services, we would support less fortune people and organizations that are not that recognized yet, even though they deserve it
4:02
Yes, and our organization AF42 is sponsored by Microsoft and Miles, and we're very grateful for that
4:11
But we also have some other contributors. We have Levento Pongor, who has been our graphic designer
4:17
And we also have our intro music, which is by Mina Marie. And we're also in close collaboration, both with C Sharp Corner and with the global AI community, so that you can see our content also on their channels
4:29
And we would also like to thank Nicolette Toth, who has been providing us with the text material, both for our website and also for our sessions and for our advertisements
4:41
And if you want to know more about us and if you want to follow us, you can find us both on Facebook, on Instagram and on Twitter
4:50
So please follow us so that you can get the latest news on all our events
4:57
And if you would like to see our previous sessions, they are all stored on our YouTube channel
5:02
So you can just scan the QR code here to get to the AI42 YouTube channel
5:06
And if you want to be able to register and get notified when we have new events
5:11
you can scan the other QR code to get to our meetup page
5:15
And as well, we really value everyone. So we have a code of conduct
5:23
Do everyone feel welcome and acceptable in our events? So if you have any questions, you can go to our website to find what we have written in code of conduct
5:36
Yes. Yes, I think let's go back here to Sandeep. yes so sandip is there anything special that you think people should think about or should
6:00
they be prepared for anything for this session uh so the the session itself is very uh hands-on
6:07
um and you should be able to follow along with me what i've also done is if you go to my website
6:15
I've written a few blog posts about this very topic. So if you have any questions
6:24
and maybe if I gloss over any of the things in my presentation here
6:30
you can refer back to my blog there and then read about it
6:34
And if you still have any questions, just contact me, and then I'll be happy to help you
6:39
That's nice. And then also for the viewers, we will have a Q&A session after the presentation
6:48
So please feel free to pose any questions that you have in our chat or in the YouTube chat or wherever you look
6:57
So then I think we will just give the stage here to you, Sandeep
7:01
Thank you. Hello all, welcome to the second session in Power BI for Advanced ytics
7:24
My name is Sandeep Pawar Before we get started again thank you AI42 for having me for the third time I really enjoying these sessions watch the other speakers as well
7:36
So great to be here. Hopefully you were able to join me last time as well
7:41
So in the first part, we looked at some of the Power BI Advanced ytics things
7:49
So for example, we started with the clustering, bivariate clustering and multivariate clustering
7:56
how it works in Power BI, what are the limitations and how you can use it to identify
8:03
patterns in your data. We also looked at Key Influencer, which is a visual that's natively
8:12
available in Power BI to yze categorical or continuous data. It gives you factors or
8:21
columns that are influencing a metric of your choice. We also discuss how it's done
8:29
and some of the algorithms behind it. After the last session, some of you reached out to me
8:38
and you expressed your interest in forecasting. I was going to cover a couple of more additional things
8:47
but since there was some interest in forecasting, I thought it would be best if we could dedicate a session to forecasting itself
8:59
So today what I want to do is we will look at a practical demo
9:03
of how forecasting can be done in Power BI. And at the same time, we will take a deeper look
9:11
at exactly how it's done in Power BI, which is way more important
9:17
If you look at the documentation, Power BI's documentation on this topic of forecasting
9:24
you will see that there is a page, but there is not much to it
9:29
In fact, there is very little to it. It just shows you how to create the forecast
9:35
but it does not go into a lot of details. So a couple of years ago
9:41
I used forecasting a lot in my work. And forecasting is something that no matter
9:48
what type of a business you are in, the company always needs it
9:53
You're always trying to forecast something, and it's probably one of the easiest
9:57
and most used method of predictive modeling. So I was in a similar situation and I could not find a whole lot
10:06
So I spent some time understanding how forecasting works in Power BI
10:14
So what I want to do now is share that with you and hopefully you will take something meaningful
10:21
out of it and use it in your own practice. data that I'm going to use in this demo it's a quarterly sales data for a French
10:31
retailer I've included the link here which obviously you can see here but I'll
10:37
put it in the we'll put it in the description of the video where you can
10:42
get the data and we will look at we'll yze the data and we'll also see how
10:50
we can use this data to create a forecast our business goal here is to create a forecast for
10:58
