Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=

# Geoeconomic Pressure

Authors:   
Discussant: None  
Video: https://www.youtube.com/watch?v=3YzXcC2W46s&t=5936s


## Talk (01:38:56 – 02:20:28)

[01:38:56] you know a first pass. Great. So now let's talk about a very different way to do this. Okay. So what I did so far in some sense is very traditional. It's the bread and butter of trade and and a lot of macro. We were interested in getting  
[01:39:11] at this also from a different perspective. Now originally where we started from was the idea that a lot of these threats, a lot of this pressure might be off path. Um, if you go back to the example of ASML, of course they're  
[01:39:25] going to comply because the US has a huge threat on them and they're going to comply and I'm not going to see it being cut off in equilibrium. If the US tells a European bank, stop trade financing Iran or you lose access to the US  
[01:39:39] financial markets, they tend to comply. Uh, and so there's a there's a simple endogenity on some of the most powerful threats. You're never going to see them realizing the data because the target complies. Also we started thinking a lot of this might be uncertainty but then it  
[01:39:54] became more interesting in our head. A part that became very interesting to us is what they're asking you to do might be pretty difficult to predict exam like what what are those toss and that with hard data it's hard to  
[01:40:08] prespecify and we might not always see it. So we started thinking about text.  
[01:40:13] Now why is text a good place? Well there is at least two reasons. Uh the first is uh text by being narrative and nonstructured potentially captures a lot of things. You can imagine that the CEO  
[01:40:26] of ASML is gonna get asked why did you stop [clears throat] selling stuff to China and he's gonna mumble something along the lines of we kind of have to and is not going to want to be maybe too precise about the pressure but it's going to mention something or a bank  
[01:40:40] that stopped trade financing Iran or a Chinese CEO that is being asked not to sell something to the US. The other thing that it does is it's available where you let them tell you what the ask  
[01:40:55] was. You don't have to prespecify it in categories. And so that's what we started thinking about. Okay.  
[01:41:03] So this is obvious to all of you now because you're using AI a lot. But just in case it's not obvious to you, um clearly one massive shift of the last two or three years has been the ability to analyze text. in particular the  
[01:41:16] ability to extract from the text uh potentially very uh sort of dense information. So you know like our colleague Nick Bloom has been a pioneer of this all of his work on text and  
[01:41:29] uncertainty. I'm gonna mischaracterize a bit of the sort of evolution of the literature but the way I think of it as a simple-minded thing is you know I could do common find so you give me a document I can do common find for the  
[01:41:44] word tariff is does the document mention tariff now that's pretty low quality uh a much better natural language processing is like biograms we're  
[01:41:56] looking for a combination of words that are of interest to us and we're comparing them to some baseline frequency in either the English language or in text of that type. So a lot of the work that Nick has done as a feature of  
[01:42:10] that. Okay. And we're looking at how frequently do they mention the word uncertainty or or tariff.  
[01:42:17] You can also go a different route which wasn't very popular in economics um with some exception but it certainly is popular in political science which is humans. You write down a long set of instruction that say please read this  
[01:42:31] document and tell me put put in a spreadsheet all of the following categorization of the document. Now the obvious problem is hard to do. You know we did it without a one of these  
[01:42:44] document took about 25 minutes. If you're going to do 1.2 million documents that's not going to work very well. What AI has really done in shifting the frontier is it is essentially allowing us to do what we would have liked to do  
[01:42:58] with a human. Give them a long set of instructions to extract potentially very complex information out of the text but relatively speaking with the simplicity of the biogram approach with an algorithm that just simply looks for it.  
[01:43:12] Now why is AI and particularly large language models very good at this?  
