Complete Intelligence


The Unbeatable Artificial Stock Market

Show Notes

MG: The Lead Lag Report joining us for the hour here is Tony Nash of Complete Intelligence has found a lot of people that I respect following. Tony, I saw a few people saying they were excited to hear what Tony has to say. So hopefully we’ll have a good conversation here.

Tony for those who aren’t familiar with your background talk about who you are how’d you get involved in the data side of markets and forecasting in general. And what you’re doing with Complete Intelligence.

TN: Sure, Michael. First of all, thanks for having me. I have followed you for probably 10 or 15 years.

MG: I am very sorry for that I am very very sorry for that.

TN: But yeah so, I got involved in data way back in the late 90s when I was in Silicon Valley and I built a couple of research firms focused on technology businesses. I then took about probably eight years to become an operator. I did a turnaround in Asia of a telecom firm. I built a firm in Sri Lanka during the Civil War and then I started down the research front again. I was the Global Head of Research for the Economist and I was the Asia Head of Consulting for a company called IHS Markit which is now owned by S&P and then after that I started Complete Intelligence.

So, you know my background is really all about data but it’s also all about understanding the operational context of that data. And I think it’s very hard for people to really understand what data means without understanding how people use it.

MG: Okay. So that’s maybe a good direction to start with that point about context with data because I think part of that context is understanding what domains data is more appropriate for forecasting and others. Right? So, I always made this argument that there are certain domains in particular when it comes to, I would argue investing that have sort of a chaotic system element to them. Right? Where small changes can have ripple effects. So, it’s hard to necessarily to sort of make a direct link between a strong set of variables and the actual outcome because there’s always a degree of randomness. Whereas, something that’s more scientific right that doesn’t have that kind of chaos theory element is it’s clearer.

So, talk about that point about context when it comes to looking at data. And again, the kind of domains where data is more appropriate to really have more conviction in than others.

TN: Yeah. Okay. So, that’s a great place to start. So, the first thing I would say is take every macro variable that you know of and throw it out the window. It’s all garbage data 100 of it. Okay? I would never trade based on macro data.

We’ve tested macro data over the years and it’s just garbage. It doesn’t matter the country. You know we hear people saying that China makes up their data. Well, that may be true you can kind of fill in the blank on almost any country because I don’t know how much you guys understand about macro data. But it is not market clearing data. Okay? Like an equity price or a commodity price.

Macroeconomic data is purely academic made-up data that is a proxy for activity. It’s a second or third derivative of actual activity by the time you see, say, a CPI print which is coming out tomorrow. Right? And it’s late and it’s really all not all that meaningful. So, I wouldn’t really make a trade or put a strategy together based on macro data even historical macro data. Every OECD country revises their data by what four times or something.

So, you see, a print for CPI data tomorrow that’s a preliminary print and that’s revised several times before it’s put on quote-unquote actual. And so, you know, you really can’t make decisions using macroeconomic data beyond a directional decision. Okay? So, if you follow me on Twitter, you see I’m very critical macro data all the time. I’m very sarcastic about it.

I think the more specific you can get… You know if you have to look at say national data or macroeconomic data, I would look at very low-level data the more specific you can get the better. Things like household surveys or you know communist and socialist countries. Chinese data at the very specific level can be very interesting. Okay? Government data the high-level data in every country I consider it garbage data in every country. So, you’re looking at very low-level very specific government or multilateral data, that’s interesting.

The closer you get to market clearing data the better because that’s a real price. Right? A real price history on stuff is better and company data is the best. And of course, company data is revised at times but that really helps you understand what’s happening at the kind of firm level. And what’s happening at the transaction level. So, you know, those are the kind of hierarchies of data that I would look at.

MG: So, okay this is a great. That’s a great point you mentioned that it’s you said very these variables is macro variables they’re proxies for activity. Right? They’re really more proxies for narratives. Right? Because and that’s where I think… You mentioned sarcasm almost 99 of my tweets at this point are sarcasm because when Rome is burning, what else I’m not going to do except joke about it. Right? Because I can’t change anything. Right?

So, and to that point I share a lot of that cynicism around data that people will often reference in the financial media that sounds really interesting, sounds like it’s predictive but when you actually test it to your point, you throw it out because it doesn’t work. Right? There’s no real predictive element to it.

