WEBVTT - Kearns: Machine Learning & Monetary Policy 0:00:03.840 --> 0:00:08.840 This is Bloomberg Surveillance. The UK needs its own sovereignty, 0:00:08.920 --> 0:00:11.880 it doesn't need rules set out for it by the 0:00:11.960 --> 0:00:16.040 European community. That decisions to raise rates are not being 0:00:16.120 --> 0:00:19.000 driven by inflation, so what is driving at There is 0:00:19.040 --> 0:00:21.599 still only one global oil market and the price goes 0:00:21.680 --> 0:00:24.439 up because of outages in Nigeria, and whether we're importing 0:00:24.440 --> 0:00:27.360 oils from Nigeria not to reporting to be reflected here 0:00:27.360 --> 0:00:30.960 at Bloomberg Surveillance, your link to the world of economics, 0:00:31.160 --> 0:00:35.120 finance and investment on Bloomberg Radio. Good morning everyone, Michael 0:00:35.200 --> 0:00:37.720 McKee and Tom Kane. The drag press conference will see 0:00:37.720 --> 0:00:43.000 that in thirty minutes, low expectations, the official surveillance radar 0:00:43.120 --> 0:00:46.720 is up for what Mr drog may say, not only 0:00:46.720 --> 0:00:49.040 in the first five ten minutes of the press conference, 0:00:49.520 --> 0:00:53.159 usually obligatory a yawn, But then it always seems to 0:00:53.200 --> 0:00:56.280 get interesting. The Euro one eleven eighties seven, I'm gonna 0:00:56.280 --> 0:00:59.800 call it weaker euro over the last few days worldwide, 0:01:00.240 --> 0:01:03.840 across the nation. In New York, Bloomberg Surveillance vatch Vy 0:01:03.920 --> 0:01:08.119 Cone Resnick, look Ahead, Gain insight, imagine more, get forward 0:01:08.160 --> 0:01:12.880 thinking advice that can help turn business possibilities in the 0:01:13.000 --> 0:01:17.120 business opportunities. Find out more at Cone Resnick dot com. 0:01:17.280 --> 0:01:19.959 This is gonna be interesting, Michael McKee on the Bloomberg 0:01:19.959 --> 0:01:23.800 There is t y L go, that Taylor rule go, 0:01:24.560 --> 0:01:27.640 and you can plug in and hug the Newtonian mechanics 0:01:28.400 --> 0:01:32.759 of the algebraic function, which approximates John B. Taylor's great 0:01:32.760 --> 0:01:36.399 work out at Stanford. But there's other ways to do 0:01:36.480 --> 0:01:40.440 monetary policy, aren't there approximates? Of course John always said 0:01:40.440 --> 0:01:43.040 it wasn't a way to forecast where rates should be, 0:01:43.120 --> 0:01:45.919 just a way to double check rate where rates should 0:01:45.920 --> 0:01:48.800 have been in the past. But the question is can 0:01:48.880 --> 0:01:51.480 you come up with rules? Can you come up with 0:01:52.120 --> 0:01:58.320 some sort of reliable way of setting rates based on 0:01:58.480 --> 0:02:01.160 data that takes human emotion sort of out of it 0:02:01.720 --> 0:02:04.640 um And that's been a focus of the work of 0:02:04.640 --> 0:02:07.920 Michael Currn's. He's a professor and a National Center Chair 0:02:07.960 --> 0:02:10.880 at the University of Pennsylvania down in Philadelphia, where, of 0:02:10.960 --> 0:02:14.760 course Janet yellen Um, the real Janet you know, not 0:02:14.800 --> 0:02:18.040 the computerized version, is speaking on Monday, and that highly 0:02:18.080 --> 0:02:23.080 anticipated speech are Christopher Connon talked with you, professor, not 0:02:23.160 --> 0:02:27.400 long ago, about the possibility of using artificial intelligence to 0:02:28.120 --> 0:02:32.280 improve economic forecasting. And I guess that's the key, uh, 