[01:43:17] Again, many of you will know more than I do about this. Uh, but it comes from the cross attention mechanism. Now, what is that? It's essentially a very nonlinear network that is putting weights on words  
[01:43:30] tokenize that are potentially very far apart in the text and has been trained to recognize that um in the English language, the sequential nature of the language means that words that might be potentially far apart in that context  
[01:43:44] are related to each other. So when we all chat with chat GPT or code the reason why we really like it it's really deep down coming from that mechanism okay and that it feels a lot more human than a common fine okay so that's what  
[01:43:58] we did so basically the corpus of text is the text of CEOs and CFOs or public firms worldwide speaking to investors and it's a text that has been used a lot  
[01:44:12] before in economics for example Nick has used it a lot. Um, by the way, if you're a student, this is easy text. It's available routinely uh in by data providers even a stand for other  
[01:44:25] universities. We also included analyst reports. Why? Because in this particular context, you have to worry that because it's politically sensitive, the CEOs might be very careful about what they say. basically had in mind that if  
[01:44:38] you're a US CEO, you can bash China all you like, but you got to be careful when you talk about the US administration. If you're a sitting Chinese CEO, you might have opposite incentives. But you might think that an analyst that is arms  
[01:44:52] length, somebody that is paid to cover the company, might not have that problem. That might be more than willing to say they're doing this because they're getting pressure from the foreign office. Okay, so we did that. Um the prompts are nothing else than the  
[01:45:07] instructions that if you went for the human approach you would have given the array. Uh so and you know we published them online. Uh and this is a paper by the way with Antonio Copa. Antonio the first time he showed me three years ago that we could do this. I was totally  
[01:45:22] blown away because I was massively under ambitious. I thought that we were going to give you like a few words that said here's the instruction. You showed up with like three pages of text. I didn't even know that this was possible. Um, but they're very detailed instructions.  
[01:45:36] And ultimately what this is is you take the text, you take the prompt, you feed it to an LLM and you're really using the LLM as a classifier. So ultimately what comes out of it is a  
[01:45:50] structured data set. In the particular case that what we did a lot of it is 01 variables. Is the firm affected by the tariff? Is the firm positively affected or negatively affected? Some of it is less structured. tell me which product  
[01:46:03] is affected by the export control. So it might give you an answer that is in words. Uh but most of this is going to be hard classification. Now if you're a PhD student, let me spend two seconds on  
[01:46:15] this. Um there is at least two big considerations.  
[01:46:20] One consideration is open weights versus closed weights models.  
[01:46:24] So the way the landscape of LLMs is right now is changing of course very quickly. As we speak today uh a lot of the frontier models are closed weights.  
[01:46:35] What it means is that uh code chpt the algorithm and their weights are proprietary to them. You don't see them.  
[01:46:43] You might have some ability to control features of the model. In particular, you might have the ability to make the model deterministic.  
[01:46:52] Okay? But ultimately, you don't control it. If they change something in the background, you're not going to recover your results.  
[01:47:01] Open weights models are different. They're the same architecture, but you can download the weights from hugging face. You can put them on the Stanford supercomput. In a lot of these cases,  
[01:47:14] you are a, you know, there are features that make the model deterministic.  
[01:47:18] Essentially, you're going to be able to do narrow replicability in the context that you're familiar with in Python STA or MATLAB, which is same code, same data. If I press the button a million times, I get the same result. Okay. The  
[01:47:33] other advantage, of course, if you have the weights is that you might be able to fine-tune yourself. You might be able to modify, add things that you cannot do if you don't have the weights. What are the disadvantages? There's at least two. The first is the computational  
[01:47:47] infrastructure goes up a lot because now you have the weights and you have to do them. You have to run them. Our experience has been that over the last 3 years that computational cost and problem has gone down a ton. Quite  
[01:48:01] frankly 3 years ago I didn't know how to do this. It seemed like a complicated problem. Now you just tell code please run this model and give me the answers.  
[01:48:10] So that is also changed a lot. the it is still cumbersome compared to an API call. Okay, it's a lot cheaper and in particular Stanford unfortunately I believe right now not this to this not  
[01:48:24] to be true for students that you don't have access to the new supercomput uh for GPUs u but you have access to parts of it it's generally a lot cheaper than the commercial services okay this is also changing so rapidly that whatever  
[01:48:39] I'm saying if it's not obsolete already it's going to be obsolete in two weeks um but it does have these are some of the advantages and disadvant advantages in practice. Who knows? Uh in the US, most of the frontier models have gone  
[01:48:54] closed weights, although Google keeps releasing close to frontier open weights. In China, they were doing mostly open weights. Uh but Alibaba just switched to closed weights. It's changing. Uh but I strongly encourage  
[01:49:08] you to experiment with this. Okay. Now, eventually we get to this. Now one interesting econometric problem that I'm going to return to is that this is generated data and you might have to  
[01:49:23] worry about not only the narrow replicability can I get it back again uh but also like does he have biases does he have measurement error that is nonclassical how I'm going to do econometric based on this and a lot of  
[01:49:36] the recent papers in econometric are tackling this it's a very active area even if you're you know purely taking the applied path I'm going to give you a sense of this so let's Go back to the example I started with. Okay, let's go  
[01:49:48] back to ASML. They have US suppliers. They have Chinese customers or they used to and there's a US government pressuring them. They're a public firm.  