So, we’ll get into some of the predictive stuff that you talk about but I want to hit a little bit on this market clearing phrase you kept on using. Explain what you mean by market clearing.

TN: Data is where there is a buyer and a seller.

MG: To actual prices of some asset class or something like that.

TN: Yep. That’s right.

MG: Okay. So, that makes sense. Okay. Now again I go back to the certain domains that data is more clear in terms of cause and effect and getting a sense of probabilities the challenge with markets. As we know is that the probabilities change second by second because not only does that mean meaningless data change second by second but the market clearing data changes second by second. Right? Going back to that point.

So, with what you do with Complete Intelligence, talk us through a little bit. What are some of the variables that you tend to find have some predictive power? And how do you think about confidence when it comes to any kind of decision made based on those variables?

TN: Sure. Okay. So, before I do that let me get into why I started Complete Intelligence because if none of you have started a firm before don’t do it. It’s really really hard so…

MG: From the people in the back because I got to tell you I’m an entrepreneur, I’m going through. And all you got is people on Twitter kicking you when you’re down when it’s the small sample anyway.

TN: Absolutely. So, I was where I had worked for two very large research firms The Economist and IHS Markit. And I saw that both of them claimed to have very detailed and intricate models. Okay? Of the global economy industries, whatever. Okay? For all of the interior models. And I have never spoken with a global research firm a data firm that is different from this. And if I’m wrong then somebody please correct me. But at the end of that whole model pipeline is somebody who says “no that’s a little bit too high” or “a little bit too low” and they change the number. Okay? To whatever they wanted it to be in the first place. So, and I tell you 100% of research firms out there with forecasts today have a manual process at the end of their quote-unquote model. A 100% of them. Again, if there’s somebody else that doesn’t do that, I am happy to be corrected. Okay? But I had done that for a decade and I felt like a hypocrite when I would talk to clients.

So, I started Complete Intelligence because I wanted to build a 100% machine driven forecasts across economics, across market, across equities, across commodities, across currencies. Okay? And we’ve done that. So, we have a multi-phase, multi-layer machine learning process that takes in billions of data items. We’re running trillions of calculations every week when we reforecast our data. Right? Now the interval of our forecast is monthly interval forecast. So, if people looking at daily prices that’s not what we’re doing now. Okay? We will be launching daily interval forecasts. I would say probably before the end of the year to be conservative but we’re doing monthly interval forecasts now.

Why is everything I’ve said is meaningless unless we measure our error. Okay? So, for every forecast that we do. And if you log into our website, you can see whether it’s the gold price, the S&P 500, USD, JPY, molybdenum or whatever. We track our error for every month, for everything that we do. Okay? So, if you want to understand your risk associated with using our data it’s there right in front of you with the error calculations. Okay? It’s only fair, If I’m gonna say sell you a forecast, you should be able to understand how wrong we’ve been in the past, before you use that as a decision-making input.

MG: Well, maybe just add some framework on that because I think that’s interesting. So, what you call error I call luck. Right? Because luck is both good or bad. I always make that point that with any equation any set of variables you’re going to have that error is the luck component that you can’t control. And that doesn’t necessarily mean that the equation is wrong. Right? It’s just means that for whatever reason that error in that moment in time was higher or lower than you might otherwise want. Okay?

TN: There is no such thing as zero error. And anybody who tells you that they have zero error is obviously they’re an economist and they don’t understand how markets work. So, there is always error in every calculation.

So, the reason we track error is because that serves as a feedback loop into our machine learning process. Okay? And we have feedback loops every week as we and what we’re doing right now is every Friday end of day. We will download global data process over the weekend have a new forecast on Monday morning. Okay? And so all of that error whether it’s near-term error, short-term error or say medium-term error, we feed that all back in to help correct and understand what’s going on within our process. And we have like I said, we have a multi-phase process in our machine learning platform. So, error is simply understanding the risk associated with using with using our platform.

MG: Right, which is basically how apt is a thing that you’re forecasting to that error which is again luck good or bad. I’m trying to put in sort of a qualitative framework also because I think… Yeah, there’s errors in life obviously, too. Right? And so, when they’re good or bad. But you know those elements.