0:02:32.400 --> 0:02:34.760 to to improve forecasting, and then you can use a 0:02:34.840 --> 0:02:37.520 rule based on the forecast to figure out where your 0:02:37.560 --> 0:02:41.760 rate should be. Yeah, that's right. Chris and I had 0:02:41.760 --> 0:02:45.519 a long conversation about the possible application of machine learning 0:02:45.560 --> 0:02:49.160 to sort of macro economic forecasting in general and kind 0:02:49.160 --> 0:02:53.040 of policy setting more specifically, and UM, you know what 0:02:53.080 --> 0:02:56.960 I told Chris was that, UM, I'm sort of a 0:02:57.040 --> 0:02:59.200 machine learning advocate in the sense that I think that 0:02:59.400 --> 0:03:02.720 such a thing as possible in principle, but probably pretty 0:03:02.760 --> 0:03:06.960 far from being practical at this point. Yeah. I mean, 0:03:07.400 --> 0:03:12.359 it's not like the FED and other commercial enterprises don't 0:03:12.400 --> 0:03:15.680 have models, but they find that there are so many 0:03:15.760 --> 0:03:19.840 millions of influences that it's almost impossible to find tune 0:03:19.840 --> 0:03:23.240 them enough. Yeah, that's right. And I think that, you know, 0:03:23.320 --> 0:03:25.840 with all of the kind of media frenzy about machine 0:03:25.919 --> 0:03:29.680 learning these days, we forget that, um many people in 0:03:29.720 --> 0:03:32.720 many fields have been doing what has now been rebranded 0:03:32.760 --> 0:03:35.960 machine learning for a very long time. I think in 0:03:36.080 --> 0:03:39.760 terms of kind of macroeconomic policy prediction, the thing that's 0:03:39.800 --> 0:03:44.240 particularly difficult at this point for machines is just incorporating 0:03:44.480 --> 0:03:48.160 knowledge about you know, how markets work, how policy interacts 0:03:48.160 --> 0:03:53.320 with markets, how international events shape you know, the future. UM, 0:03:53.440 --> 0:03:57.840 and it's just very difficult to get clean data that 0:03:58.280 --> 0:04:01.360 you know kind of UM relates all of those working 0:04:01.440 --> 0:04:04.440 parts over a long enough history that you would have 0:04:04.480 --> 0:04:07.600 seen cycles, you would have seen market crashes, you would 0:04:07.600 --> 0:04:10.080 have had a lot of examples of how economy has 0:04:10.160 --> 0:04:14.120 changed and responses to policy. That goes to the heart 0:04:14.120 --> 0:04:17.040 of the matter. And course this goes to artificial intelligence 0:04:17.080 --> 0:04:20.719 and the rest. Where what have we accomplished in the 0:04:20.800 --> 0:04:24.320 last thirty years? If Governor Tarula who was just on 0:04:24.440 --> 0:04:28.240 with us, Cherry Yellen, Mario Doroggi, who's gonna speak in 0:04:28.279 --> 0:04:32.280 twenty five minutes. If people like that are working with 0:04:32.360 --> 0:04:38.440 basic algebra, Newtonian mechanics, and maybe something is fungible, is dynamics, 0:04:38.520 --> 0:04:44.760 stochastic general equilibrium theory, How can AI help them? I mean, 0:04:44.800 --> 0:04:48.520 it's very to me. It's your world is very nonlinear, 0:04:49.240 --> 0:04:53.160 it's it's got huge degrees of freedom issues. How can 0:04:53.360 --> 0:04:58.400 AI assist economists to do a better job? Well, I 0:04:58.440 --> 0:05:02.159 mean I think A can help, and in machine learning 0:05:02.160 --> 0:05:06.200 more specifically, can help in any domain in which you 0:05:06.279 --> 0:05:10.320 have massive amounts of historical data, including very high