[01:49:57] So their CEO and CFO had do public calls that get transcribed.  
[01:50:03] If I feed their transcripts through the algorithm, what comes out is that the country imposing the pressure is the US and the Netherlands. The country receiving the pressure is China. The  
[01:50:17] firm products are extreme ultraviolet and deep ultraviolet. So you can see that these ones are all R values. This is essentially free text that the algorithm will reply and then you have to classify it if you want to go to an  
[01:50:29] HS6. U the overall impact on the firm is negative. This is a costly action that they were asked to do. What was the action that they were asked to do? Lower sales to China. Okay. And we're going to have 55 of these answers in in many  
[01:50:43] dimensions. We're going to ask did did you change supplier? Did you increase the price? Did you uh lower the inputs?  
[01:50:50] All of this. Okay. Good. What are two advantages of this that I haven't mentioned? The first is uh particularly for questions in trade and geoeconomics. Uh one advantage which surely is familiar to you from the trade  
[01:51:04] class is that trade is very very heavily dominated by multinationals.  
[01:51:10] from the finance classes you know that multinationals and large firms are disproportionately likely to be public firms and public firms produce text. So in this particular context uh text is  
[01:51:22] friendly because uh it comes mostly from public firms but public firms are not everything you like to know but they're a big chunk of the global trade activity. If I were doing a paper on domestic employment that might be different but for these questions the  
[01:51:36] public uh text is good. The other thing that I think it's interesting and we only started to exploit it now is that of course I have text from the suppliers from the firm itself and from the potential customers. So one of the  
[01:51:50] things that I want to know going back to the equations on the counterfactual is okay now this is telling me that the US threatened ASML that SML complied they did take on board the costly action they  
[01:52:04] stop selling to China but now I can go to the previous Chinese customers and I can say was this expensive or cheap for you? Remember that we have predictions depending on the lasio substitutions and the shares whether this affects the  
[01:52:17] target a lot or very little. Um and so we can ask how are they reacting and what are they doing. Okay. So it's kind of nice because you have text all around the supply chain and you can try to chase it. So we're going to play around a little bit with that. Okay. So let's  
[01:52:32] start from the aggregates. Okay. So let's look at export controls. This is a percentage of uh all publicly listed firms for which we have text. Um the mention being affected by export controls. So you can see that originally  
[01:52:47] this was like nonopolic was being used a lot and then in recent years there is a big trend up. Where does it come from? A big chunk is the US pressure in China that will turn out to be mostly semiconductors. But then you can see  
[01:53:02] here towards the end a green area opening up. That's the Chinese firms uh pressuring the US with export controls on rails. Now at this level you could have probably gotten away with a lot of  
[01:53:15] this which much simpler methods at NLM. You know for example in the aggregate biograms would have worked pretty well for does the firm mention export controls. Now, our own experience has been if you focus on a policy where the  
[01:53:28] there's a key word like tariff, anybody that is affected by tariff tends to use the word tariff, then it's easier.  
[01:53:35] Export controls are surprisingly nasty because there's just lots of ways to phrase what an export control is and AI is quite good at that. But then again, you could have spent a lot of time with the biograms to figure out all the right combination of words. Okay, so I  
[01:53:49] wouldn't do AI for this. If you do it for sanctions, what do you find? the two episodes that you would imagine. These are the sanctions on Russia following the earlier invasion of Crimea and the war in Ukraine. And then in the middle, there's probably an episode that is less  
[01:54:03] familiar to you. Those are the US imposing sanctions on Huawei and Zerat in the Trump one. Okay, if you look at tariffs, it's Mr. Trump one and two. Uh they're large tariffs and everybody else  
[01:54:17] um that you that you can see. But now let's start to go deeper.  