TN: Right. But here’s what I would and I don’t know, I don’t want to dispute this too much but I think there is. So, you use the word luck and that’s fine but I think luck has a bit to do with the human element of a decision. Okay? We’re using math and code there’s zero human interaction with the data and with the process. And so, I wouldn’t necessarily call it luck. I mean, it literally is error like our algorithms got it wrong. So, if you want to call luck that’s absolutely fine but I would say luck is more of a human say an outcome associated with a human decision. More than something that’s machine driven that’s iterating. Again, we’re doing trillions of calculations every week to get our forecasts out there.

MG: Yeah, no that’s fair and maybe for the audience, Tony. Explain what machine learning is now.

TN: Sure.

MG: I once developed an app called “How Edition”. I was having dinner with the head developer once and he said he just came back from a conference about machine learning and he was just basically well, having drinks with me laughing and joking saying everybody use this term machine learning but it’s really just regression analysis. Right? So, talk about machine learning what is actual machine learning? How important is recent data to changes in the regression? Because I assume that’s part of the sort of dynamic nature of what you do just kind of riff on that for a bit.

TN: Okay. So, when I first started Complete Intelligence, I was really cynical about AI. And I spoke to somebody in Silicon Valley and asked the same question: what is AI? And this person said “Well AI is everything from a basic I say, quadratic equation upward.” I’m not necessarily sure that I agree that something that simple would be considered artificial intelligence. What we’re really doing with machine learning is there are really three basic phases. Okay? You have a preprocess which is looking at your data to understand things like anomalies, missing data, weird behavior, these sorts of things. Okay? So, that’s the first phase that we look at to be honest that’s the hardest one to get right. Okay?

A lot of people want to talk about the forecasting methodologies and the forecasting algorithms. That’s great and that’s the sexy part of ML. But really the conditioning and the pre-process is the is the hardest part and it’s the most necessary part. Okay? When we then go into the forecasting aspect of it, we’re using what’s called an ensemble approach. So, we have a number of algorithms that we use and let’s say they’re 15 algorithms. Okay? That we use we’re looking at a potential combinatorial approach of any individual or combination of those algorithms based on the time horizon that we’re forecasting. Okay?

So, we’re not saying a simple regression is the way to go we’re saying there may be a neural network approach, there may be a neural network approach in combination with some sort of arima approach. We’re saying something like that. Right? And so, we test all of those permutations for every historical period that we’re looking at.

So, I think traditionally when I look back at kind of quote-unquote building models in excel, we would build a formula and that formula was fairly static. Okay? And every time you did say a crude oil forecast you had this static formula that you set your data against and a number came out. We don’t have static formulas at all.

To forecast crude oil every single week we start at obviously understanding what we did in the past but also re-testing and re-weighting every single algorithmic approach that we have and then recombining them based upon the activity that happened on a daily basis in that previous week. And in the history. Okay?

So, that’s phase two the forecasting approach and then phase three is the post process. Right? And so, the post process is understanding the forecast output. Is it a flat line? Right? If it’s a flat line then there’s something wrong. Is it a straight line up? Then that there’s something you know… those are to use some extremes. Right? But you know we have to test the output to understand if it’s reasonable. Right? So, it’s really an automated gut check on the reasonableness of the outcome and then we’ll go back and correct outliers potentially reforecast and then we’ll publish. Okay?

So, there are really three phases to what we do and I would think three phases to most machine learning approaches. And so, when we talk about machine learning that’s really what we’re talking about is that that really generally three-phase process and then the feedback loop that always goes back into that.

MG: Yeah. No that makes sense. Let’s get…

TN: That’s really boring after a while.

MG: No, no, no but I think that’s it’s part of what I want to do with these spaces is try to get people to understand you know beyond sort of just the headline or the thing that is thrown out there. As a term to what does that actually mean in practice you don’t have to know it fully in depth the way the that you do. But I think having that context is important.

