1
00:00:00,990 --> 00:00:05,850
In this lecture, we will learn how to include seasonality.

2
00:00:05,990 --> 00:00:07,130
No, Eddie, my model.

3
00:00:08,390 --> 00:00:11,800
Or how do you study my model in Python?

4
00:00:13,440 --> 00:00:15,920
Satima essense for seasonal Atima.

5
00:00:16,830 --> 00:00:23,600
And this model also incorporate seasonality while fitting or training or model.

6
00:00:25,240 --> 00:00:31,440
Now, one of the limitation of Eyadema is that it can only handle Ukraine in your data.

7
00:00:33,450 --> 00:00:39,750
If your data have seasonality, you first served to the IT and then use that in my model on their DePrince

8
00:00:39,750 --> 00:00:40,140
data.

9
00:00:41,880 --> 00:00:50,740
To overcome this limitation, we have Satima model where we include both grand and seasonality while

10
00:00:50,760 --> 00:00:51,320
modeling.

11
00:00:54,360 --> 00:00:58,560
So in Rhema you just have to provide beebee you variables.

12
00:00:59,190 --> 00:01:04,680
What in Satima you have to provide seven different parameters.

13
00:01:05,010 --> 00:01:12,870
First you have to broyd lowercase beebee Q for data and then a better case.

14
00:01:12,870 --> 00:01:21,570
BBC you and M for the seasonal data lowercase B essense for a number of domes.

15
00:01:22,380 --> 00:01:28,020
We want to consider for auto regression be censored number of times.

16
00:01:28,170 --> 00:01:33,480
We want to do depressing cuz sense for number of lag times.

17
00:01:33,510 --> 00:01:36,750
We want to consider for moving average.

18
00:01:37,880 --> 00:01:45,630
Now all these mortgages we need you is four grand that we haven't discussed in a remodel so.

19
00:01:49,840 --> 00:01:57,920
No, next, we have four other parameters, capital p sense for seasonal auto regressive order.

20
00:01:58,180 --> 00:02:05,140
So this is the number of times you want to consider for auto regression, for seasonality.

21
00:02:05,890 --> 00:02:15,340
So if you are giving capital P equal to one, this means that the model will also consider one seasonality

22
00:02:15,430 --> 00:02:15,810
lag.

23
00:02:17,140 --> 00:02:26,020
So if you're seasonality is of value time period or you have a yearly seasonality, it will consider

24
00:02:26,020 --> 00:02:28,600
the last year value for auto regression.

25
00:02:29,320 --> 00:02:35,320
Similarly, B is the number of defensing you want to be seasonal as your data.

26
00:02:37,780 --> 00:02:46,060
Q Is the number of seasonal times you want to include in your moving average and M is the number of

27
00:02:46,060 --> 00:02:49,120
times steps for a single seasonal period.

28
00:02:49,330 --> 00:02:57,970
So if you have a monthly data and and yearly seasonality, you have to give AM equates to an.

29
00:03:00,780 --> 00:03:06,750
If you have a daily data and you have a weekly seasonality, you have to give Amik way to seven.

30
00:03:11,330 --> 00:03:14,840
The implementation is almost similar to Eyadema.

31
00:03:16,310 --> 00:03:25,880
So instead of Eyadema, this time you have to import Satima X from his set Smardon and we will be using

32
00:03:26,450 --> 00:03:28,990
U.S. airline monthly aircraft data.

33
00:03:29,060 --> 00:03:32,990
Since this data have both grindin seasonality on.

34
00:03:32,990 --> 00:03:35,300
Let's look at the head of this data.

35
00:03:36,200 --> 00:03:37,960
The first column we have is for one.

36
00:03:38,060 --> 00:03:40,770
The second volume V is for the minus.

37
00:03:41,910 --> 00:03:44,030
Let's look at the less.

38
00:03:44,030 --> 00:03:44,270
Well.

39
00:03:47,330 --> 00:03:52,930
So we have data for on 17 years from 1963 to 1970.

40
00:03:58,190 --> 00:04:05,370
Let's change the index as well so that we can decompose this data to look at the trend and seasonality,

41
00:04:06,190 --> 00:04:11,480
we'll be using the seasonally decompose function from insects for the library.