[01:54:22] In particular, you can start tracing out who's imposing the pressure using what sector and towards whom.  
[01:54:32] And you can see that a lot of the pressure is coming from the US on China using the semiconductor supply chain and China on the US using rare earths. But you can go a lot more disagregated.  
[01:54:45] Okay. Also bear in mind that one of the nice things of text is that we might be getting some of this information from a firm that isn't involved.  
[01:54:55] Think of for example the sanctions of the US or Europe on Russia and oil. A lot of the information might be coming from Chinese firms or Indian firms that are saying we're very positively  
[01:55:08] affected by the US uh sanctions on Russian oil because we got cheap oil as a result. Okay. So you get a lot of third party effects.  
[01:55:19] Okay, good. Now go back to the previous formula. Okay, let's simplify it a bit.  
[01:55:23] So we mentioned this nonlinearity. I get a lot of power or a lot of pressure on the target when 100% of what they were using comes from me or when the elysio substitution is coast to cop Douglas.  
[01:55:37] Okay, so the the the plot before here I just did it in logs. It's here. If you look at the US hard data versus China, it looks like this. Clearly, it's off because there is other things that matter. In particular, these big omegas  
[01:55:51] here that in a threedimensional plot I'm not capturing. And then we started thinking, okay, can we use AI to get a sense about whether these models are capturing the world? Here's an example. Okay, and this is very provisional, but it's kind of  
[01:56:06] fun. um we started thinking if I pick firms that are in these particular sectors what is the probability that their report being used to target China.  
[01:56:19] So here what we're doing is we're saying is the US policy maker using the model to pick which sectors to use or is he picking at random and here they are. So a red scale, a red  
[01:56:33] and bigger bowl means firms in that sector are much more likely to report being used to pressure China. And you can see that it's a bit of a mixed bag, but overall they're kind of using the right thing. I think of this as you know  
[01:56:47] Jake Salman who was the national security adviser to Biden saying for export controls we want a very high fence on a small garden. We don't want to affect every relationship with China.  
[01:57:00] We want to pick a few sector and put very very high controls. This looks a lot like this. Like I think semiconductors is one of this.  
[01:57:08] This is aircrafts. There's also misses. For example, the US has never used autos uh visav China. But they're not picking a lot that is down here. Now imagine that I did this the other way around.  
[01:57:21] Imagine that I looked at countries exporting to the US and I looked at tariffs.  
[01:57:29] it will look like a Christmas tree, right? We tariffed everybody including, you know, islands with three penguins.  
[01:57:35] Uh we didn't say, well, let's only tariff a few countries that really were a huge customer for them and they cannot find an alternative. One thing that we haven't discussed is in one of the papers that I've assigned you, but I  
[01:57:48] didn't cover it today. Suppose that we switch the threats to not being in threats not to sell you something, but with it to be in threats not to buy something from you. It's essentially the threat of imposing a tariff. If you  
[01:58:01] think of the wedge implementation or if you want to do a full ban as an infinite tariff, then what is the theory there?  
[01:58:08] The theory there is that in matters that I if to you in matters that I'm telling you a tariff, if once I cut you off from exporting to me, you will have to move the price index of your sales as in you  
[01:58:22] will have to depress prices at which you sell to find a new customer. So what we're looking for is the elasticity of how easy it is for you to find another customer. So if I'm a big huge market, I  
[01:58:35] have some power because if you lose me in order to sell the same volume, you're going to have to lower prices and hence you might be willing to give me a concession.  
[01:58:45] That's a the theory of a tariff. If you think of for example Russia, if they cannot sell to Western Europe and the US, the example that you had in mind, um they they need to sell to India and  
[01:58:58] China, but India and China are a relatively small market for Russian exports of oil and so you're going to depress prices. So allegedly the estimates are that Russia sells somewhere 30% uh at a 30% discount.  
[01:59:12] Okay, imagine that we did that version and then looked at which firms report that they're being used for tariffs.  
[01:59:21] That's going to say they're not using the model. They're picking all sectors.  