the next eight quarters okay next eight quarters meaning for the next two years
11:06
in this case in addition to just creating the forecast in power bi what we also want to
11:13
understand is how is this forecast created in Power BI? What are some of the algorithms that
11:21
are used to create these forecasts? What are the assumptions behind how Power BI creates these
11:28
forecasts? Importantly, if you create a forecast, how accurate is it? How do you assess the accuracy
11:36
of a forecast that Power BI creates. Can we improve it? And then when does it really work
11:43
and when does it fail? And this is very important for you to understand and not just blindly create
11:50
the forecast in Power BI. Let's get started. The data is actually very simple. I've already loaded
11:56
it in Power BI. So let's go over here. And the data in this case, let me show you, is right over here
12:09
I'm going to pull it up on my screen. So let's go to some date. And then this is my sales
12:18
okay let me on the right hand side over here it created a hierarchy what I want to do is get rid
12:25
of that hierarchy so I'll click on this triangle button and then get rid of the hierarchy so we'll
12:31
go to dates only and it will convert to dates so we have dates now let me increase the font size
12:38
so you can see it easily okay good enough so it created the we have the data here as you can see this is from 2012 to
13:03
2017 and the data is in quarters. So if you see the numbers here, first is the end of March
13:13
then the end of June, then the end of September, so every three months. We can create a line chart
13:20
out of this. So let's create a line chart and you see the data over here. A few things that you want
13:26
observe here the first thing is that when you block something you want to
13:32
maybe first look at what is the overall trend in your data and you can do that
13:37
by clicking on here and then go to the Advanced ytics pane over here and
13:43
then click on the trend line okay when you get a trend line here and then add
13:49
it will create a trend line which shows us it's it's increasing and visually we
13:54
can see that but this just helps us you know put it on a chart and understand that next thing we
14:02
want to observe is these peaks if you see these peaks they appear in a very regular order right
14:11
and that's because this is a quarterly data so it looks like it increases or we have sales that go
14:20
up based on the season end of September. So this is a quarterly data. That's why it's in regular
14:28
three months period. And there is a nice seasonality to it. To create a forecast
14:33
it's actually very simple. To create a forecast, we select the data. We create the... it always
14:40
works in a line chart. If you have it in a tabular or maybe a bar chart, it won't work. So it has to be
14:46
a line a line chart and another thing that you have to be careful about is if
14:52
we go to the x so i selected that and if you go to the format pane and then click on the x let click on the x you will see two options here continuous and categorical you have to make sure that you have
15:10
continuous selected if it is categorical the the forecasting algorithm won't work it has to be
15:18
continuous okay so make sure it is continuous then we click on we go to the ytics pane again
15:27
and so this is the ytics pane and right at the bottom here is the forecast option okay so
15:34
we will click on the forecast option and it will give you we'll click on add and it will give you
15:41
a bunch of different options as soon as you click on that it will create a forecast for you but there
15:47
are various options over here it's first asking us what's the length of your forecast by default it
15:53
just pick 10 points um and then confidence interval took we'll look at all these options later
16:00
but for now i just want to create something and show you how it works and then the seasonality
16:06
uh and then for now we'll just accept the defaults and in this case the only thing i'm going to change
16:12
is 8 forecast length is 8 and then click on apply and as soon as you create it there you go we have
16:19
the forecast that's created it doesn't show you the label so even if we go here and then turn on
16:27
the labels so i went here and then turn on the labels here you'll see that it created the forecast
16:35
but it did not add any of the labels to the forecast if you want to see the labels to the
16:41
forecast that's applied to the forecast you have to click on these three dots and then click on
16:49
show as stable and once you show as stable let's go over here show as stable you will see this is
16:57
the sales and then you have the forecast values that are created and these are the forecast
17:04
values there's no way to export these out so you when you create a forecast you see the forecast
17:11
and that's about it. Okay, so this is simple, right? I mean, you just have to click on
17:18
one single button and it created the forecast for you. Now, the question is, how is this forecast
17:28
created? So let's first quickly go over what I have, what I found as to how Power BI creates
17:37
this forecast and there is no official documentation on this topic and this is just me yzing and
17:44
