Do media narratives influence financial markets?
Ronnie Sadka: Hi, everyone. Good afternoon. Nice to see you folks. How's everyone feeling? Are we good? I know it's been a long day. I've had a long day. It's always hard to, I think, present at this time of the day. I think the good news is there's some drinks after this to have something to look forward to. Look, this paper I'm going to present today. I have to admit, it's probably the closest paper to a true kind of academic paper that are presented at statewide conferences. I'm super excited about it because we've been working on narratives for a while. And Gideon, my co-author, I think, is here. We've been doing this for so long and we always wanted to write a paper about how narratives drive the markets. And not only we're going to show you, it drives the markets because we talked about it over the years. And today we'll also had a paper talking about FCS and how it drives effects. But we're going to show you that it prices the cross section of stock returns. I've always wanted to write a paper like that. And and I think the results are very strong there, there. So I'd like to share that with you. Okay. So let's let's begin. So some motivation here. I'm glad you think about me when when you think about narratives. I think sometimes about Schiller. He gave a talk here four years ago, and I think that helped us formulate some of our hypotheses. So Schiller also, he cites Samuelson Samuelson's dictum, which basically says that the market or investors are very good in understanding the micro.
So there's so many analysts they're looking at, they're looking at stocks and looking at financial statements, and they're looking at every piece of news comes out on a given stock, but sometimes they just miss the big picture. So they miss the housing bubble. They miss the tech bubble. Okay. They miss these narratives that might be long lasting, the kind of macro, but they're not always paying attention to it or they don't always understand that it's happening. And so one of the things that we think about our work is that we're able to quantify these narratives that maybe people miss. Okay. And Shiller calls this narratives go go viral. And so someone talks about like meme stock is a good example, right? The meme stock, we remember GME and AMC and something goes viral and it affects asset prices. But but the thing is, with these stories or these ideas, it's hard to quantify. It's not like a momentum factor or a value factor that you sort firms on book to market ratio and well, you have a factor is these ideas and how do you quantify them? Shiller looks at this with Ngrams. He looks at it on long over Long horizons. It looks over years and he has observations once a year. We have observations once a day so we can really understand dynamics and how some themes are new. Some things are evergreen, but they change.
And so that's kind of the stuff that we try to quantify to give investors a better idea of how they look like. So in this talk, what I'm going to do is I'm going to focus on quantifying narratives similar, I think, to what to what we showed earlier. But then I'm going to take it to the cross section of stocks. To my understanding, this is really the first paper that does this. So just to give you an idea of what we're going to be talking about today, I'm going to first introduce narratives. Some of it will be repetitive, repetitive, but that's fine, because think about it this way. If you can remember 10% of what I say, I just need to pick one thing and just say it ten times, right? So I'm just going to talk a little bit more about narratives, more ideas, more examples, so that we're comfortable, again with quantifying narratives. And then I'm going to jump to pricing narratives and the cross section. I'm going to form portfolios that track narratives. I'm going to have hundreds of narratives I'm going to do. I'm going to do narrative mimicking portfolios. I'm going to have hundreds of these portfolios, and we're going to see whether they predict returns in the cross section of stocks. And then we're going to talk about the economic interpretation of these findings. All right, so let's start from the basics. What is the what is the data? Do we use what is the Mqtt media stat? What do they do? Well, we gather data every day.
We look at many, many sources of digital media have about 150,000 sources, 5 million articles every week that are relevant to the assets that we are that states and clients care about. So we look at media that talks about firms. We call it corporate reservoir. I'm going to talk about reservoirs in a second. We'll talk about firms. We'll talk about country equity effects. General Press, We also have some social media. Et cetera. So many, many, many different sources. One of the things that we need realized early on is each source may come with their biases. So we need to understand the source bias. We spend so much time trying to understand that I'm not going to spend too much time. I'm just going to give you a flavor, right? So you have, let's say, a day of the week effect. Some firms are always mentioned on a particular day of the week, like technology stocks typically are more mentioned on the weekends. There's some evidence that there's Monday stocks, there's Tuesday stocks, there's some stocks happen in every every month. On a Wednesday, they show up. We have we don't know the particular stocks, but we see the seasonality. We see evidence that there's some repetition. And I think it makes sense because journalists not always have something new to say. So but news sells, so they have to write something. So if a journalist knows, let's say someone in IBM, they have a good idea of what's happening there.
