1 00:00:11,060 --> 00:00:16,250 OK, so in this lecture, we are going to look at applying the whole Winsor's model to a new data center, 2 00:00:16,520 --> 00:00:18,170 specifically stock prices. 3 00:00:18,740 --> 00:00:23,840 Now, before we begin this lecture, I want to give you an opportunity to stop this video and to try 4 00:00:23,840 --> 00:00:24,970 to do this on your own. 5 00:00:25,580 --> 00:00:30,440 Since you've already learned all the code you need to complete this exercise, you technically do not 6 00:00:30,440 --> 00:00:31,200 need my help. 7 00:00:31,760 --> 00:00:37,520 So if you want to treat this as an exercise, please download the data set from the euro given in this 8 00:00:37,520 --> 00:00:38,120 notebook. 9 00:00:38,420 --> 00:00:42,590 But after you get the euro, close the notebook and build a model on your own. 10 00:00:43,250 --> 00:00:47,130 OK, so if you want to try this as an exercise, please do so now. 11 00:00:47,330 --> 00:00:48,830 Otherwise, let's continue. 12 00:00:49,770 --> 00:00:52,910 OK, so let's skip the imports since you have seen this all before. 13 00:00:56,690 --> 00:01:02,870 The next step is to download our data, so clearly this is a CSFI of stock prices for stocks in the 14 00:01:02,870 --> 00:01:04,220 S&P 500. 15 00:01:08,310 --> 00:01:11,910 Well, now run the head commands in order to check what's inside our CSFI. 16 00:01:16,740 --> 00:01:22,860 Note that unlike a typical stock price, GSV, this contains multiple tickers at once, you can see 17 00:01:22,860 --> 00:01:26,910 that it's essentially multiple ticker's concatenated in the same data frame. 18 00:01:27,430 --> 00:01:29,610 You can tell them apart by the name column. 19 00:01:34,040 --> 00:01:37,690 The next step is to load in our data using Pediatrics V.. 20 00:01:41,970 --> 00:01:45,630 The next step is to call the head to see what's in our data frame. 21 00:01:49,710 --> 00:01:51,390 OK, so nothing unexpected. 22 00:01:54,680 --> 00:01:57,480 The next step is to grab the clothes prices for Google. 23 00:01:57,860 --> 00:02:00,350 Please feel free to choose your own ticker if you like. 24 00:02:04,280 --> 00:02:07,660 The next step is to call the plot function to see what our data looks like. 25 00:02:11,330 --> 00:02:15,020 OK, so we see that Google is performing very well in the stock market. 26 00:02:18,340 --> 00:02:23,950 The next step is to take the log transform of the close price, we'll also plot this new column to see 27 00:02:23,950 --> 00:02:24,850 what it looks like. 28 00:02:30,470 --> 00:02:33,410 Interestingly, we can now see a pretty linear trend. 29 00:02:37,280 --> 00:02:40,700 So you'll notice that I've commented on this line, which does not work. 30 00:02:41,960 --> 00:02:46,040 So the problem with stock prices is they only exist for trading days. 31 00:02:46,520 --> 00:02:49,370 However, trading days are not evenly spaced in time. 32 00:02:49,940 --> 00:02:54,470 Furthermore, they don't correspond with business days, which is the closest thing in panties. 33 00:02:55,460 --> 00:02:57,580 Unfortunately, this line throws an error. 34 00:02:57,830 --> 00:03:01,730 So we'll just have to make do without having a frequency for the index. 35 00:03:03,740 --> 00:03:09,560 The next step is to split our data into training tests I've chosen and testicles 30, which might be 36 00:03:09,560 --> 00:03:10,160 a bit long. 37 00:03:10,160 --> 00:03:12,640 So feel free to change this to whatever you like. 38 00:03:16,450 --> 00:03:18,180 OK, so you've seen this before. 39 00:03:23,360 --> 00:03:28,950 OK, so the next step is to instantiate our model, since I don't believe this data has any seasonality. 40 00:03:29,090 --> 00:03:31,470 I've said Seasonale to none again. 41 00:03:31,490 --> 00:03:34,580 Feel free to test different parameters and see what works best. 42 00:03:37,480 --> 00:03:42,730 Note that we technically didn't have to take the log ourselves, since that's models can do this internally, 43 00:03:43,390 --> 00:03:48,490 also observe that we get a warning since our data frame index doesn't have a frequency. 44 00:03:51,490 --> 00:03:54,970 The next step is to assign our predictions to the Google data frame. 45 00:03:58,470 --> 00:04:02,550 For the test set, you see that I've converted the result into an umpire, Ray. 46 00:04:03,330 --> 00:04:09,390 This is because not having a frequency for the index messes up the forecast indices and this will mess 47 00:04:09,390 --> 00:04:11,240 up where it goes in the data frame. 48 00:04:16,200 --> 00:04:18,990 OK, so the next step is to plot our predictions. 49 00:04:26,230 --> 00:04:29,330 Notice how nice they look for the train set, but don't be fooled. 50 00:04:29,650 --> 00:04:31,930 Of course, what really matters is the test said. 51 00:04:35,300 --> 00:04:40,310 OK, so in order to get a better picture of what's going on, we're going to plot only the last one 52 00:04:40,310 --> 00:04:41,330 hundred data points. 53 00:04:45,530 --> 00:04:47,670 So the result is not too surprising. 54 00:04:48,140 --> 00:04:53,700 We see that our model only appears to do well on the train set because it copies the last value. 55 00:04:54,080 --> 00:04:57,620 This makes sense since the log price nearly follows a random walk. 56 00:04:58,100 --> 00:05:03,230 And of course, the forecast is a straight line since we've used holds linear trend model to fit our 57 00:05:03,230 --> 00:05:06,740 data set as an exercise. 58 00:05:06,770 --> 00:05:12,680 Try to compute one of our forecasting metrics for this prediction and compare it to the night forecast. 59 00:05:13,130 --> 00:05:18,380 Remember that for the naive forecast, we have to propagate the last known value through the forecast 60 00:05:18,380 --> 00:05:19,020 horizon. 61 00:05:19,430 --> 00:05:24,890 So don't copy the last value within the test set, but only copy the last value from the train set. 62 00:05:25,490 --> 00:05:27,380 This should give you a horizontal line. 63 00:05:28,010 --> 00:05:31,280 So in other words, see if a horizontal line is better than Holtze. 64 00:05:31,340 --> 00:05:32,390 Linear trend line.