when reverse engineering a lot of these things and then based on my experience how forecasts work
17:52
i create a lot of forecasts as well more advanced forecasts using python and r
17:57
and so this is just based on my own experience knowing how forecasting works
18:02
The algorithm that Power BI uses to create this forecast, it's called ETS. ETS stands for error
18:12
trend, and seasonality. There are two variations of that. The first is, let me make this bigger
18:20
for you, it's easier to see. So the first variation of that is, the algorithm is seasonal. If the data
18:27
is seasonal, which in our case it is seasonal, we have quarterly seasonality, then it uses ETS-AAA
18:36
And if it is non-seasonal, it uses ETS-AAN. And we will look at exactly what that means and we'll
18:43
decode that. It's one of the things that I hope that you will take out of this session
18:50
exactly how understand how this ETS works at a very high level and whether you should use Power BI
18:58
forecast or not will hinge upon understanding the pros and cons the advantages and the
19:08
limitations of this ETS algorithm. If there is missing data if there are missing data in your
19:17
time series, then Power BI will do linear interpolation and then fill the missing values
19:24
You will look at that. It will detect the seasonality automatically for you. So this is one of the
19:30
things that if you use Python or R, you have to input the seasonality value yourself
19:39
In Power BI, it detects, and that's the novelty here, it detects the seasonality for you
19:46
You don't have to enter it and then it will create the forecast based on that identified seasonality
19:55
And then the very last thing is that it will create the forecast based on the horizon that you have selected
20:02
In our case, the horizon is eight quarters into the future. Let's go to the next page
20:12
So exponential smoothing, we talked about over here ETS. And ETS, as I said earlier, is error trend and seasonality
20:23
It's an exponential class of algorithms. If you are familiar with forecasting, you might know that there are other types of forecasting algorithms
20:35
ETS is one of them. ARIMA, SARIMA, GAR, CHUBAR and lots of other many many different types of
20:46
forecasting algorithms. What Power BI uses is ETS. Now if you Google on you know
20:53
Google Power BI forecast and try to find what forecasting algorithm it uses you
21:03
may find some forums where it says it uses Sarema which is not it's completely inaccurate so
21:12
just know that RBI uses ETS. The ETS-AAA is nothing but a fancy name for
21:24
single exponential smoothing class of algorithm whereas ETS-AAA is a Hort-Winter's method
21:32
so before I use too many jargons here let's actually understand what that
21:38
means what is a a an what is ets and what all of this has to do with it being
21:46
exponential class I'm gonna open my excel sheet over here the first thing
21:50
that you want to understand is the concept of lag it's pretty easy to
21:55
understand lag is let's say you have a time series like this a lag is just a
22:01
first lag is just its first we shift the series by one by one period so it's
22:10
fast that's just offset by one periodic value so the first the 286 we it comes down so we lose 200
22:20
so we lost the the recent past right then we lack two would be we lose last two values If you see 372 372 is a town over here
22:35
And the importance of understanding that is when you are creating a forecast
22:41
The forecasting algorithm will look at how your time series is correlated with its past
22:53
So in statistics, you know the concept of correlation. You take variable X and variable Y, and you find a correlation between the two
23:02
It's exactly the same top. It's exactly the same thing. Instead of correlation, we calculate autocorrelation
23:08
Auto means with itself. So if we can, here, we can do correl, for example
23:16
Go here, and then take this. And then it will tell us how this is correlated within the past
23:23
So this is negatively correlated and 0.85. So not much in this case
23:29
Well, the exponential part comes in is, let's look at this chart over here
23:36
Let's understand what the exponential and what the single exponential means. So imagine this is our same quarterly data that we have
23:46
In the exponential smoothing algorithm, but the algorithm, what we do in the algorithm
23:52
is we assign a value alpha, also called as weight. We assign a weight, like in percentages
23:59
between zero and one to each of the past values, to each of the lags
24:08
So let's imagine we are over here, we are over here, and we want to predict the next value, right
24:18
We will look at the last value, assign the value alpha to it, that value can be 0 to 1
24:25
Let's say that value is 0.8. Then it will look at a value before that value
24:33
So that would be two quarters in the past. And then the same 0.8 value now will be 0.8 multiplied
24:45
by 1 minus alpha, which will be 0.2. So if the value here was 1000, let's see
24:56
then our first value will be 0.8 multiplied by 1000. So which would be 800
25:04
If the next value here is, let's say 800 here, 800, 0.8 multiplied by 0.2 is 0.16
25:14