dimensional 0:05:10.400 --> 0:05:14.960 data um that's sort of relatively clean and and sort 0:05:15.000 --> 0:05:18.000 of drawn under you know, kind of similar conditions over 0:05:18.040 --> 0:05:21.040 a long period of time. The more sort of you know, 0:05:21.520 --> 0:05:25.480 heterogeneous your data is, the more diversity is, the harder 0:05:25.560 --> 0:05:27.640 it is to kind of apply machine learning. And this 0:05:27.680 --> 0:05:31.799 is why you see machine learning especially succeeding in either 0:05:31.880 --> 0:05:35.160 cases where there's a massive amount of clean data, as 0:05:35.160 --> 0:05:39.360 in areas like speech recognition or exactly, or in sort 0:05:39.360 --> 0:05:42.839 of very closed environments like game playing like chests or go, 0:05:43.520 --> 0:05:47.200 which is difficult as those problems are. You know you're 0:05:47.200 --> 0:05:49.440 in a closed world, right, you know that the rules 0:05:49.480 --> 0:05:52.320 are very Yeah, well there's I get the beautifully explained. 0:05:52.320 --> 0:05:55.680 There's sixty four squares and chess, I get it. Can 0:05:55.760 --> 0:06:00.359 you take your world and help Janet Yellen with the 0:06:00.480 --> 0:06:03.680 data that goes back to Lawrence client at pen In 0:06:05.000 --> 0:06:08.080 or whatever. I mean, was that data then clean or 0:06:08.160 --> 0:06:11.679 is that data now clean? I think in the case 0:06:11.800 --> 0:06:14.800 I mean, and and mind you, I'm not an economist 0:06:14.800 --> 0:06:17.479 per se, you know, and even less we take that 0:06:17.600 --> 0:06:23.240 as an advantage. Certain I appreciate that, but I think 0:06:23.279 --> 0:06:26.360 my short answer to that is no, and it's not 0:06:26.480 --> 0:06:28.320 so much whether the data is clean or not, it's 0:06:28.360 --> 0:06:30.240 just sort of the length of history. I mean to 0:06:30.560 --> 0:06:33.599 give an analogy, right, Um, you know, if you go 0:06:33.720 --> 0:06:36.680 back thirty years before sort of the automation of Wall 0:06:36.720 --> 0:06:39.440 Street more generally, there were a lot of you know, 0:06:39.560 --> 0:06:44.440 trading problems that at that time really required human expertise, intuition, 0:06:44.560 --> 0:06:47.080 knowledge of how markets work, how they might react to 0:06:47.200 --> 0:06:50.400 certain exogenous events, and so on and so forth. And 0:06:50.440 --> 0:06:52.280 now a lot of that stuff has been automated, right, 0:06:52.279 --> 0:06:54.320 I mean, sort of the brokerage business has become a 0:06:54.400 --> 0:06:57.440 very difficult business, um, at least being a sort of 0:06:57.440 --> 0:07:00.599 the human sort of specialist brokerage. But this has become 0:07:00.680 --> 0:07:04.279 very difficult as algorithmic trading has risen. And the real 0:07:04.320 --> 0:07:07.400 reason now that the trading has taken over is partly 0:07:07.400 --> 0:07:11.720 the automation. But the automation has you know, generated years 0:07:11.720 --> 0:07:16.240 and years a very clean data. Um. That makes machine 0:07:16.320 --> 0:07:19.560 learning a much more practical, you know, sort of practical approach. 0:07:19.880 --> 0:07:22.440 And by the way, you tend to see the you know, 0:07:22.520 --> 0:07:26.960 the greatest uses of machine learning in trading or economic 0:07:27.040 --> 0:07:31.040 settings where the most data is generated, so high frequency 0:07:31.120 --> 0:07:33.760 trading for instance, right, just because of the speed at 0:07:33.800 --> 0:07:36.840 which thing is things are happening, you know, in the 0:07:36.920 --> 0:07:40.000 same minute. Um, I get way more data about high 0:07:40.040 --> 0:07:43.200 frequency trading that I do about macro economous policy predictions. 