1 00:00:11,040 --> 00:00:15,780 So in this lecture, we will be summarizing what we learned in this section, which was all about a 2 00:00:15,780 --> 00:00:22,260 Facebook time series library called Profit, we observe that this library is a lot like the ETS methods 3 00:00:22,260 --> 00:00:26,880 we learned about before and that they both explicitly model trend and seasonality. 4 00:00:27,570 --> 00:00:31,470 But profit has many differences compared to the methods we've studied before. 5 00:00:32,400 --> 00:00:35,340 Firstly, we saw that profit is not auto regressive. 6 00:00:35,610 --> 00:00:37,230 Instead, it's only regressive. 7 00:00:37,230 --> 00:00:38,600 It is time itself. 8 00:00:39,180 --> 00:00:42,750 Because of this, we saw that profit handles missing data with ease. 9 00:00:43,980 --> 00:00:48,180 Furthermore, we saw that profit allows you to explicitly handle holidays. 10 00:00:48,540 --> 00:00:53,910 You can either enter these yourself or you can use profits built in functions for standard holidays 11 00:00:53,910 --> 00:00:54,750 in each country. 12 00:00:55,740 --> 00:00:59,790 We also saw that profit model seasonality at multiple scales at once. 13 00:01:00,000 --> 00:01:06,330 Unlike other methods, it's able to model yearly, weekly and daily seasonal patterns using a four year 14 00:01:06,330 --> 00:01:07,090 series. 15 00:01:08,160 --> 00:01:11,500 Finally, we learned that the profit model is a Bayesian model. 16 00:01:11,910 --> 00:01:17,190 Thus, when you're working with clients, you can help them tune the model by modifying its priors. 17 00:01:18,150 --> 00:01:23,520 Because of these factors, we learn that profit is an excellent choice for business use cases such as 18 00:01:23,520 --> 00:01:26,190 modeling, sales, CPU usage or ads. 19 00:01:26,190 --> 00:01:32,250 But we also learn that it is not very good at predicting stock prices, as some blog articles would 20 00:01:32,250 --> 00:01:33,300 have you believe. 21 00:01:38,060 --> 00:01:41,430 In this section, we learned how easy it is to use. 22 00:01:41,900 --> 00:01:45,440 Let's go over the steps one more time to summarize what we've learned. 23 00:01:46,490 --> 00:01:49,640 So we always start by instantiating a profit object. 24 00:01:50,120 --> 00:01:55,070 At this point, we can choose various options like the growth pattern, the types of seasonality and 25 00:01:55,070 --> 00:01:55,910 the Pryors. 26 00:01:56,900 --> 00:01:58,580 The next step is to fit the model. 27 00:01:59,120 --> 00:02:06,560 As you saw, this requires a specially formatted data frame with the specific columns D and Y, optionally, 28 00:02:06,560 --> 00:02:09,500 you can add more columns as exoticness or grassers. 29 00:02:10,490 --> 00:02:16,040 However, you have to be aware that when you make a forecast, you must have the values of those exoticness 30 00:02:16,040 --> 00:02:18,500 progressors for the forecasting dates. 31 00:02:20,630 --> 00:02:24,470 The next step is to make predictions, which actually involves multiple steps. 32 00:02:25,190 --> 00:02:28,530 We start by creating a placeholder data frame, which we call future. 33 00:02:29,210 --> 00:02:35,060 This contains the dates we want to make predictions for when we call, predict profit, takes these 34 00:02:35,060 --> 00:02:38,160 dates and makes predictions for the given dates. 35 00:02:39,290 --> 00:02:43,820 Once we have our predictions, we can easily plot them with a single call to the plot function. 36 00:02:44,480 --> 00:02:48,590 Those plots, both the sample predictions and the out of sample forecast. 37 00:02:49,580 --> 00:02:53,650 And with just one additional function call, we can plot the change points as well. 38 00:02:55,210 --> 00:03:00,040 The next step is to applaud the individual components of our prediction, which, again, is just a 39 00:03:00,040 --> 00:03:00,830 single call. 40 00:03:01,870 --> 00:03:07,540 So this gives us the trend that seasonal and holiday components and external progressors if you had 41 00:03:07,540 --> 00:03:08,060 them. 42 00:03:08,950 --> 00:03:14,110 We also learned that we can do cross-validation and just a single line of code which works precisely 43 00:03:14,110 --> 00:03:17,110 like the walk forward validation I previously described. 44 00:03:17,680 --> 00:03:22,600 This returns a data frame of predictions from which we can compute performance metrics.