So it will be 800 multiplied by 0.16. okay whatever that number uh comes out to me okay and then we will keep doing that so the quarter
25:27
that three quarters in the past we will again take 0.8 into 1 minus uh alpha which is 0.2
25:36
and the square of that then a cube of that and the fourth power of that like this if we actually plot
25:42
all of these weights so i if i took 80 percent 80 is this 0.8 that we talked about and then
25:50
plot these out you will see that these weights diminish very rapidly so we assign 0.8 value to
25:59
our last quarter then 0.16 value 0.16 value to two quarters away and then we are taking only three
26:08
percent of the value three quarters away and so on and so forth and if you look at this decay this
26:16
This is exponential, right? And that's where the exponential part comes in
26:22
If we were to use maybe a value like 0.25, it won't decay as much
26:27
It's decaying very linearly. If we do, let's say 0.5, for example, again
26:33
the same thing. If we take 0.8, then that decay, okay, something, I messed up something over there
26:42
Yeah, let's do that, 0.5. decay is not as much and that's where the exponential part comes in and this is for
26:53
the single exponential part of it. Single exponential meaning there is no seasonality to it
27:02
and if there is no seasonality that's associated with it it will use this single exponential
27:07
smoothing AAN. If there is a seasonality it will use Holt-Wenther's method and the way the Holt-Wenther's
27:14
method works yes RBI will decompose the time series into the composer time series into trend
27:24
seasonality and cycle trend meaning how is the long-term average it is monotony it could be going
27:32
up or going down there is no up and down it can only go up or it can go down so the red line over
27:37
year is the trend. Seasonality, on the other hand, is it changes frequently and it's fixed in length
27:46
In our case, the seasonality is quarter seasonality. You have monthly data, the seasonality
27:52
would be 12. If you have a weekly data, the seasonality would be 52 because there are 52
27:57
weeks in a year. In the ETSAAA algorithm, it works exactly similar to the single exponential
28:10
smoothing that we looked at. Power BI will look at the trend, it will decompose the trend
28:17
and then apply that exponential weighting that we just created. It will extract the
28:24
seasonality and then again go through the same process of applying different ways and how does
28:30
it know which weight to use well it's it's an optimization problem it will go from it will just
28:36
like a hyper parameter tuning it will go from values let's say 0.1 to 0.9 it will look at
28:44
all the possible values calculate an error and whatever whichever value gives
28:51
the least error is what that alpha value is. So ETSAAA works the same way. It will
29:03
decompose the time series into trend cycle seasonality, apply that exponential weighting average to it and then calculate the forecast. It will add all
29:16
of those up together and then it will create a forecast the aaa that we looked at it means
29:24
additive error additive trend and additive seasonality what that means is if your data is such that if your data is such that the trend is linearly increasing
29:41
or it could be linearly decreasing doesn't matter if it is linearly increasing or decreasing
29:46
It's called an additive print. If it is increasing at a rapid pace
29:51
or if it is decreasing at a rapid pace, like this shown over here, like this
29:57
or it is decreasing like this, then it's called multiple trend. If it is increasing but it is increasing at a slower pace then that means it's a damp trend
30:10
So the A stands for additive it could be multiplicative so ETS would be M as well or
30:16
it could be D for damp. So if it is like this then it is increasing for sure but it is decelerating
30:24
quite a bit. Same thing with the seasonality we won't go into it but the thing that you have to
30:29
know here is that power bi only includes uh additive uh it only includes uh the additive
30:39
trend as we saw here uh look at that yeah aaa aaa so it only includes additive trend and
30:50
additive seasonality so all these different flavors of forecasting that are available or how
30:59
a time series could be model as it could be man it could be a a a a m a so on and so forth these
31:08
are all the different combinations that are possible but power bi only supports aaa the
31:15
additive trend and additive seasonality which is these two so this is aan because this is no
31:24
seasonality and this is AAA so additive seasonality and additive trend so your time series is increasing
31:32
but increasing linearly so to speak and same thing with your seasonality so this is very important
31:41
for you to understand that what that means because the implication of that is if you are working
31:49
with a time series and if you plot it and that's why i i created a trend earlier if you plot it
31:57
and if you see that you uh the the trend is one of these two the multiplicative or tamp
32:07
then power bi for power bi won't be able to capture that trend uh that non-linear trend for you and
32:14
whatever forecast is created will be it will create a forecast still but it the forecast
32:22
