1
00:00:01,350 --> 00:00:04,140
How far in the future should we forecast?

2
00:00:06,180 --> 00:00:10,200
Should we generate all the forecasts at a single time point?

3
00:00:11,020 --> 00:00:14,250
Or can we forecast on an ongoing basis?

4
00:00:16,350 --> 00:00:20,190
These are important questions to be answered at the gold definition stage.

5
00:00:21,900 --> 00:00:28,860
But before we try to find these answers, let us discuss some basic notation that we are going to use.

6
00:00:28,950 --> 00:00:29,340
And this.

7
00:00:31,270 --> 00:00:40,590
We are going to use four types of symbols to denote time period data as forecasts and forecast errors.

8
00:00:42,900 --> 00:00:46,170
So B will be used to denote the time period.

9
00:00:47,830 --> 00:00:54,950
This is equal to one is the first time period indices and these equal to N means d end it.

10
00:00:55,050 --> 00:01:04,040
Time period, the whity will be used to denote the actual value of the variable of interest in the pit

11
00:01:04,170 --> 00:01:04,740
time period.

12
00:01:06,420 --> 00:01:14,820
For example, if we start measuring daily average temperatures at the start of the week, then these

13
00:01:14,820 --> 00:01:22,980
equal to one denotes day one or Monday B is equal to two, denotes day two or Tuesday and so on.

14
00:01:24,420 --> 00:01:28,900
And Vivan denote the average temperature of Monday.

15
00:01:30,240 --> 00:01:30,690
And why?

16
00:01:30,690 --> 00:01:33,840
To denote the average temperature of Tuesday and so on.

17
00:01:36,240 --> 00:01:37,340
Next is F.T..

18
00:01:38,130 --> 00:01:41,180
This will be the forecasted value for time period.

19
00:01:42,990 --> 00:01:44,100
Not the actual value.

20
00:01:44,450 --> 00:01:46,320
Actual value is denoted by whity.

21
00:01:47,160 --> 00:01:49,850
Forecasted value will be denoted by F.T..

22
00:01:51,360 --> 00:01:59,050
Similarly, F.T. plus K will then denote D gay step ahead forecast when the forecasting time.

23
00:02:01,200 --> 00:02:08,160
If we are currently a time Beretti, the forecast for the next time period that is D plus one is denoted

24
00:02:08,160 --> 00:02:09,600
by F people as one.

25
00:02:10,230 --> 00:02:17,280
So if you forecast the average temperature of Wednesday on a Monday, you are forecasting for peoples

26
00:02:17,310 --> 00:02:18,600
tuart time period.

27
00:02:22,050 --> 00:02:25,940
Lastly, E.T. is the forecast error for time purity.

28
00:02:27,570 --> 00:02:32,480
That is it is the difference between the actual value and defocused value at 90.

29
00:02:33,870 --> 00:02:38,460
So mathematically, it will be equal to Viotti minus F.T..

30
00:02:40,200 --> 00:02:43,530
This is basically the inaccuracy of our forecast.

31
00:02:46,350 --> 00:02:54,570
With this notation, let us denote the forecast horizon with key basically gauge the number of time

32
00:02:54,570 --> 00:02:57,900
period ahead that we will be forecasting.

33
00:02:58,950 --> 00:03:05,430
For example, if Gaitskell equal three four monthly data, this means that we will be forecasting for

34
00:03:05,430 --> 00:03:09,550
the next three months in our Amtrak ridership.

35
00:03:09,570 --> 00:03:17,970
Example, one month Ed forecast, that is if people's money will be sufficient for revenue management.

36
00:03:18,420 --> 00:03:20,520
That is for flexible pricing.

37
00:03:22,320 --> 00:03:32,520
Whereas longer term forecast such as dreman three months ahead forecast or F, P plus three are more

38
00:03:32,520 --> 00:03:36,300
likely to be needed for procurement and inventory purposes.

39
00:03:37,980 --> 00:03:42,150
Knowing forecast horizon, as does an important part of goal setting.

40
00:03:43,710 --> 00:03:47,830
One of the factors that affect forecast horizon is the decency of data available.

41
00:03:49,050 --> 00:03:55,360
Forecasting next month's ridership is much harder if we do not have the data for last two months.

42
00:03:56,910 --> 00:04:01,740
If we do not have the data for last two months, it actually means that we are generating forecasts

43
00:04:02,550 --> 00:04:06,210
of F.P. plus three rather than F.T. plus one.

44
00:04:09,850 --> 00:04:18,250
Now, what is the impact of large or small forecast horizon on forecasting accuracy the further into

45
00:04:18,250 --> 00:04:19,000
the future?

46
00:04:19,120 --> 00:04:26,140
We are trying to forecast the more likely that the forecasting context will change and certainty will

47
00:04:26,140 --> 00:04:26,680
increase.

