1
00:00:11,220 --> 00:00:16,710
So in this video, we are going to summarize everything we learned in this section, this section was

2
00:00:16,710 --> 00:00:20,690
about how to apply machine learning methods, two time series analysis.

3
00:00:21,270 --> 00:00:26,070
Basically, the purpose of this section was not to discuss any particular machine learning method in

4
00:00:26,070 --> 00:00:26,580
depth.

5
00:00:27,180 --> 00:00:33,690
Instead, the main goal was to understand how to convert time series data into a format which is usable

6
00:00:33,990 --> 00:00:39,690
by supervised machine learning methods, although we did discuss the intuition behind the models we

7
00:00:39,690 --> 00:00:40,200
used.

8
00:00:40,410 --> 00:00:44,100
I hope that your attention was on how to apply those models in the code.

9
00:00:48,840 --> 00:00:54,600
What a nice side effect of using generic machine learning models is that we can do multi output multi-step

10
00:00:54,600 --> 00:00:55,350
forecasts.

11
00:00:55,830 --> 00:01:00,690
This makes it easier to predict multiple time steps ahead by making all the predictions at the same

12
00:01:00,690 --> 00:01:01,170
time.

13
00:01:02,250 --> 00:01:07,800
We learn that this can also improve forecast performance by avoiding error propagation that can occur

14
00:01:08,040 --> 00:01:10,920
when you try to make predictions from other predictions.

15
00:01:11,550 --> 00:01:17,100
We also learn that by using machine learning, it's just as easy to do classification as it is to do

16
00:01:17,100 --> 00:01:18,000
a forecast.

17
00:01:18,870 --> 00:01:24,330
With stock prices in particular, one common task is to simply do classification instead of trying to

18
00:01:24,330 --> 00:01:25,660
predict an exact number.

19
00:01:26,310 --> 00:01:28,020
Although this task is much simpler.

20
00:01:28,260 --> 00:01:32,820
We found that it is still difficult to do much better than predicting the result of a coin toss.

21
00:01:33,690 --> 00:01:38,250
In other words, this also lends more evidence to support the random walk hypothesis.

22
00:01:42,880 --> 00:01:47,890
The main insight you should learn from this section is that it's very easy to build an auto regressive,

23
00:01:48,070 --> 00:01:54,160
non-linear machine learning forecaster that makes it more powerful than the simple linear methods we

24
00:01:54,160 --> 00:01:55,250
learned about before.

25
00:01:56,080 --> 00:02:00,220
Now, whether or not that flexibility is useful is dependent on the data.

26
00:02:00,880 --> 00:02:03,820
The data will decide which technique should be applied.

27
00:02:04,300 --> 00:02:07,660
So use these techniques wisely and don't forget to benchmark.

28
00:02:08,140 --> 00:02:11,830
If a linear model gives you the best error, then that is the best choice.

29
00:02:12,310 --> 00:02:17,650
However, it's always worth trying these non-linear methods in case you can achieve superior results.