[01:59:25] But what we're planning to do, what we're working on is go the other way.  
[01:59:29] Then say, okay, start with this one. Suppose that we are randomized and that we have pick sectors here that the model says it shouldn't matter for the target. That should be very easy  
[01:59:42] for them to find an alternative. Then we can go to the text of the target and say do they indeed report that being cut off didn't really matter. They could find it easily. And similarly for the tariff, you can you can go to sectors and say  
[01:59:56] the sell to the US and say look were they bothered did they lose a lot of revenue from the tariff? Well some sectors might if you're selling for example I don't know things that only the US buys like oversized SUVs. um that  
[02:00:11] might be a problem. But if you're selling plain stuff where the US is 10% of your customers, you can reallocate pretty quickly. Okay, so that's part of what we're planning to do. And this gives you like a a first sense,  
[02:00:25] but here's some things that we did do. So, okay, the these one on the export control says the US picked a few sectors that look like they use the model. Okay, then you can say go to China and how do  
[02:00:38] they react? So this is this panel here. First of all, the firms are reporting that this was expensive. There are many more firms that report that this is negative for them than is positive for them. It's affecting profits. But also  
[02:00:52] look at the response. What is the main response? Domestic R&D.  
[02:00:58] They're saying, look, we lost the semiconductors from the US.  
[02:01:03] We're doing large investments to replace them with a domestic source because we don't want to be cut off in the future.  
[02:01:09] So in some sense it tells you exactly what you would have you would have imagined to see a switch to an alternative source. When you look at the US firms it's expensive to them. That goes to your question of in in the  
[02:01:22] theory we made it totally inexpensive as you're cutting off a single firm. But clearly when you're cutting off China, China is a big customer. For Nvidia it's pretty expensive not to sell to them.  
[02:01:32] There is an R&D response. But the funny thing is when you look at the text, what are they doing? This is Nvidia saying we're doing R&D to develop a cheap that the Chinese want and they're just below the cut off for export controls in the  
[02:01:45] US. So it's just bypassing the legislation. Uh and there's a lot of that.  
[02:01:53] What else? So let's go to tariffs. Okay. So I gave you at least an informal rendition of a tariff as a geioeconomic negotiating tool. Um when you look at tariffs you can split the firms into  
[02:02:07] positive these are US firms into positively and negatively affected. The first thing that you find is that US firms overwhelmingly report being negatively affected by US tariffs.  
[02:02:19] But there is a bunch of firms here this blue line that report being positively affected by the US tariffs and I want to show you who they are. Okay. Now you've all taken the trade class. What is a  
[02:02:31] tariff? A tariff is other than attacks on imports. We think of it as a combination of two policies. What are the two policies?  
[02:02:45] Fine. I'll take you out of your miss. It's a consumer tax and a domestic producer subsidy. One of the reasons why economists tend not to like tariffs is because it's an instrument that really is two combined policies and you might want to have separate instruments that  
[02:02:59] affect one margin but not the other. For example, you might want to have a domestic producer subsidy but without taxing the consumer.  
[02:03:06] So let's have a look at this. If I take the US firms and I divide them into those that report being positively affected and those that report being negatively affected. First, as a as a sanity check, you can see that it corresponds to the profit margin. Those  
[02:03:20] that are positively affected report an increase in profit margins. Those that are negatively affected report a decreasing profit margin. But then focus on the other oh sorry the other side.  
[02:03:32] Start with the input price. The firms that are negatively affected are much more likely to report that as a consequence of the tariff they're facing higher input prices.  
[02:03:45] Okay. So, think of a world in which the foreigners are not cutting their prices at all. You're not successfully manipulating the terms of trade. You face as an importer the old price plus the advalorum tax, the tariff, and so  
[02:03:59] you're facing a higher input price. If you look at the sales price, what you sell at, they are reporting that they're passing on some of this. As a consequence, they're increasing their own sales  
[02:04:13] price. But you can see that the effect the difference between the firm that are positively affected and the firms that are negatively affected is much more muted when it comes to increasing their own prices.  