42
00:04:14,960 --> 00:04:18,410
This will decompose our data frame and pull forward parts.

43
00:04:19,220 --> 00:04:20,660
First one is the observe data.

44
00:04:20,690 --> 00:04:23,060
You can see that on the raw data itself.

45
00:04:23,090 --> 00:04:25,170
You can see some kind seasonality.

46
00:04:28,280 --> 00:04:35,390
Then we have the trend line tech, which is showing us how the trend is changing with time.

47
00:04:36,170 --> 00:04:44,810
You can see that we can say that this is a linear crane and we only need one defensing to decode our

48
00:04:44,810 --> 00:04:45,170
data.

49
00:04:46,010 --> 00:04:47,960
Then we have the seasonality graph.

50
00:04:48,950 --> 00:04:54,200
You can see that the seasonality varies from point eight to one point two.

51
00:04:54,480 --> 00:05:01,310
That's mean if we are using multiplicative model, there is a change of 20 percent due to seasonality.

52
00:05:01,490 --> 00:05:02,450
And that's huge.

53
00:05:02,600 --> 00:05:06,020
So we can say that our data have high seasonality.

54
00:05:07,670 --> 00:05:12,380
And the next thing you can notice is that the seasonality is a little seasonality.

55
00:05:12,650 --> 00:05:14,090
So this is 1964.

56
00:05:14,120 --> 00:05:15,290
This is 1963.

57
00:05:15,320 --> 00:05:16,560
This is 1965.

58
00:05:17,250 --> 00:05:21,380
But you can see that the cycle is repeating year after year.

59
00:05:22,520 --> 00:05:25,190
And then we have that as a real graph for our data.

60
00:05:28,310 --> 00:05:35,640
So from this graph, we can say that our lower case, D. That is the defensing we need to be trained.

61
00:05:35,640 --> 00:05:44,150
The data should be one and M should be well, since we have a monthly data and then yearly seasonality.

62
00:05:48,540 --> 00:05:50,850
So this is the parameters we are going to use.

63
00:05:52,020 --> 00:05:58,200
So, no, this is the parameters that we are going to use while creating the model object.

64
00:05:58,350 --> 00:05:59,900
We have to give two parameter.

65
00:06:00,180 --> 00:06:01,500
First one is the order.

66
00:06:01,590 --> 00:06:03,360
And the second one is the seasonal order.

67
00:06:03,840 --> 00:06:06,350
Order is the same as what we give them a rhema.

68
00:06:06,690 --> 00:06:09,930
That is lowercase B, B and Cuba loose.

69
00:06:12,150 --> 00:06:18,300
And seasonal order is the additional parameter that we have to give in Satima where we have to give

70
00:06:18,360 --> 00:06:18,840
capital.

71
00:06:18,840 --> 00:06:19,890
B, capital.

72
00:06:19,900 --> 00:06:20,750
B, capital.

73
00:06:20,810 --> 00:06:21,500
Q Aleuts.

74
00:06:21,870 --> 00:06:23,220
And then the M value.

75
00:06:25,950 --> 00:06:27,990
So we are willing to train set.

76
00:06:27,990 --> 00:06:36,840
He makes more than on minus data with B as five, B as one and lowercase Q as three.

77
00:06:37,560 --> 00:06:44,310
Whereas A part is B, B, Q and M as one comma, one comma, one comma 20 respectively.

78
00:06:46,290 --> 00:06:54,200
So let's create this model object and let's for this data we have over our data.

79
00:06:54,270 --> 00:07:01,950
We can also check out the rest rules of this model so that we can say that modulo extricated all the

80
00:07:01,950 --> 00:07:04,230
information that they they in the data.

81
00:07:06,300 --> 00:07:07,140
So let's plot.

82
00:07:07,610 --> 00:07:08,220
That's real.

83
00:07:08,550 --> 00:07:16,170
You can see that the restaurants are centred around zero and that is no Grendon seasonality in the resolution.

84
00:07:18,420 --> 00:07:20,610
This that bulls are like a white noise.