won't be as accurate if you were to use the multiplicative or damped trend
32:30
in most practical cases it's actually neither so it's never like completely additive or completely
32:40
multiplicative or completely damp. It's some kind of a mixture to that and there are things that are
32:46
done like transformations of like box cost or box cost transformation, lock transform or things like
32:53
that to convert it from multiplicative to additive or from damp to additive. Damped trend typically
33:01
using the damp trend method usually gives better forecasts because there is nothing usually that
33:10
just keeps on increasing. There is some deceleration, there is some amount of
33:16
component to that trend or seasonality where it just doesn't keep on increasing
33:24
So this is where that AAN comes in. So as we discussed earlier, this is for you
33:33
this is very critical for you to understand. So always when you get your hands on a time series
33:38
always plot it first then and and always create a trend uh unfortunately over here it will always
33:45
create a linear trend for you but this at least gives you a good understanding uh if it is following
33:52
a linear trend or not if it is it if it does follow the linear trend um then you know it's good
34:00
just to give you a quick example of what a non multiplicative seasonality would look like
34:07
it would something look like this where it increases right and then it these
34:18
peaks the distance between these peaks it increases at a very higher rate okay
34:29
And that's where the multiplicative seasonality comes in. Okay. So we created that
34:38
So this ETS, is it good? It is actually. So if I blow this up a little bit here
34:46
this is from a competition called M3 competition. Usually every couple of years or so
34:54
where they share a lot of different time series and forecasting practitioners they participate and they apply various
35:03
forecasting algorithms and a winner is announced. If we look at the E3 M3 data
35:11
and M3 results ETS if you look down here this is on the y-axis is the error so
35:21
smaller the better. ETS is actually one of the best algorithms to use. It is in
35:28
fact as good as, I mean in many cases even better than let's say using the
35:37
random forest or using RNN, the neural networks or even CART which is the
35:44
decision tree and random parts. So ETS is actually very good, it's very versatile
35:53
and it's used in practice quite a bit for a univariate time series. It's very robust as well
36:01
It's robust meaning it is very robust to the outliers. If there are any outliers in your data
36:09
it works slightly better than using an ARIMA model. And it really easy to forecast even on large time series data The limitations of using ETS is that it cannot be used with high frequency data And again this is one of those things that you have to understand
36:35
If you get your hands on your data, and if the data is monthly or yearly, you are good
36:45
Or even it is, you know, maybe weekly data, you're good. If the data is hourly data or daily data and if you apply ETS algorithm to it, you will not get
37:00
accurate forecasts. And the reason for that is hourly or daily forecast usually has multiple
37:08
seasonalities in it. So imagine you have data that shows the daily production rate for a
37:19
manufacturing company it most likely has some weekly seasonality pattern to it or maybe monthly
37:28
seasonality pattern to it and ets we can only use one seasonality value so if you have our your daily
37:38
data and this is one of those things that i see many rbi practitioners make an error in this that
37:47
that they have the data and then they just use it and then they see the forecast is just showing
37:54
one single line. And the reason for that is because if it does not detect
37:59
a seasonality in your data, it will switch over to that AAN algorithm, which is single exponential smoothing
38:06
Single exponential smoothing always creates a single line forecast. It will not give meaning
38:15
Let's go over here. if we if we were to apply a single exponential smoothing in this case it will just be one single
38:23
line basically an extension of this trend line because that's what uh it will just extend the
38:30
line out so if your data has um high frequency count in it or it is high frequency then don't
38:40
use it okay or if it is intermittent data as well intermittent meaning uh let's see
38:47
Intermittent meaning you have something and then no sales for a few days or a few weeks
38:59
and then some other and then, you know, and then so this is very sporadic in nature
39:04
And when you have data like that, ETS won't work either. Imputation meaning if you have missing values in your data, then Power BI will just impute those missing values by interpolation
39:30
So in this case, I manually deleted some of the data. So let me put it here
39:39
And now look at this. The way this works is, let's actually go over here
39:45
I manually deleted these two values from this. And the way interpolation will work is it will take the value before and after and then just average it and then create a value in between the two
40:02
So it will take, for example, let's go here. It will take this value plus this value and then divide it by 2