TN: I would say on the idea generation side and on the risk management side right now. Okay? Now the other thing that I didn’t cover is obviously we’re doing markets but we also do… we use our platform to automate the budgeting process within enterprises. Okay? So, we work with very large organizations and the budget process within these large organizations can take anywhere from say four to six months. And they take hundreds of people. And so, we take that down to really interacting with one person in that organization and we do it in say less than 24 hours. And we build them a continuous budget every month.

Once accounting close happens we get their new data and then we send them a new say 18-month forward-looking forecast for them. So, their FPA team doesn’t have to dig around and beg people for information and all that stuff. So, some of this is on the firm event could be on the firm evaluation side, as well. Right? How will the firm perform? Nobody’s using us for that but the firms themselves are using that to help them automate their budgeting process. So, some of that could be on this a filtering side and the idea generation side, as well.

So, we do not force our own GL structure onto the clients. We integrate directly with their SAP or Oracle or other ERP database. We take on their GL structure at whatever levels they want. We have found that there is very little deterioration from say, the second or third level GL to say the sixth or seventh level GL, in terms of the accuracy of our forecast. And when we started doing this it really surprised me. We do a say a team level forecast for 10, 12 billion organizations, six layers down within their GL. And we see very little deterioration when we go down six levels than when we do it at say two levels. Which is you know it really to me it speaks to the robustness of our process but would we consider Anaplan a competitor not really, they’re not necessarily doing the kind of a budget automation that we’re doing at least, that I’m aware of. I know that there are guys like Hyperion who do what we’re doing but again their sophistication isn’t necessarily. What we’re doing and they do a great job and Hyperion is a great organization. I think Oracle gave them a new name now but they’re not necessarily using the same machine learning approaches that we’re using. And our clients have told us that they don’t get the same result with using that type of say ERP originated or ERP add-on budgeting process.

Yep. So, I would say we can’t we can do company-specific information for a customer if that’s what they want. Okay? We don’t necessarily have that on our platform today aside from say individual ticker symbols. Okay? But we’re not forecasting say the P&L of Apple or something like that or the balance sheet of Apple. Something we could do in a pretty straightforward manner but we do that on a customer-by-customer basis.

So, what we’re forecasting right now are currency pairs, commodities about 120 commodities and global equity indices. Okay? We are Beta testing individual equity tickers and we probably won’t introduce those fully on the platform until we have our daily interval forecast ready to go to market. But those are still we’re still working some kinks out of those and we’ll have those ready probably within a few months.

MG: Okay. So, let’s talk about commodities here for a bit tonight. Obviously, this is where a lot of people’s attention has gone to. What kind of variables and I know you said you have a whole bunch of variables that are being incorporated here but are there certain variables in particular when it comes to oil and other commodities that have a higher predictive power than others.

TN: There are I think one of the stories that I tell pretty often and this really shocks people is when we look at things like gold. Okay? I’m not trying to deflect from your oral question but just to you know we’ve spoken with the number of sugar traders over the years. Okay? And so, we tell them that say the gold price and the sugar price there may not necessarily be a say short term say correlation there but there is a lot of predictive capability there and we talk them through why. And I think the thing that we get out of the machine learning approach and we cast a wide net. We’re not forcing correlations is that we’ll find some unexpected say drivers. Although drivers implies a causal nature and we’re not trying to imply causality anywhere. Okay?

We’re looking at kind of co-movement in markets over time and understanding how things work in a lead lag basis with some sort of indirect causality as well as say a T0 or current state movement. So, with crude oil you know there are so many supply side factors that are impacting that price right now, that I can’t necessarily point to say another commodity that is having an impact on that. It really is a lot of the supply side and sentimental factors that are impacting those prices right now.

MG: That makes a lot of sense. And I’m curious how did you mention it’s I think the intervals once a month. Right? So, given the speed with which inflation has moved and yields have moved how does a machine learning process adapt to sudden spikes or massive deltas in in variable movement. Right? Because there’s always a degree of randomness going back to error. Right? And you can make an argument that the larger move is the that may actually be more error but I think that’s an interesting discussion.

