1
00:00:11,090 --> 00:00:16,250
So in this lecture, we're going to summarize everything we learned in this section, this section was

2
00:00:16,250 --> 00:00:22,250
all about how to use Adewusi forecast, which is an industrial strength time series, forecasting products.

3
00:00:22,730 --> 00:00:26,960
In this section, you learned about the data model that is used by US forecast.

4
00:00:27,350 --> 00:00:32,030
In some ways, it's much more complicated than what we'd been working with in this course prior.

5
00:00:32,480 --> 00:00:38,540
Instead of just a data frame or an umpire, we had to build covid files with a specific schema to fit

6
00:00:38,540 --> 00:00:40,040
what AWI expects.

7
00:00:40,490 --> 00:00:43,760
In other ways, this makes HWC forecasts very flexible.

8
00:00:44,420 --> 00:00:49,600
We know that adding more features or more items is merely just a matter of adding more CSV files.

9
00:00:50,060 --> 00:00:53,330
So there's a tradeoff between flexibility and complexity.

10
00:00:53,960 --> 00:00:58,820
We learn that in order to train the model, it wasn't just a matter of passing our data and do some

11
00:00:58,820 --> 00:00:59,560
function.

12
00:01:00,020 --> 00:01:04,000
Instead, there were several steps that took a few minutes of just waiting.

13
00:01:04,670 --> 00:01:08,060
We had to create data set groups which then held data sets.

14
00:01:08,450 --> 00:01:13,770
There were different kinds of data sets like Target Time series and related time series and metadata.

15
00:01:14,300 --> 00:01:19,400
There were also different data set domains so that the model could be more tailored to your particular

16
00:01:19,400 --> 00:01:20,260
application.

17
00:01:20,780 --> 00:01:26,030
For example, retail forecasting, supply chain forecasting and web traffic forecasting.

18
00:01:26,930 --> 00:01:33,530
We learned about the six different model types available, including a rhema profit CNN, QR and Depay

19
00:01:33,640 --> 00:01:35,810
Plus, which was the one we used.

20
00:01:37,280 --> 00:01:42,710
Finally, we learned how to create a predictor and how to use modern features like Auto Amelle and hyper

21
00:01:42,710 --> 00:01:48,590
parameter optimization, once our predictor was trained, we saw how to use the predictor to make a

22
00:01:48,590 --> 00:01:49,410
forecast.

23
00:01:49,910 --> 00:01:55,100
We saw that the forecast was actually pretty good, although we did only try a single time series.

24
00:01:59,800 --> 00:02:04,540
To close this section, I want to leave you with a few thoughts about how this solution compares to

25
00:02:04,540 --> 00:02:05,950
the other models we've studied.

26
00:02:06,580 --> 00:02:11,740
One thing we didn't explore was whether or not the predictor can make forecasts outside of the specified

27
00:02:11,740 --> 00:02:12,490
time window.

28
00:02:13,030 --> 00:02:16,030
I encourage you to try this on your own, to verify what I'm saying.

29
00:02:16,280 --> 00:02:18,400
But in fact, this is not possible.

30
00:02:19,040 --> 00:02:24,590
As you recall, at one point we defined our forecast horizon and then pass this into our predictor.

31
00:02:25,120 --> 00:02:29,700
So if you want to forecast beyond that window, you want to be able to use the same model.

32
00:02:30,190 --> 00:02:34,360
In fact, if you want to make enough sample predictions, that's also not possible.

33
00:02:36,170 --> 00:02:41,210
This brings me to my next point, which is something worth considering, you actually have no access

34
00:02:41,210 --> 00:02:42,470
to the model at all.

35
00:02:43,130 --> 00:02:47,870
This is really a solution that seems like it's meant for the enterprise crowd rather than practicing

36
00:02:47,870 --> 00:02:48,870
data scientists.

37
00:02:49,340 --> 00:02:55,070
In fact, Amazon advertises that this service can be used without any machine learning experience.

38
00:02:55,490 --> 00:03:01,280
So it's not surprising that the only interface between you and the model is uploading your data to S3

39
00:03:01,490 --> 00:03:03,850
and launching a few jobs using Boso three.

40
00:03:04,580 --> 00:03:10,880
If you did have access to the model, it would be easy to do things like in sample predictions or forecasting

41
00:03:10,880 --> 00:03:12,680
an arbitrary number of time steps.

42
00:03:13,220 --> 00:03:18,980
This is a real disadvantage because it makes it more difficult to debug your model if you want to apply

43
00:03:18,980 --> 00:03:20,500
explainable A.I. techniques.

44
00:03:20,630 --> 00:03:23,060
You can't do that without access to the model.

45
00:03:27,960 --> 00:03:33,330
One practical thing you have to consider is this if you're using this in production and you need to

46
00:03:33,330 --> 00:03:36,540
generate forecasts pretty frequently, what does this entail?

47
00:03:37,230 --> 00:03:42,810
Well, if you did have access to the model, you just call model predict or some other analogous python

48
00:03:42,810 --> 00:03:43,370
function.

49
00:03:44,010 --> 00:03:47,520
But with US forecasts, we've seen that this is not possible.

50
00:03:48,420 --> 00:03:51,670
Instead, you just have to go through the entire process again.

51
00:03:52,470 --> 00:03:55,940
So imagine that you need to generate a new forecast every week.

52
00:03:56,460 --> 00:03:59,870
Then you would have to upload all the data for the most recent week.

53
00:04:00,330 --> 00:04:03,340
Then you have to add your new data set to the data set group.

54
00:04:03,840 --> 00:04:07,320
Then you have to create your new predictor, which might take a few hours.

55
00:04:08,130 --> 00:04:11,780
And only after you've done all that work can you generate a new forecast.

56
00:04:12,090 --> 00:04:15,230
And this model you trained is only good for this week.

57
00:04:15,930 --> 00:04:18,620
Next week you have to go through the same process again.

58
00:04:19,290 --> 00:04:21,090
And remember, this all costs money.

59
00:04:21,540 --> 00:04:27,270
I encourage you to check out the AWB forecast pricing page to see how much this really costs in terms

60
00:04:27,270 --> 00:04:28,520
of dollars and cents.

61
00:04:29,070 --> 00:04:35,240
So if you want to use AWB forecast, it's not just a one time cost, but rather an ongoing cost.

62
00:04:35,610 --> 00:04:38,640
That is, if you want to make forecasts on an ongoing basis.

63
00:04:39,240 --> 00:04:42,660
Then again, this applies to essentially all of Amazon's services.

64
00:04:43,170 --> 00:04:48,240
At the same time, it is a powerful solution and it's really excellent for those who have programming

65
00:04:48,240 --> 00:04:50,820
experience, but no machine learning expertise.