1
00:00:11,020 --> 00:00:16,180
So in this lecture, we will be starting the next section of this course, which is on Facebook's private

2
00:00:16,180 --> 00:00:16,810
library.

3
00:00:17,410 --> 00:00:20,080
I really like this package for various reasons.

4
00:00:20,710 --> 00:00:25,250
The first reason is that you'll see it's a package built with business data sets in mind.

5
00:00:25,660 --> 00:00:31,270
So really, anyone who runs a business or works for a business as a data scientist can derive value

6
00:00:31,270 --> 00:00:32,370
from Facebook profit.

7
00:00:33,070 --> 00:00:35,800
For example, you can track and forecast sales.

8
00:00:36,040 --> 00:00:38,430
You can track and forecast CPU usage.

9
00:00:38,710 --> 00:00:42,030
You can track and forecast your advertising spend and so forth.

10
00:00:42,520 --> 00:00:46,290
If you are running a business, then likely you have something to count.

11
00:00:46,630 --> 00:00:53,200
If that count varies over time, then you have a time series and you can apply these methods in addition,

12
00:00:53,200 --> 00:00:54,890
because it's meant for business data.

13
00:00:55,090 --> 00:01:00,850
It's very easy to take into account different factors that businesses deal with, like weekends, holidays

14
00:01:00,850 --> 00:01:03,290
and even special events like the Super Bowl.

15
00:01:03,850 --> 00:01:07,300
It's also very easy to add your own custom exogenous inputs.

16
00:01:07,750 --> 00:01:13,270
For example, you might add a special input to the NFL Sundays, which you're not every Sunday, but

17
00:01:13,300 --> 00:01:15,520
only for specific portions of the year.

18
00:01:16,090 --> 00:01:21,310
So if you're a consultant or a freelancer and your client needs a time series, forecasting solution,

19
00:01:21,520 --> 00:01:23,320
profit can be an excellent choice.

20
00:01:28,010 --> 00:01:34,040
OK, so the number two reason I like this package, it's very easy to use with just a few lines of code

21
00:01:34,190 --> 00:01:38,660
and it's relatively fast, you'll find that models like Arima can struggle.

22
00:01:38,840 --> 00:01:44,720
When your time series is long with profit, you'll find that most things can be accomplished with just

23
00:01:44,720 --> 00:01:50,030
a single function call even cross-validation, which, as we've seen, can take some work when you write

24
00:01:50,030 --> 00:01:50,800
it yourself.

25
00:01:55,460 --> 00:01:59,280
The number three reason I like this package is because it's very interpretable.

26
00:01:59,930 --> 00:02:03,860
The model used by profit is most reminiscent of exponential smoothing.

27
00:02:04,070 --> 00:02:09,650
If you want to compare it to the other models we've studied, essentially your model can be broken down

28
00:02:09,650 --> 00:02:16,100
into error, trend and seasonality after fitting a model profit gives you the tools to automatically

29
00:02:16,100 --> 00:02:19,550
plot these components again with just a single line of code.

30
00:02:20,120 --> 00:02:25,430
But unlike its profit, can actually model seasonality at multiple frequencies.

31
00:02:25,880 --> 00:02:30,350
So you might have a pattern at the weekly level and you might have a pattern at the yearly level.

32
00:02:30,980 --> 00:02:35,950
For example, sales might slow down on Sundays, but sales might increase on Black Friday.

33
00:02:36,530 --> 00:02:41,900
So one is an example of a weekly recurring pattern and the other is an example of a yearly recurring

34
00:02:41,900 --> 00:02:42,520
pattern.

35
00:02:47,110 --> 00:02:53,320
OK, so I hope you're convinced that Facebook profit is a valuable addition to your time series toolkit,

36
00:02:53,800 --> 00:02:59,410
if you're a consultant or a freelancer, your clients will truly appreciate such a simple, accurate

37
00:02:59,410 --> 00:03:00,970
and interpretable solution.

38
00:03:01,390 --> 00:03:05,950
If you run your own business, then I'm sure you'll find that these features are useful for yourself

39
00:03:05,950 --> 00:03:06,420
as well.