[02:04:27] So what is going on is the following. Imagine that I have a steel mill in the US that doesn't import steel from abroad or from Mexico. The US puts very high  
[02:04:39] tariff on imported steel. that steel mill is going to report in the data that they're positively affected by the tariff because their competitors are facing higher input prices because they were sourcing from abroad. Their  
[02:04:53] competitors are passing on some of those prices to the consumer by increasing the price of steel.  
[02:05:02] This steel meal now has an option. He can either grab market share by maintaining lower prices so multiply the markup by bigger quantities profits go up or he can increase prices and increase the markup because he hasn't it  
[02:05:15] doesn't face a higher uh input cost. That's a domestic producer subsidy the tariff and so you can see it here. Now the only ones that we found investment domestic investment responses where they were planning to increase  
[02:05:30] domestic investment was coming from the positively affected firm. So you can imagine doing you know here I'm only going to do a few but you can imagine doing this systematically for every firm around the world for different policy instruments. So this is for us is an  
[02:05:45] ongoing project. Um, but it's part of getting at some of the same mechanism, but in a very different way and figuring out other ways to test the theory that isn't just the hard data. Uh, it has  
[02:05:58] lots of other issues. So, let me mention a few. There are advantages and disadvantages. What are pretty obvious ones? Well, first, a lot of this isn't quantitive.  
[02:06:08] I'm planning to increase the prices. Yeah, but by how much? Very often we're interested in a more precise assessment of okay, you're passing on some of the higher input prices, but is the pass through 5% 90% that's pretty different.  
[02:06:22] In general, we find that um that isn't great in this setup. um not or at least not primarily because of a limitation of AI in the quality of being able to  
[02:06:34] extract the information from the text but just because the text itself tends to not to be that quantitive. When CEOs speak, they tend to speak in we plan to increase prices. Maybe they might  
[02:06:46] qualify a lot, little, slow, fast, but they don't tend to speak like an economist that says with a pass through of 7. That that doesn't tend to happen.  
[02:06:56] And so that's one limitation. Um the other limitation which I think I want I you know this is not a class in econometrics and I'm not going to have time to do it justice but it should be very interesting to you both if you want  
[02:07:09] to pursue applied econometrics as a field uh but also if you're just going to be a user like me you have to think hard about the noise and the bias in the data um this has always been there like I wrote a paper with surveys 10 years  
[02:07:24] ago and human surveys felt like this is just a lot of good information in there and there's a lot of noise and bias and unfortunately is unlikely that it's classic measurement error. If it's classic measurement error, there's plenty of well-known techniques to deal  
[02:07:38] with it. But this is not entirely likely. In fact, I would say it's probably pretty unlikely. And when you look at the noise in this data, it's not the level of noise that you put a footnote early in the paper and say, well, all large data sets have some  
[02:07:53] noise. I'm going to ignore it. This is substantial. Um so there is many different ways to deal with it and it's an open question right now. There are structural techniques where you try to  
[02:08:05] characterize the bias and correct. It's not easy. Um there are more informal techniques of run these 50 different ways with different models just to get get a sense of whether the  
[02:08:19] results are reliable. They have the disadvantage that it's like doing robustness checks in papers. How much did you push it? It's a pretty soft sense. There isn't a pre-anned notion of did you explore it  
[02:08:32] extensively enough nor any guarantee that the particular variations that you did are enough. Maybe it breaks in other ways. And particularly with AI, that's important because one of the things that I'm sure has become obvious to you is  
[02:08:45] that the way these models work doesn't have a lot of, you know, human generalizability. Like if you're a PhD student at Stanford and you pass Daryl Duffy's class, uh, you tend to pass mine. Uh, we're pretty sure if you tend  
[02:09:00] to be very good at doing some super high level stuff, uh, you're pretty good at doing other things. We tend to be able to predict that. AI is not like that.  