48
00:04:27,340 --> 00:04:36,940
Therefore, for larger reason, we can expect large uncertainty and less accuracy for such a situation.

49
00:04:37,480 --> 00:04:45,640
We should probably examine our model and try and updated, as in when new data or information is available.

50
00:04:48,510 --> 00:04:57,270
For example, if we have a three month ahead forecast for April of 2020, which we have generated in

51
00:04:57,330 --> 00:05:06,110
January 2020, if possible, we should update this forecast in February and again in March of this year.

52
00:05:09,610 --> 00:05:14,080
Lastly, we will discuss two points, forecasting use and level of automation.

53
00:05:16,060 --> 00:05:17,920
How will the forecast be used?

54
00:05:18,910 --> 00:05:26,080
Understanding how forecasts will be used, perhaps by different stakeholders is critical for generating

55
00:05:26,080 --> 00:05:27,600
forecasts of the right type.

56
00:05:28,060 --> 00:05:32,980
And with a useful accuracy limit, consider this question.

57
00:05:33,580 --> 00:05:39,250
Does over prediction cost more or it cost less than the under prediction?

58
00:05:40,540 --> 00:05:46,510
For example, if you're organizing an event and you want to forecast how many people will attend it,

59
00:05:48,610 --> 00:05:56,200
if you all predict and arrange for 120 people instead of hundreds who actually turn up, there will

60
00:05:56,200 --> 00:05:58,780
be some cost incurred for 20 empty seats.

61
00:06:00,460 --> 00:06:07,270
On the other hand, if you under predict and arrange for 80 people only, you will lose revenue of 20

62
00:06:07,270 --> 00:06:07,720
people.

63
00:06:08,350 --> 00:06:14,500
And there will be some emotional loss also for those 20 people who came but were not allowed to enter.

64
00:06:16,570 --> 00:06:23,620
Similarly, this question will be forecast and forecasting method is to be presented to management or

65
00:06:23,710 --> 00:06:30,940
to the technical department, answer to this question will tell us the level of complexity in our model

66
00:06:31,030 --> 00:06:31,960
that we can use.

67
00:06:32,740 --> 00:06:36,430
Similarly, we ask, should the forecast be numerical or binary?

68
00:06:37,660 --> 00:06:44,740
For example, if you're forecasting whether it will rain or not tomorrow, it is a binary forecast.

69
00:06:46,660 --> 00:06:51,820
On the other hand, if you are forecasting the average temperature for the next day, that is a numerical

70
00:06:51,820 --> 00:06:52,360
forecast.

71
00:06:54,040 --> 00:06:56,530
Different models work for these two types of problems.

72
00:06:57,640 --> 00:07:03,940
We will cover both types of models then, of course, but you must know which one is to be applied for

73
00:07:03,940 --> 00:07:04,900
the problem at hand.

74
00:07:06,100 --> 00:07:11,950
Next is a set of questions to determine the level of automation required in the forecasting project.

75
00:07:14,170 --> 00:07:20,170
The level of required automation depends on the nature of forecasting tasks and on how the forecast

76
00:07:20,290 --> 00:07:21,460
will be used in practice.

77
00:07:23,140 --> 00:07:27,690
For example, when we ask how many cities need to be forecasted.

78
00:07:28,240 --> 00:07:35,800
If for us in Adio there are many cities which are do we forecasted and not much forecasting expertise

79
00:07:35,800 --> 00:07:36,820
in the organization.

80
00:07:37,210 --> 00:07:38,890
We must automate the solution.

81
00:07:40,780 --> 00:07:48,430
A classic example is forecasting at the point of sale for purpose of inventory control across many stores.

82
00:07:50,530 --> 00:07:55,540
Organizations do not depend on these salesperson depuis us to manage inventory.

83
00:07:56,910 --> 00:07:59,290
It is, it added, managed automatically by the system.

84
00:08:00,610 --> 00:08:03,640
Several consulting firms provide such software solutions.

85
00:08:07,040 --> 00:08:11,910
The question is, is default casting an ongoing process or a one time event?

86
00:08:13,020 --> 00:08:16,530
If it is a one time event, we need not implement automation.

87
00:08:17,820 --> 00:08:22,800
We also might want to know which daytimes software will be available during the forecasting period.

88
00:08:24,150 --> 00:08:30,600
Word forecasting expertise will be available at the organization during forecasting period because lack

89
00:08:30,600 --> 00:08:32,610
of expertise favours automation.

90
00:08:33,390 --> 00:08:39,900
So if we do not have a lot of expertise in the organization, we may want to automate the forecasting

91
00:08:39,900 --> 00:08:43,950
process so that it does not depend on deep people in the organization.

92
00:08:45,900 --> 00:08:53,190
So you can see different answers will lead to different choices of data forecasting methods and evolution

93
00:08:53,190 --> 00:08:53,790
schemes.

94
00:08:55,980 --> 00:09:00,990
And all these questions must be answered at the gold definition stage.