TN: So, I’ll tell you where we were say two years ago when 2020 hit versus today. Okay? So, in March of 2020, April 2020 everything fell apart. I don’t think there were any models that caught what was going to happen. It was an exogenous event that hit markets and it happened very quickly. So, in June, I was talking with someone who is with one of the largest software companies in the world and they said “Hey has your AI caught up to markets yet because ours is still lost” And you guys would be shocked if I told you who this was because you would expect them to know exactly what’s going to happen before it happened. Okay? I’ll be honest I think it was all of them but the reality is you know Michael you where you were saying that ML is just regression analysis.

I think a lot of the large firms that are doing time series forecasting really are looking at regression and derivatives of regression as kind of their only approaches because it works a lot of the time. Right? So, we had about a two-month delay at that point and part of it was because… So, by June we had caught up to the market. And we had started in February to iterate twice a month, we were doing once a month; I hope you guys can understand with machine learning two factors are we’re always adjusting our algorithms. Okay? We’re always incorporating new algorithms. We’re always you know making sure that we can keep up with markets because you cannot be static in machine learning. Okay? The other thing is we’re always adding capacity why? Because we have to iterate again and again and again to make sure that we understand the changes in markets. Okay?

So, at that time we were only iterating twice a month and so it took us a while to catch up. Guys like this major technology firm and other major technology firms they just couldn’t figure it out. And I suspect that some of them probably manually intervened to ensure that their models caught up with markets. I don’t want to accuse any individual company but that temptation is always there. Especially, for people who don’t report their error. The temptation is always there for people to manually intervene in their forecast process. Okay?

So, now, today if we look for example at how are we catching changes in markets. Okay? So, if I look at the S&P 500 for April for example, our error rate for the S&P 500 for April I think was 0.6 percent. Okay? Now in May it changed it deteriorated a little bit to I think four or six percent, I’m sorry I don’t remember the exact number offhand but it deteriorated. Right? But you know when there are dramatic changes because we’re iterating at least once a week, if not twice a week we’re catching those inflections much much faster. And what we’re having to do, and this is a function of the liquidity adjustments, is where in the past you could have a trend and adjust for that trend and account for that trend. We’re really having to our algorithms are having to select more methodologies with recency bias because we’re seeing kind of micro volatility in markets. And so again…

MG: So, kind of like the difference between a simple moving average versus like an exponential moving average. Right? Where you’re waiting the more recent data sooner.

TN: It could be. Yeah.

MG: Right.

TN: Yeah. That’s a very very simple approach but yeah it would be something like that, that’s right. Yeah. What so when we work with enterprise customers that level of engagement is very tight because when we’re getting kind of the full set of financial data from a client obviously, they’re very vested in that process. So, that’s different from say a small portfolio manager subscribing to RCF futures product where we’re doing forecasts and they have their own risk process in place. And they can do whatever they want with it. Right? But again, with our enterprise clients we are measuring our error so they can see the result of our continuous budgeting process. Okay?

So, if we’re doing let’s say, we launch with a customer in May, they close their mate books in June get them over to us redo our forecast and send it over to them and let them know what our error rate was in May. Okay? So, they can decide how we’re doing by department, by team, by product, by whatever based upon the error rates that we’re giving at every line item. Okay? So, they can select and we’re not doing kind of capital projects budgets we’re doing business as usual budgets so they can decide what they want to take and what they don’t want to take. It’s really up to them but we do talk through that with them and then over time they just start to understand how we work and take it on within their own internal process.

MG: So, back a little bit Tony. So, you mentioned you do this machine learning forecasting work when it comes to broad economics, markets and currency; of those three which has the most variability and randomness in other words which tends to have a higher error? Whenever you do any kind of machine learning to try to forecast what comes next?

TN: I would say it depends on the equity market but probably equity markets when there are exogenous shocks. So, our error for April of 2020 again, we don’t hide this from anybody it was not good but it wasn’t good for anybody. Right? And so, but in general it depends on the equity market but some of the emerging equity markets, EM equity markets are pretty volatile.

We do have some commodities like say rhodium for example. Okay? Pretty illiquid market, pretty small base of people who trade it and highly volatile. So, something like rhodium over the years our air rates there have not necessarily been something that we’re telling people to use that as a basis to trade but obviously, it’s a hard problem. Right? And so, we’re iterating that through our ML process and looking at highly volatile commodities is something that we focus on and work to improve those error rates.