40:16
So 612. So notice the original value was 582. The interpolated value is 612
40:24
And the forecast created based on that would look something like this
40:31
which is not, it doesn't look accurate, right? So you have to understand if there are any
40:38
if there are any, so yeah, so if there are any missing values in your data
40:47
if there are missing values, then Power BI will automatically interpolate those values for you
40:54
It could also be that maybe it is, the values are missing
40:58
because maybe it is zero. You did not have anything to report for that period
41:06
which in that case, zero, or it's not actually missing, right? It's just that we did not have anything
41:15
If par bi, if you have null in that scenario, then you should replace that null with a zero value
41:24
So you just keep that in mind. Again, knowing how the ETS algorithm works, we know that when we are looking at the ETS algorithm, it will assign weightage to the past value, right
41:41
So if the missing value is towards the recent past, like over here or here, then the interpolation, if it is inaccurate, it will lead to greater inaccuracies in your forecast
42:03
If the missing values are in the distant past, because how the ETS works is it doesn't look at the past data as much, then you are C
42:17
So you have to understand why you have missing values in your data
42:22
And the second thing you have to understand is where the missing values are
42:27
If there are more than 40% missing values in your time series, Power BI will not create any forecast at all
42:41
If the missing values are less than 40 which is pretty generous I think it will use the interpolation method that we just talked about and then create the forecast based on that The limitations of that obviously that we been talking about is
43:04
you know, one big thing is that it will not give you any sort of error
43:08
If you have any missing values in your data, RBI will just impute those for you
43:15
I wish it gave some error that, hey, I'm going to impute those missing values and this is what it looks like
43:23
It doesn't do anything of that sort. It just imputes the missing values and then it gives you the forecast based on that
43:35
Seasonality detection. We won't go into the details of this. You can go to my blog at PawarDI and then I've written about it
44:15
Thank you
44:44
Hi, good evening everyone
45:05
We are sorry about the technical issues. We are aware that the sound was not perfectly fine during this session
45:15
I'm not sure if we have our speaker on board. Yes, we do
45:22
Hi, Sandeep. So that was a little issue with the session. So we started, sorry
45:32
I'm not sure if you have any questions from the audience. There was one for sure about if you could share the link that you've shown in the beginning
45:43
about data that you used? Can you hear me? Yes, we can hear you
45:50
Okay. Sorry, could you please repeat the question? So the question was that you were talking about that
45:58
that you were sharing, you were using, sorry, some data for this session. And if you can share that information
46:06
Sure, yeah, I can do that. Let me, so for those of you who want to look
46:12
at the data, the best way would be if you go to my blog at PalwarBI and then go to the
46:19
first page and I'm going to put it in the chat window here and the organizers can get
46:27
the data. Let me give you one second. Do you see that in the chat window
46:40
Yes. Yes. Okay. You can share it. You're going to share it in a sec. Yeah
46:46
In the meantime, we have another question. Okay. From James, who is asking, will you advise to use Power BI for advanced forecasting
46:59
So, it depends on what your goals are. And as we looked at in the presentation, if it depends on the data as well, if the data is low frequency, meaning if the data is, let's say, weekly, monthly, quarterly, or yearly, then you can use, and if there is some seasonality to it, then you can use forecasting in Power BI
47:26
And the reason for that is, as we saw, it uses the ETS algorithm in Power BI
47:35
And the ETS algorithm works with low-frequency data. If you are working with high-frequency data, like hourly data or daily data, then ETS won't work as good
47:48
And the reason for that is when you have high-frequency data, you may have more seasonality in your data
47:55
so you won't be able to capture that seasonality accurately. And I guess the second part of that is it depends on your goal as well
48:06
If the goal is to just show the forecast to your business users
48:11
then, yeah, it's just more of a visualization thing. If the goal is to extract your forecast values and then forecast them
48:22
and then maybe create some sort of a DAX measure, then you can do that
48:29
You will have to do that outside of Power BI. Thank you for your answer
48:37
and for the great question as well, of course. So for those who are just joining us now
48:48
like in the last few minutes, and that I just wanted to mention that this session is going to be recorded
48:55
and will be available on our team's channel. So you can check out again
49:02
so you can see the session whenever you want to rewatch it
49:11
