1 00:00:11,040 --> 00:00:16,730 OK, so in this lecture, we will be discussing a surprising observation we have made in this section, 2 00:00:17,520 --> 00:00:23,310 this observation was that Arnon's don't seem to perform particularly well when it comes to TIME series 3 00:00:23,310 --> 00:00:24,160 analysis. 4 00:00:24,870 --> 00:00:29,810 So before we even begin discussing the main point of this lecture, I have a question for you. 5 00:00:30,120 --> 00:00:36,300 The students of this course, I know that many of you have assumed that ListBox must be excellent for 6 00:00:36,300 --> 00:00:37,170 Time series. 7 00:00:37,740 --> 00:00:41,520 My question is, why do you think this what gave you this idea? 8 00:00:42,330 --> 00:00:47,970 So as an exercise for this lecture, if this was one of your assumptions coming into this course, please 9 00:00:47,970 --> 00:00:51,060 let me know on the Q&A why you believe this to be true. 10 00:00:55,950 --> 00:01:00,420 So your next question might be, why do Arnon's seem to perform so poorly? 11 00:01:01,350 --> 00:01:05,190 Unfortunately, answers to questions like these are harder than you think. 12 00:01:05,640 --> 00:01:10,670 A question like this might take up a whole Ph.D. thesis that is years of work. 13 00:01:11,190 --> 00:01:16,710 My personal belief is that it's usually not worth your time to invest in the theory of deep neural networks 14 00:01:16,980 --> 00:01:19,830 unless you have a very strong background in mathematics. 15 00:01:20,970 --> 00:01:25,680 Furthermore, it's often the case that when people come up with theories about why things work or don't 16 00:01:25,680 --> 00:01:29,440 work, they are likely to be superseded by new theories in the future. 17 00:01:30,150 --> 00:01:35,340 So unless you want to devote years of your life studying a niche topic only for it to be obsolete a 18 00:01:35,340 --> 00:01:39,150 few months later, don't do it unless it's truly what you love to do. 19 00:01:40,050 --> 00:01:43,710 Now, my students know that I love inventing motto's about machine learning. 20 00:01:44,340 --> 00:01:49,790 The relevant model for this situation is machine learning is experimentation and not philosophy. 21 00:01:50,550 --> 00:01:54,660 What this means is that you often don't care about why something works or does not work. 22 00:01:54,900 --> 00:01:56,470 You just care about the results. 23 00:01:56,970 --> 00:01:58,560 The result is easy to check. 24 00:01:59,130 --> 00:02:02,100 Simply do the experiment and look at what happens. 25 00:02:06,530 --> 00:02:10,260 The nice thing about this is someone has already done these experiments. 26 00:02:10,730 --> 00:02:17,090 There is a paper included in extra reading called Statistical and Machine Learning, Forecasting Methods, 27 00:02:17,450 --> 00:02:19,010 Concerns and ways forward. 28 00:02:20,000 --> 00:02:26,000 So in this paper, the authors compare both classical times, various methods like Hetson Arima, along 29 00:02:26,000 --> 00:02:31,610 with machine learning and deep neural networks, including trees, support vector machines and LSD. 30 00:02:32,360 --> 00:02:34,070 So let's look at the results. 31 00:02:35,000 --> 00:02:41,110 So this is a chart showing estimate on the Y axis in computational complexity, on the X axis. 32 00:02:41,600 --> 00:02:44,550 Basically, the higher you are, the worse error you have. 33 00:02:45,170 --> 00:02:47,540 Well, which model is the highest on this chart? 34 00:02:47,960 --> 00:02:49,300 That would be the Helstone. 35 00:02:50,270 --> 00:02:55,730 Sadly, many of the machine learning methods perform poorly, showing that although these methods may 36 00:02:55,730 --> 00:03:00,650 be powerful for other types of data, this is not the case for Time series forecasting. 37 00:03:01,580 --> 00:03:07,160 Note that for this paper, the authors looked at the M3 competition, which is a data set with over 38 00:03:07,160 --> 00:03:09,010 one thousand monthly time series. 39 00:03:09,320 --> 00:03:13,600 So they certainly had a lot of data to work with to justify these results.