[02:09:10] You might be great at a particular activity and then fail on trivial dimensions and that we've seen all the time. I can give you a couple of good examples. One that occurred to us was  
[02:09:24] very early we were comparing uh this graph actually a simpler version of here this one fraction of firms that report  
[02:09:37] being affected by tariffs and Jesse my co-author who has written with t before had done a very simple let me do it with bagrams and see if we can get a similar result and we were finding that there  
[02:09:49] were many more flags in biograms and so we said oh we must have a mistake AI isn't working what had actually happened is that AI are idiot proof the two of us uh because we had given it a precise  
[02:10:01] definition we had said a tariff is an attacks on import imposed by a government I can't believe this because I'm Italian so I should know but what is a very  
[02:10:14] common use of the word tarifa if you're a Latin country a tarifa is a fee schedule. Mobile phones, Netflix, they all talk about tariffs because they're talking about the fee schedule to  
[02:10:28] customers. So if I simply look for number of firms mentioning tariffs, I would find many many more and they were all utility companies. AI had figured out that that use of the word tariff  
[02:10:41] isn't attacks on imports and I excluded them. Instead, when we were simply doing biograms, a biograms as you mentioned the word tariff, that's it. So that's a great example of this was substantially more sophisticated than even the two  
[02:10:54] authors were and sort of figured it out. On the other hand, he failed spectacularly in dimensions that you wouldn't expect. For example, I played around with increasing and decreasing sanctions,  
[02:11:08] but the precise wording that people use in documents for the word increasing assumption, it's hard to pick up and it just struggles. Or uh it might struggle  
[02:11:20] with export controls versus um impediments to exports that come from uh climate legislation or consumer safety.  
[02:11:30] there's ways in which he fails are hard to predict given how good it does in other things. So it's just a good topic but particularly here at Stanford I hope you're experimenting with this and the ability as a student to generate your  
[02:11:44] own data for papers it's a pretty incredible opportunity this is going to change the field I mean I remember you know look look Schiller was doing surveys by paper in the 1980s in finance that's why he has a Nobel Prize uh by  
[02:11:58] the 2010s the technology to do surveys on humans online had scaled up enormously And you could do them like you know Stephanie for example Stanchva has done amazing work on this with the  
[02:12:10] changing technology the ability to use AI as a classifier and generate data out of unstructured video text it's just enormous and as a student that's a it's just a great area um but he has to you  
[02:12:24] know he comes with his own baggage and things that you need to understand and we're at the beginning of that process but I I I hope you get involved I think it's I find it tremendously interesting Interesting.  
[02:12:36] Okay. So, actually this data we did put it online. So, if you're interested in playing around um with research on tariffs, export controls, and sanctions, all of the data is online and we're going to keep putting it online. Uh you know, again, the the last thing I will  
[02:12:51] mention is the cost to doing this kind of research um is going to keep varying a lot. When we started this was very expensive, has already come down tremendously.  
[02:13:02] probably not close to a PhD student budget but actually pretty affordable.  
[02:13:08] Uh but one of the things that I'm keen on particularly here at Stanford is to facilitate your work on this. So a there's lots of limitations that come from proprietary data servers that you cannot have access to  
[02:13:22] but there's a lot that you do for example a lot of text is routinely available. It's public. The congressional record, you can download it. Uh earning calls, they're on commercial data sets are pretty access to you. The APIs, the server cost is  
[02:13:37] becoming much more routine. So, don't get discouraged from everything looks like I need to be, you know, anthropic to do it. A lot of this can be done on your laptop  
[02:13:50] being smart and just thinking about good questions and good ways to implement it.  
[02:13:55] So if you're interested, please come and speak to me, to other people here.  
[02:13:58] There's lots of people that are interested, but it it really I I want to encourage you not to think of it as off limits, to think of it instead as exactly what you should be doing if you're interested and to jump on it.  
[02:14:09] Okay, let me u stop here. So look, I just cannot think of a question of this size, at least during my career, this is one of the most exciting things I've done. It just feels like it's a topic that no matter what this is going to  
[02:14:23] shape the next 20 years of economics like the competition between the US and China how we're going to rearrange the world order this is going to be pretty consequential and it's one where there's just a ton of good economics to do a lot  
[02:14:36] of it is you know using basic good tools of macro trade political economy the data this is just the beginning like there's going to be a whole wave of papers and agenda in in this area. And  
[02:14:50] so today I wanted to give you a sense of the research that I'm doing. A lot of the class has been things that are, you know, the class six. This for me is what I do every day, all day. And so if you're interested, come and speak to me and or go speak to Steve. There's lots  
[02:15:05] of good people. Go to Hoover. There's lots of interesting people there around campus. There are just a lot of people that are interested in these issues and are very good. Uh and hopefully you you get excited and you want to contribute.  