Do we have any more questions from the audience? No no more questions as I can see All right Then maybe we can bring in our last few One more question actually here So we have a question here from Gabriel
49:30
He says, in your experience, would you change the confidence interval of 95%
49:36
What other parameters would you adjust? Great question. So there is a little bit of the confidence interval here
49:47
is a little bit misnomer. And I'm not sure why Power BI chose that confidence interval
49:54
What Power BI is actually calculating is prediction interval. And confidence interval in forecasting
50:02
doesn't help you a whole lot. So the difference between confidence interval
50:07
and prediction interval is that prediction interval calculates uncertainty around your mean forecast
50:15
Whereas confidence interval is more of a way to calculate for your mean forecast
50:24
what is the bank, the confidence interval bank. And in most cases, in business scenarios, we are interested in the prediction interval
50:33
because it calculates the uncertainty around those mean estimates. So just one clarification there
50:39
So even though it says confidence interval, and you're reading it in a book
50:43
and it says prediction interval, just know that RBI is actually calculating prediction
50:48
Now answering your question about would you change the confidence interval of 95%
50:56
that also depends on your business context. In most companies, it depends on the risk you're
51:04
trying your business can take. If you can, if you want to buffer the risk, meaning you cannot
51:12
take a lot of risk then 95 you can leave it at 95 because you want to see all the capture you want
51:18
to see whatever is the uncertainty in your estimate but in most scenarios your company is likely
51:26
it will likely want to take some risk right because unless there is risk there is no reward
51:33
and you want to see uh what that reward is and in that's in those scenarios in most likely
51:38
scenarios you would reduce that confidence interval from 95 to 80 percent and that's a very typical
51:45
80 20 rule and then you would reduce it to 80 percent as far as what other parameters would you
51:52
adjust there are no parameters at least in power bi that you can adjust so the only thing that you
51:59
can change here is the band around the confidence interval which you can change it to 80 percent
52:05
the seasonality factor is there but that will depend on what type of seasonality you have in
52:13
your data if you're working with let's say monthly data then the seasonality number is 12
52:20
if you're working with quarterly data that seasonality number is 4 so on and so forth
52:25
and but that's not a parameter of the forecasting algorithm itself it's just you know whatever it is
52:32
And the third thing there is how many data points you forget or you ignore the past data points that you want to ignore
52:43
And in that case, if you want to check how accurate the forecast would be
52:52
So let's imagine a case where you want to forecast for the next four quarters
52:57
then you would ignore last points, you would include four and then see if you were disregard
53:05
the last four quarters, what would your forecast look like. So that in a way is a way for you to
53:11
understand how much error there is in your forecast. Unfortunately again Power BI does not give you
53:20
any estimate of the errors in your forecast so that is an approximate way of understanding
53:29
what it is if you were doing it in python or r for example then yes you would be able to change
53:36
the parameters like we saw in the presentation instead of additive you could go to damp trend
53:41
and ramp seasonality or maybe you could look at quantitative seasonality and things like that
53:47
so in Python under R yes you can do that in Power BI
53:51
unfortunately not so much well thank you for the information I mean you have shared a lot of information with us
53:59
lately so it's really cool to learn so much about these topics
54:04
so thank you a lot for being here with us again and I hope
54:11
you also enjoyed it a little bit and we hope we will
54:15
welcome you soon back to AI42. Thank you. So with that, maybe we can
54:23
have our closing notes. Hogan? Yes. A moment there. Yes. So thank you again, everyone, for joining us
54:33
It was amazing to see so many of you together with us
54:39
So, and we also hope that we're going to join us next time
54:43
when we're going to talk about Apache Spark in Azure Databricks, with Terry Mekken
54:48
So I think it's a really good idea to join us next time again
54:53
And then Laila is going to talk about Azure Machine Learning Workspace
54:57
and what are the capabilities of this amazing tool. I can't wait really, Håkan
55:03
Which is the favorite one of yours? I'm actually looking forward to all of them
55:09
But I'm really looking forward to the Azure ML Workspace and see what that entails
55:15
That sounds really interesting. Yes, I agree with you, Håkan, totally. So, yeah, so we hope to meet you again in two weeks on the 17th of November
55:26
And with all that, we want to say a big, big, big thank you for being with us tonight
55:31
And, yeah, see you soon again. Thank you so much. Bye-bye. Bye
#Business & Industrial
#Computers & Electronics