MG: Here, I hope you find this to be an interesting conversation because I think it’s a part of the of the way of looking at markets, which not too many people are themselves maybe using but is worth sort of considering. Because I always make a point that nobody can predict the future but we all have to take actions based on that unknowable future. So, to the extent that there might be some data or some conclusions that at least are looking at variables that historically have some degree of predictive power. It doesn’t guarantee that you’re going to necessarily be better off but at least you have something to hang your hat on. Right? I think that’s kind of an aspect to investing here.

Now, I want to go a little bit Tony to what you mentioned earlier you had lived abroad for a while in Europe. And when I was starting to record these spaces to put up on my YouTube channel the first one, I did that on was with Dan Arvis and the topic of that space was around this sort of new world order that seemed to be shaping up. I want you to just talk from a geopolitical perspective how you’re viewing perhaps changing alliances because of Russia, Ukraine. And maybe even dovetail that a little bit into the machine learning side because geopolitics is a variable. Which is probably quite vault in some periods.

TN: Yeah, absolutely. Okay. So, with the evolving geopolitical order I would say rather than kind of picking countries and saying it’s lining up against x country or lining up with x country or what country. I would say we’ve entered an era of opportunistic geopolitics. Okay? We had the cold war where we had a fairly static order where people were with either red team or blue team. That changed in the 90s of course, where you kind of had the kind of the superpower and that’s been changing over the last say 15 years with say, China allegedly becoming kind of stronger and so on and so forth. So, but we’ve entered a fairly chaotic era with say opportunistic macroeconomic relation or sorry, geopolitical relationships and I think one of the kinds of top relationships that is purely opportunistic today is the China-Russia relationship.

And so, there’s a lot of talk about China and Russia having this amazing new relationship and they’re deep. And they’re gonna go to war together or whatever. We’ve seen over the past say three, four months that’s just not the case. And I’ve been saying this for years just for a kind of people’s background. Actually, advised the Chinese government the NDRC which is the economic planning unit of the central government on a product or on an initiative called the belt and road initiative. Okay? I did that for two years. I was in and out of Beijing. I never took a dime for it. I never took expense reimbursement just to be clear, I’m not a CCP kind of pawn. But my view was, if the Chinese Government is spending a trillion dollars, I want to see if I can impact kind of good spend for that. So, I have seen the inside of the Chinese Government and how it works and I also in the 80s and 90s spoke Russian and studied a lot on the Russian Government and have a good idea about how totalitarian governments work.

So, I think in general if we thought America first was offensive in the last administration then you really don’t want to learn about Chinese politics and you really don’t want to learn about Russian politics because they make America first look like kindergarten. And so, whenever you have ultra-ultra-nationalistic politics, any diplomatic relationship is an opportunistic relationship. And I always ask people who claim to be China experts but say please tell me and name one Chinese ally. Give me one ally of China and you can’t, North Korea, Pakistan. I mean, who is an ally of China there isn’t an ally of China.  There is a transactional opportunistic relationship with China but there is not an ally with China.

And so, from a geopolitical perspective if you take that backdrop looking at what’s happening in the world today it makes a whole lot more sense. And a lot of the doomsayers out there saying China is going to fall and it’s going to have this catastrophic impact. And all this other stuff, the opportunism that we see at the nation-state level pervades into the bureaucracy. So, the bureaucracy we hear about Xi Jinping. And Xi Jinping is almost a fictional character. I hate to be that extreme on it but there is the aura of Xi Jinping and there is the reality of Xi Jinping, just a guy, he’s not Mao Zedong. He doesn’t have the power that supposed western Chinese experts claim that he has. He’s just a guy. Okay?

And so, the relationships within the Chinese bureaucracy are purely transactional and they are purely opportunistic. So again, if you take that perspective and you look at what’s happening in geopolitics, hopefully you can see things through a different lens.

MG: Now, I’m glad you’re framing that in those terms because I think it’s very hard for people to really understand some of these dynamics when it’s almost presented like a like the story for a movie. Right? For what could be a conflict to come by the media because and it’s almost overly simplified. Right? When you hear this type of talk. So again, I want to go back into how does that dovetail into actual data. Right? Maybe it doesn’t at all. When you have some of these dynamics and you talk about market clearing data, you’re going to probably see mark movement somewhat respond off of geopolitical changes. Talk about anything that you’ve kind of seen as far as that goes and how should investors consider geopolitical risk or maybe not consider geopolitical risk?