85
00:07:20,640 --> 00:07:21,960
And we can say that.

86
00:07:22,260 --> 00:07:26,040
I would more than ever extracted all the information from this dataset.

87
00:07:28,840 --> 00:07:29,150
No.

88
00:07:29,330 --> 00:07:32,820
Here also, just like Remar, we can use Daudt forecast.

89
00:07:33,170 --> 00:07:37,050
My turn to get forecasts value of the next budget.

90
00:07:37,800 --> 00:07:41,780
So let's clear the variable output.

91
00:07:42,180 --> 00:07:45,040
And we are saving the forecast value of the next puttered.

92
00:07:45,120 --> 00:07:46,430
And that's all very well.

93
00:07:48,570 --> 00:07:49,490
Let's look at those.

94
00:07:49,880 --> 00:07:52,170
Well, you can see that we are getting there.

95
00:07:53,850 --> 00:07:59,580
So if you see, here's the deal.

96
00:08:00,030 --> 00:08:04,280
We have data in 1970, December.

97
00:08:05,430 --> 00:08:10,710
And now we are getting the forecast for January 1971.

98
00:08:10,800 --> 00:08:14,030
And the forecast value is eleven thousand four hundred and seventy five.

99
00:08:15,450 --> 00:08:19,190
You can also pass parameters and this forecast matter.

100
00:08:19,950 --> 00:08:25,500
So if I just write well, it will give me the forecasts of 30 days.

101
00:08:27,510 --> 00:08:30,480
You can see we have the forecast for next from this.

102
00:08:31,750 --> 00:08:37,560
Now, just like Remar, we have Dort Kraddick function as well and don't predict Fanshen.

103
00:08:37,590 --> 00:08:40,290
We have to provide a start date and end date.

104
00:08:40,800 --> 00:08:46,200
And if we don't broyd any date, it will give prediction values for all the values that are available

105
00:08:46,980 --> 00:08:47,760
in your data.

106
00:08:48,900 --> 00:08:52,860
So I am sorting this information in my head very well.

107
00:08:53,880 --> 00:09:01,960
So now we have predicted values from this model, from e're 1963 to e're 1970.

108
00:09:02,100 --> 00:09:03,500
And this variable Y ahead.

109
00:09:03,870 --> 00:09:06,420
And we have the predicted value for next year.

110
00:09:06,690 --> 00:09:11,640
And this variable that is more than underscore for Dort forecast.

111
00:09:12,780 --> 00:09:15,640
So if you just look at why had.

112
00:09:20,080 --> 00:09:28,000
You can see that we have the forecast said value for first five records here using a lot more than that.

113
00:09:28,590 --> 00:09:33,510
Now, we can also plot this forecast set value against the actual values.

114
00:09:33,950 --> 00:09:41,690
So we are winning put Lord D.F. Mind, Simon, and we are plotting way ahead on the same plot using

115
00:09:41,690 --> 00:09:42,220
that killer.

116
00:09:43,370 --> 00:09:45,610
So let's look at the plot.

117
00:09:46,580 --> 00:09:50,450
Well, you can see this blue line are the oldest known values.

118
00:09:50,540 --> 00:09:53,830
And this red line is the predicted value using the word model.

119
00:09:54,920 --> 00:09:57,830
The blue line and the red line almost mowing.

120
00:09:58,190 --> 00:09:58,680
And then.

121
00:09:59,540 --> 00:10:03,530
And that is not much variation in two values.

122
00:10:06,380 --> 00:10:09,580
So that Sölvi Element, Satima and Biton.

123
00:10:10,370 --> 00:10:12,590
And we predict values using ceremony more.

124
00:10:15,470 --> 00:10:18,560
We can also use walk forward validation in Satima.

125
00:10:19,610 --> 00:10:24,590
We already have executed walk forward validation for each other.

126
00:10:24,830 --> 00:10:30,590
And remember that you can use the same code for setting my model as well.

127
00:10:30,980 --> 00:10:34,700
Just do some necessary changes in the model creation.

128
00:10:37,220 --> 00:10:39,470
So that's all for this lecture.

129
00:10:39,830 --> 00:10:40,280
Thank you.