0:07:43.240 --> 0:07:45.000 So can your model tell me if there's going to 0:07:45.080 --> 0:07:52.400 be a junior July rate increase? I couldn't tell you 0:07:52.440 --> 0:07:55.240 if I knew good. No one else can either, So 0:07:55.360 --> 0:07:59.680 can like, like one more question? Please? Well, I just 0:07:59.720 --> 0:08:02.680 car is because Bill Dudley and others are de fitive 0:08:02.720 --> 0:08:06.560 noted that one failure of their models is that doesn't 0:08:06.600 --> 0:08:10.880 take financial markets into account. Financial markets, much of it 0:08:10.920 --> 0:08:13.000 is high frequency trading, but much of it is also 0:08:13.200 --> 0:08:19.520 market psychology. How does machine learning account for psychological factors? Yeah, 0:08:19.560 --> 0:08:22.040 so that's a great question. And um, one of you 0:08:22.080 --> 0:08:25.440 mentioned equilibrium a little while ago, and you know, the 0:08:25.560 --> 0:08:30.200 term equilibrium kind of refers to taking the strategic considerations 0:08:30.360 --> 0:08:33.480 of the parties into account, you know, or or you know, 0:08:33.559 --> 0:08:35.800 kind of counter factuals, if you like, sort of not 0:08:35.960 --> 0:08:39.480 just how did the data look historically, but how might 0:08:39.559 --> 0:08:42.360 it have looked different if we had done something differently 0:08:42.400 --> 0:08:45.840 and something different had happened. And one thing I like to, 0:08:46.200 --> 0:08:47.520 you know, one way I like to put it when 0:08:47.520 --> 0:08:51.120 I talk to people about machine learning in finance or 0:08:51.120 --> 0:08:54.439 economics versus in other domains, which is, you know, as 0:08:54.520 --> 0:08:57.840 hard as the problem of let's say, recognizing whether there's 0:08:57.840 --> 0:09:00.480 a cat in a video on YouTube or not might be, 0:09:00.720 --> 0:09:03.640 and you know, you might think I'm joking, but actually 0:09:03.640 --> 0:09:07.000 that's not an easy problem. As hard as that problem 0:09:07.080 --> 0:09:10.000 might be, it one one advantage you have an applying 0:09:10.000 --> 0:09:12.319 machine learning to such a problem is that you can 0:09:12.320 --> 0:09:15.959 be pretty sure that you know the world is not 0:09:16.720 --> 0:09:19.920 you know that your very effort to decide whether there's 0:09:19.960 --> 0:09:22.520 a cat in the video or not is not you know, 0:09:22.640 --> 0:09:26.400 changing the nature and casts um you know themselves right. 0:09:26.880 --> 0:09:28.960 Where this is not true in trading, you can be 0:09:29.040 --> 0:09:32.440 quite sure that you know your very effort to predict 0:09:32.559 --> 0:09:36.400 something in financial markets and then act on that prediction 0:09:37.120 --> 0:09:39.880 link fact change the markets in a way that makes 0:09:40.160 --> 0:09:43.240 what you're doing less effective. Right, So it's that kind 0:09:43.240 --> 0:09:46.839 of adaptive dynamic, the market reacting to what you're doing 0:09:47.120 --> 0:09:51.320 because of the strategic considerations the other parties involved, and 0:09:51.360 --> 0:09:54.280 there there are branches of machine learning that are trying 0:09:54.280 --> 0:09:59.280 to seriously take those kind of strategic considerations