TN: Yeah, I think, well when you see geopolitical adjustments today all that really is… I don’t mean overly simplified but it’s a risk calibration. Right? So, you know Russia invades Ukraine, that’s really a risk calibration. How much risk do we want to accept and then what opportunities are there? Right?

So, when you hear about China, you have to look at what risk is China willing to accept for actions that it takes? Keeping in mind that China has a very complicated domestic political environment with COVID shutdown, lockdowns and all of this stuff. So, having worked with and known some really smart Chinese bureaucrats over the years, these guys are very concerned with the domestic environment. And I don’t although there are idiot you know generals and economists here and there who say really stupid stuff about China should take over TSMC and China should invade Taiwan, these sorts of things. My conversations over the years have been with very pragmatic and professional individuals within the bureaucracy.

So, do I agree with their policies? Not a lot of them but they are well thought out in general. So, I think just because we hear talk from some journalist in Beijing who lives a very sheltered life about some potential thing that may happen. I don’t think we necessarily need to calibrate our risk based on the day-to-day story flow. I think we need to look at like… so there’s a… I’m sure you all know who Leland Miller is in China beige book like?

MG: Yeah, he’s not too long ago.

TN: Yeah. He has a proxy of the Chinese economy and that’s a very interesting way to look at an interesting lens to look through China or through to look at China or whatever. But so, I think that the day-to-day headlines, if you follow those, you’re really just going to get a lot of volatility but if you try to understand what’s actually happening, you’ll get a clearer picture. It’s not necessarily a connection of a collection of names in China and the political musical chairs, it’s really asking questions about how does China serve China first. What will China do to serve China first and are some of these geopolitical radical things that are said do they fit within that context of China serving China first? So, that’s what I try to look at would I be freaked out if China invaded Taiwan? Absolutely. I think everybody would right but is that my main scenario? No, it’s not.

MG: In terms of the data inputs on the machine learning side how granular is the data meaning? Are you looking at where geographically demand might be picking up or is it simply this is what the price is and who cares the source? Because again with hindsight if you knew that the source of China and kind of had a rough sense of the history of Russia-Ukraine maybe that could have been an interesting tell that war was coming.

TN: Yes or No. To be honest it had more to do with the value of the CNY. Okay? And I’ll tell you a little bit about history with the CNY. We were as far as I know, the only ones who called the CNY hitting 6.7 in August of 2019 with a six-month lead time. And so, we have a very good track record with USD-CNY and I would argue that China’s buying early in 2022 had a lot more to do with them from a monetary policy perspective needing to devalue CNY. So, they were hoard buying before they could devalue the CNY and I think that had a lot more to do with their activity than Russia-Ukraine. Okay? And if you notice they’ve made many of their buys by mid-April and once that happened you saw CNY, go to 6.8. Right? It’s recovered a little bit since then but China has needed to devalue the CNY for probably at least nine months. So, it’s long overdue but they’ve been working very hard to keep it strong so that they could get the commodities they needed to last a period of time. Once they had those commodities, they just let the parachute go and they let it do value to 6.8 and actually slightly weaker than 6.8.

MG: The point of the devaluation is interesting. I feel if I had enough space but we were talking about the Yen and what’s happened there. And this observation that usually China will start to devalue when they see the end as itself going through its own devaluation.

How does some of those cross correlations play out with some of the work that on machine learning you’re doing? Because there’s a human element to the decision to devalue a currency. Right? So, the historical data may not be valid I would think because you might have kind of a more humanistic element that causes the data to look very different.

TN: Well, they’re both export lab economies. Right? And we’ve seen a number of other factors dollar strength and we’ve seen changing consumption patterns. And so, yes when Japan devalues you generally see China devalue as well but also, we’ve seen a lot of other activities in on the demand-pull side and on the currency side especially with the US dollar in… I would say over the last two quarters. So, yes, that I would say that the correlation there is probably pretty high but there are literally thousands of factors that contribute to the movement of those of those currencies.

