1 00:00:11,080 --> 00:00:16,330 So in this lecture, we are going to do stock predictions once again, but this time our objective is 2 00:00:16,330 --> 00:00:17,340 slightly different. 3 00:00:17,980 --> 00:00:23,530 Instead of trying to forecast the return and effectively the price, we are going to simply try to predict 4 00:00:23,530 --> 00:00:24,380 the direction. 5 00:00:24,880 --> 00:00:26,680 So that should make the task much easier. 6 00:00:27,310 --> 00:00:32,050 Instead of trying to predict a number which has an infinite number of choices, we're just going to 7 00:00:32,050 --> 00:00:34,640 predict up or down, which is two choices. 8 00:00:35,350 --> 00:00:37,520 Furthermore, to make this even easier. 9 00:00:37,570 --> 00:00:40,390 We will only try to predict one step ahead. 10 00:00:40,900 --> 00:00:43,820 So this is effectively the easiest task possible. 11 00:00:44,380 --> 00:00:46,010 Let's see how machine learning will do. 12 00:00:50,760 --> 00:00:56,050 OK, so notice that for our imports, we are now importing classifiers and not aggressives. 13 00:00:59,790 --> 00:01:04,110 OK, so, again, we're going to download our data and choose IBM as our stock. 14 00:01:09,890 --> 00:01:15,680 OK, so the new part here is how we create the data set, so for this data set, we're again going to 15 00:01:15,680 --> 00:01:19,940 use a stationary time series as input, which comes from the log returns. 16 00:01:20,840 --> 00:01:24,890 The target is simply whether the log return is positive or negative. 17 00:01:25,670 --> 00:01:28,970 An easy way to compute there is to use greater than zero. 18 00:01:29,630 --> 00:01:35,330 Multiplying by one is technically not necessary, but I like my targets to be zero and one instead of 19 00:01:35,330 --> 00:01:36,240 true and false. 20 00:01:37,430 --> 00:01:40,050 Notice how I'm using 21 lags for this script. 21 00:01:40,580 --> 00:01:46,070 This is just to test the idea that maybe we can get better predictions if we use more lax. 22 00:01:49,880 --> 00:01:55,090 OK, so basically the only difference in our loop now is that when we want to get the target, the target 23 00:01:55,090 --> 00:01:57,170 no longer comes from the Time series itself. 24 00:01:57,550 --> 00:01:59,410 Instead, it has been pre computed. 25 00:02:00,130 --> 00:02:02,350 Other than that, it's the same as before. 26 00:02:07,890 --> 00:02:11,610 OK, so the next step is to split our data into train and test. 27 00:02:16,200 --> 00:02:18,540 The next step is to train a logistic regression. 28 00:02:22,150 --> 00:02:26,410 OK, so the train accuracy is only about 50 percent, not great. 29 00:02:30,330 --> 00:02:36,780 And the test accuracy is also only about 50 percent, so pretty much just as hard as predicting a coin 30 00:02:36,780 --> 00:02:37,380 toss. 31 00:02:40,880 --> 00:02:43,310 The next step is to train a support vector machine. 32 00:02:47,010 --> 00:02:51,140 So this does better on the train set, but let's see whether or not it does better on the tests. 33 00:02:54,930 --> 00:03:01,050 Unfortunately, it does not do better on the test set so we can conclude that the model just over fits. 34 00:03:03,960 --> 00:03:06,090 The next step is to try the random forest. 35 00:03:10,220 --> 00:03:15,770 OK, so the random forest gets 100 percent on the train set, which is not unsurprising given what we 36 00:03:15,770 --> 00:03:16,340 have seen. 37 00:03:20,770 --> 00:03:23,870 But the test set is, again, very close to 50 percent. 38 00:03:24,400 --> 00:03:26,560 So, again, just in overfit model. 39 00:03:27,830 --> 00:03:33,770 OK, so what we can conclude from this is that the linear model was the only model that did not overfit. 40 00:03:35,700 --> 00:03:40,830 Now, again, I want to mention that the exercises and things you can try in this course are nearly 41 00:03:40,830 --> 00:03:41,430 endless. 42 00:03:42,150 --> 00:03:46,580 You might be wondering what if I included other columns such as open high in volume? 43 00:03:47,160 --> 00:03:49,540 What if I included the returns of other stocks? 44 00:03:50,040 --> 00:03:52,230 So these are all interesting things to try. 45 00:03:52,620 --> 00:03:56,610 You should be able to test out these ideas with pretty much no new code. 46 00:03:57,600 --> 00:04:01,220 So, as always, keep trying new things and I'll see you in the next lecture.