and counter 0:09:59.320 --> 0:10:02.120 factuals in to account. But it's a much much more 0:10:02.120 --> 0:10:05.360 difficult problem. And I think that you know, we're very 0:10:05.440 --> 0:10:08.080 very far from kind of understanding how to deal with 0:10:08.120 --> 0:10:10.240 such problems at the at the kind of scale that 0:10:10.280 --> 0:10:12.520 you guys are talking about. Michael Currents, thank you so much. 0:10:12.559 --> 0:10:14.839 At the University of Pennsylvania and machine learning and a 0:10:14.920 --> 0:10:18.760 touch there and AI as well, our machine learning, I mean, 0:10:18.760 --> 0:10:22.600 we are robots, Mike Is. We're twenty minutes away from 0:10:22.600 --> 0:10:25.360 the drug press comments. I would don't, Mike. Not much 0:10:25.400 --> 0:10:27.960 movement in the market. I guess that's not a surprise, Mike. 0:10:28.000 --> 0:10:29.719 I would know two days in a row, strong and 0:10:29.760 --> 0:10:33.280 strong Yen one O eight ninety two and a yen 0:10:33.360 --> 0:10:37.080 as well, indeed, but but no real reaction to Torulo 0:10:37.600 --> 0:10:39.720 or to the e c B. And I guess, uh 0:10:39.960 --> 0:10:42.520 Jim vot will put it, well, it's all down to 0:10:42.600 --> 0:10:45.880 Janet Yellen on Monday, the futures and negative four down 0:10:45.880 --> 0:10:52.920 futures at negative yield one point eight three. Time now 0:10:52.960 --> 0:10:54.640 to check in with Michael Barr and get caught up 0:10:54.640 --> 0:10:57.040 on world in the national headlines. Mike, Tom, thank you 0:10:57.160 --> 0:10:59.920 very much. Hillary Clinton has sat to unleash a four 0:11:00.000 --> 0:11:04.040 in policy attack on Donald Trump. The former Secretary of 0:11:04.040 --> 0:11:06.360 State will use a speech in San Diego today to 0:11:06.400 --> 0:11:11.080 cast the Republican as unqualified and dangerous. Trump accused Clinton 0:11:11.120 --> 0:11:13.839 of lying about his foreign policy plans at a rally 0:11:14.120 --> 0:11:18.240 in Sacramento, California, last night. Classes will resume next week. 0:11:18.320 --> 0:11:21.360 At u c l A yesterday, a professor was shot 0:11:21.360 --> 0:11:24.439 and killed before police say the gunman took his own life. 0:11:24.920 --> 0:11:28.120 The head of the Global Airline Industry Association says terrorist 0:11:28.200 --> 0:11:32.640 attacks will not stop surging travel demand. International air Transport 0:11:32.640 --> 0:11:37.160 Association CEO Tony Tyler told Bloomberg people don't get frightened 0:11:37.160 --> 0:11:40.480 off by these thugs. Global News twenty four hours a day, 0:11:40.520 --> 0:11:43.480 powered by our twenty four hundred journalists more than a 0:11:43.559 --> 0:11:46.360 hundred fifty news bureaus around the world. I'm Michael Barr. 0:11:46.960 --> 0:11:49.240 Michael Barr, Thanks so much, che life. If you've got 0:11:49.240 --> 0:11:51.880 a Bloomberg terminal in your car, t l i V 0:11:52.920 --> 0:11:56.240 great updates on the Draggy press conference and on OPEC 0:11:56.880 --> 0:11:59.600 in vi Enna. Futures at negative four. This is Bloomberg. 0:11:59.640 --> 0:12:07.840 Save us. The news update brought you by your Mercedes 0:12:07.840 --> 0:12:09.920 Benz tri state dealer. The star you've been wishing for 0:12:10.000 --> 0:12:12.000 is waiting at the Mercedes Benz Summer Event. Head to 0:12:12.000 --> 0:12:14.600 your Mercedes Benz Tri State dealer for more details. Hurry 0:12:14.840 --> 0:12:19.240 before this star flies by.