MG: Is there anything recently Tony in the output that machine learning is spitting out that really surprises you? That you know… And again, I understand that there’s a subjective element which is our own views on the world and of course then the pure data. But I got to imagine it’s fascinating sometimes if you’re sitting there and seeing what’s being spit out if it’s surprising. Is there anything that’s been kind of an outlier in in the output versus what you would think would likely happen going forward?

TN: Yeah. You know, what was really surprising to me after we saw just to stick on CNY for a minute because it’s the first thing that comes to mind, when we saw CNY do value to 6.8. I was looking at our forecast for the next six months. And it showed that after we devalued pretty strong it would moderate and reappreciate just a bit. And that was not necessarily what I was hearing say in the chatter. It was kind of “okay, here we go we’re going to go to seven or whatever” but our data was telling us that that wasn’t necessarily going to happen that we were going to hit a certain point in May. And then we were going to moderate through the end of the year. So, you know we do see these bursty trends and then we see you know in some cases those bursty trends continue for say an integer period. But with CNY while I would have on my own expected them. I expected the machines to say they need to keep devaluing because they’ve been shut down and they need to do everything they can to generate CNY fun tickets. The machines were telling me that we would you know we’d see this peak and then we would we would moderate again and it would kind of re-appreciate again.

So, those are the kind of things that we’re seeing that when I talk about this it’s… Oh! the other thing is this: So, in early April we had a we have people come back to us on our forecast regularly who don’t agree with what we’re saying and they complain pretty loudly.

MG: So, what do you say I talk when I hear that because whenever somebody doesn’t agree with the forecast, they are themselves making a fork.

TN: Of course. Yeah. Exactly. Right? Yeah, and so this person was telling us in early April that we’re way wrong that the S&P was going to continue to rally and you know they wanted to cancel their subscription and they hated us and all this other stuff. And we said okay but the month’s not over yet so let’s see what happens this was probably a week and a half in April. And what happened by the end of April things came in line with our forecast and like I said earlier we were like 0.4 and 0.6 percent off for the month. And so that person had they listened to us at the beginning of the month they would have been in a much better position than they obviously ended up being in. Right? And so, these are the kind of things that we see on a… I mean, we’ve got hundreds of stories about this stuff but these are the kind of things that we see on a regular basis. And we mess up guys I’m not saying we’re perfect and but the thing that we when we do mess up, we’re very open about it. Everything that we do is posted on our on our website. Every call we make, every error we have is their wars and all. Okay? And so, we’re not hiding our performance because if you’re using our data to make a trade, we want you to understand the risk associated with using our data. That’s really what it comes down to.

MG: It reminds me of back in 2011 and in some other periods I’ve had similar situations, where I was writing and I was very adamant in saying the conditions favored a summer crash. Right? I was saying that for the summer and the market should be going up and people would say oh where’s your summer crash and I would say this summer hasn’t started. Like it’s amazing how people, I don’t know, what it is, I don’t know if it’s just short-termism or just this kind of culture of constantly reacting as opposed to thinking but it is it is remarkably frustrating.

Going back to your point at the very beginning being entrepreneur don’t do it, that you have to build a business with people and customers who in some cases are just flat out naïve.

TN: That’s all right though. That’s a part of the risk that we accept. Right?

MG: Yeah, the other thing right now that happens with every industry but from the entrepreneur’s standpoint. It’s what you’re doing the likely outcome of your product of your service. You’re trying to communicate that to end clients but then in the single role of the die the guy the end client who comes to you exactly for that simply because they disagree with you know the output, now says I want out.

TN: Oh! Yeah! Well, your where is your summer call from 2011 the analogy today is where is your recession call. Right? So, that’s become the how come you’re not one of us calls right now. So, it’s just one of those proof points and if you don’t agree with that then you’re stupid.

So, I would say you never finish with that there is always a consensus and a something you’re you absolutely, must believe in or you don’t know what you’re talking about.

MG: Yeah, well, thankfully. What you’re talking about so appreciate everybody joining this space Tony the first time you and I were talking. I enjoyed the conversation because I think it said on investing and I encourage you to take a look at Tony’s firm and follow him here on twitter. So, thank everybody. Thank you, Tony and enjoy.