1
00:00:11,070 --> 00:00:16,680
So in this lecture, we will be introducing the next section of this course, which is on a Financial

2
00:00:16,680 --> 00:00:19,110
Times series technique known as Gargash.

3
00:00:19,800 --> 00:00:21,570
So what's the story behind Gurche?

4
00:00:22,530 --> 00:00:27,480
Well, I've said pretty often that in finance it's common for people who are just starting out to ask

5
00:00:27,480 --> 00:00:30,300
the question, how do I predict stock prices?

6
00:00:30,870 --> 00:00:32,400
In most finance courses?

7
00:00:32,400 --> 00:00:34,900
You quickly learn that this is not the right question.

8
00:00:35,580 --> 00:00:39,250
The correct question is how do I predict stock returns?

9
00:00:39,720 --> 00:00:41,230
Now, that's just the first step.

10
00:00:41,670 --> 00:00:46,940
You see, what happens when you try to predict stock returns is that you find it is pretty hard.

11
00:00:47,850 --> 00:00:53,100
The next step is to consider the question, well, why are stock returns so hard to predict?

12
00:00:53,700 --> 00:00:57,090
The answer is that they very closely resemble a random walk.

13
00:00:57,630 --> 00:01:04,160
In fact, random walks are so accurate that we even use them in finance, for example, to price options.

14
00:01:04,680 --> 00:01:08,150
There is even a book written about stock prices being random walks.

15
00:01:08,910 --> 00:01:13,950
Unfortunately, one characteristic of random stocks is that they are completely unpredictable.

16
00:01:18,620 --> 00:01:24,160
Now, it turns out that there are some attributes of stock returns that make them unlike random walks,

17
00:01:24,770 --> 00:01:29,220
one of those attributes has to do with what we call volatility clustering.

18
00:01:29,960 --> 00:01:31,190
So what does this mean?

19
00:01:32,000 --> 00:01:38,010
Well, notice how at certain points in time we have very large returns, both positive and negative.

20
00:01:38,720 --> 00:01:44,870
One thing to notice about these large returns is that they tend to be surrounded by other large returns.

21
00:01:45,290 --> 00:01:48,280
In other words, large movements tend to cluster together.

22
00:01:49,040 --> 00:01:55,070
So Financial Times series is really all about figuring out what you cannot model and then asking new

23
00:01:55,070 --> 00:01:57,210
questions to figure out what you can.

24
00:01:57,920 --> 00:02:03,320
We know that it's very difficult to predict returns, but clearly if we consider the variance of the

25
00:02:03,320 --> 00:02:07,040
returns, there seems to be some pattern that may be explainable.

26
00:02:11,740 --> 00:02:14,450
OK, so let's go through a quick outline for this section.

27
00:02:15,190 --> 00:02:19,570
This section will start out with the arch model, which is like an auto regressive model of squared

28
00:02:19,570 --> 00:02:20,230
errors.

29
00:02:21,040 --> 00:02:27,040
Now, the reason we want to study art first is because Gurche seems to be like a Rimma for variance.

30
00:02:27,430 --> 00:02:32,660
But there are a few details we need to consider if we really want to understand this relationship.

31
00:02:33,340 --> 00:02:37,340
The next step is to study Darch, which adds one more component to Arch.

32
00:02:38,410 --> 00:02:43,650
The next step, as usual, is to try everything in code so you can see how it works.

33
00:02:44,290 --> 00:02:48,350
We'll be using the arch library in Python, although there are others available.

34
00:02:49,120 --> 00:02:54,430
Well, then go back into theory mode and discuss how Gargash models might be improved, for instance,

35
00:02:54,430 --> 00:02:56,030
using deep neural networks.

36
00:02:56,590 --> 00:03:01,450
I'll go through a conceptual derivation of how you might apply deep learning to Gurche, although we

37
00:03:01,450 --> 00:03:04,100
won't be implementing these concepts at this time.

38
00:03:04,900 --> 00:03:10,450
However, note that I have done examples like this in the past, so feel free to send me an email if

39
00:03:10,450 --> 00:03:11,470
you want to learn more.

40
00:03:12,100 --> 00:03:15,310
They do require more advanced knowledge of deep learning libraries.

41
00:03:15,880 --> 00:03:21,120
The reason I mention this now is because it should give you a better understanding of how Gargash works,

42
00:03:21,370 --> 00:03:23,790
even though deep learning is not a large topic.

43
00:03:24,220 --> 00:03:28,620
So the purpose of this conceptual exercise isn't to actually do deep learning.

44
00:03:28,870 --> 00:03:33,800
It's more to enhance your understanding of garbage once we're done.

45
00:03:33,820 --> 00:03:39,370
We will summarize this section and I'll give you a preview of some other ways Gargash has been extended.

46
00:03:39,910 --> 00:03:43,630
There are many variations of garbage and it really could be a course on its own.

47
00:03:44,170 --> 00:03:49,060
So if you would be interested in such a course that goes in-depth into Gargash, just let me know when

48
00:03:49,070 --> 00:03:51,730
my suggestion box or contact me by email.