[02:15:18] um things that I'm I'm working on. We're working on a paper with multiple hedgeimons. Everything I've done today, it's just the one himon at a time, but clearly in the world right now there are two. In the past, there have been two or more. How does that change the game? Uh  
[02:15:31] we're thinking about political economy. One of the things that I'm very interested in is the relationship between uh the conflicts within the firms and his own government. Uh hidden in what um I discussed today, you could see the conflict. uh think of for  
[02:15:46] example Nvidia in that model. What Nvidia would like to do is sell to everybody, restrict quantities and increase the markup just a monopolist.  
[02:15:55] What the US would like to do is have Nvidia sell cheap to everybody but exclude China. So there's a clear conflict between the national security value of the firm and the value to the  
[02:16:08] shareholders. And how does that lead to lobbying? Similarly, I get asked a lot particularly on policy panels whether the domestic political economy of the country matters. There is the sense out there that China has a big advantage in  
[02:16:21] playing this game because you can directly dictate to stateown enterprises what to do. Okay. So in the formal context of the model you might think that I assume that the edge had no limits on what he could ask domestically at the firm. There might be  
[02:16:36] participation constraints there too and there might be differences in the political system. But then you also have to think about the opposite which is uh the incentives to innovate might depend on whether or not the government is going  
[02:16:51] to commandeer you later. Uh and so whether or not you have a sector that is a choke point that is technologically very advanced might depend on commitment not to overuse it. If we tell Silicon Valley that every time they're going to have a great discovery, the US will  
[02:17:05] commander it and the interest on national security, we can all agree that the incentives to innovate are going to go down at home. So there's a lot there that hasn't been done. There's a lot with the duality with the military.  
[02:17:18] Um yet I you know I intentionally like if you think of political science uh you know Joseph Nia had this thing like soft power like the ability to convince people that you're doing the right thing or you know they all like California  
[02:17:32] because they grew up watching Baywatch uh that kind of thing like then you have like economic power that's sort of that's what we explore but then there's like real hard power like military threats military intervention all the  
[02:17:46] way to like nuclear we focus on this area not because I felt that the other two are not interesting but more because that's what we had something to say uh and you know you focus on your own marginal product but there's a lot of interesting stuff there what are other things that are interesting you know  
[02:18:00] today I did a little bit of finance but it was really basic finance like payment system but there is lots of interesting stuff I mentioned at the beginning buying firms abroad think about the model the lrange multipliers on those  
[02:18:14] participation constraints tell you the marginal value of having more power on that entity. That's your shopping list.  
[02:18:22] I think of FDI as um another way to get power rather than doing this whole game of inducing you to do what I want. I can just buy your equity. If I buy your equity and I own you, I'll dictate the  
[02:18:35] answers. I'll just get you to do I mean if I want ASML not to sell to China I can bully the Netherlands government and ASML and get them not to sell or I can just buy the firm outright and once I own it I can decide not to sell. So  
[02:18:48] that's a a very different model of FDI and where do you want to buy? There's things like causal evidence. Um, you know, clearly most threats are endogenous,  
[02:19:02] but can we have clean evidence that indeed uh when I cut you off, this is the outcome? Right now, we're doing it to a structural model. We're inferring from the model that that's what happened. But can we do it uh with  
[02:19:16] reduce form methods? And then we discussed, you know, ultimately these have to be quantity question. I liked your question of you know a lot of stuff is not going to work a lot of stuff isn't strategic a lot of stuff isn't that important but we have to show  
[02:19:31] reliable sort of models to tell the government when industries come and lobby for you should subsidize me because I'm very important for the national interest we should have as economists an answer says actually no like we're pretty confident  
[02:19:45] that you're playing vanilla or similarly I think 98% to maybe of what the trade that we do. It's totally not sensitive and then we shouldn't worry about it for national security, but we need to have better statements and right now we're  
[02:19:59] not there. So hopefully this gives you a sense of um where you could get involved. Uh let me stop here. I promise you there was a long reading list so I wasn't kidding. Uh it's here. Um you can read it at your own pleasure during the  
[02:20:13] summer. Um, [music] everybody