And every third Wednesday of the month, they just write something about it. Okay, that could happen. Even today's Wall Street Journal, I saw something that reminds me of something like that. There's an article about about Robinhood and how people are trading. Stocks had nothing to do with anything that happened. I just felt, okay, this is like a place for something that to fill in. So that happens. So we need to understand these type of seasonalities and correct for biases. There's a geographical biases. So if a firm is mentioned by, let's say, a Microsoft mentioned by the Seattle Pi versus mentioned by New York Times. Okay. So if you're mentioning normally when you mention closer to the headquarter of the firm is mentioned close to the information source, then you see a more it's more positive sentiment. So you want something a little bit more more partial, impartial. You have to go, let's say, distant. You go to New York Times covering Microsoft. So that's just type of the things we've seen. We've also seen the length of the article matters. Lengthy articles are just less informative. Why? Because spend half the article explaining one on the one hand and then the other half on the other hand. And then if the day it converges to neutrality and you don't know what's going on, okay, we see it in the data, but short articles are to the point.
Okay? That's why we have these breakout sessions here, right? 20 minutes, half an hour. Boom. It's not like an hour and a half presentation. States, which gets this. Okay. So that's a little bit about biases. So we try to correct for them. Again, I'm not going to talk too much on this. We try to correct them. Let me give you an idea now of how we measure these narratives. So we have all these different sources and we have different. We take all the articles and we kind of classify them into different reservoirs. So I said, we have corporate reservoir and we have an FCS reservoir and we have country equity reservoir. We have a general reservoir politic reservoir, and we go in for each one of them. Each one of these reservoirs, we look at a given narrative and I plot it here, a few narratives. So let's say we take a narrative like Covid 19. We look at all the articles in the reservoir and we see the percent of articles that actually talk about Covid 19. Okay. And how do we do that? We have some NLP, We do it. We figure out whether the article talks about Covid. Okay. And as you can see here, let's see. Do I have a. Do I have a pointer or. No? No. I don't think I have a point there. So let me just help you navigate. Look at the top graph and you see. There's a spike in the Covid narrative around 2020, and that's about 75%.
You can see on the vertical axis, that means 75% of all articles that time were about Covid. I think it makes sense. If you recall that period three, four years ago, you remember many articles were talking about that over time, there's a drop in the coverage, in the intensity. You can see that. Another example. Oh, thank you. Here we go. Thank you very much. You see it right here? It's Covid. That's March 2020. Another interesting narrative. Right here is inflation. Inflation is interesting because you can see how there's discussion of inflation even before inflation realized. And then there was the peak about last year. Then it's been going down recently. We update this. There's a little bit of a uptick. There's more discussion of inflation. But what's interesting about it, even before. Before the spirit. Look at how this variation of inflation. I mean, if you remember since 2008, everyone's been talking about where's inflation, right? People talk about quantitative easing. They thought that was going to it's going to cause inflation and security prices. Boom. We didn't really see inflation so much in CPI, but we do see some changes in the amount of discussion. There is dynamics, there is variation. In fact, at the end of almost every year we've seen some increased discussion of inflation, even though it didn't realize. The reason I think it's interesting is because I think the discussion is what causes prices to change. The fact that inflation hadn't realized doesn't mean that people didn't price securities with the expectation of inflation.
You had a whole session on inflation this morning, right? So to me, this suggests that it can capture some changes in investor expectations of inflation, even if it hadn't realized more on this soon. When you talk about pricing the cross of narratives in the last one armed conflict, you could see the increase here in end of February 2022, and since then there's less discussion. The conflict is still happening, but people are just less interested in it. At least that's what the data is saying. Let's open up a little bit the discussion on Covid 19. You can see here we looked at the different reservoirs. I just want to highlight this because I think it just shows you the depth of the data because you can look at discussion Covid 19 in each of the reservoirs. So what I did here, I'm comparing F X to corporate, to general, to politics. So you see the first ones to discuss Covid 19 are actually the F x. I think it makes sense if traders, they need to be on top of everything all the time. If something happens in the world, you're going to see it in media. Indeed, that's the case. And by the way, for our central bank indicators, we do. We used indeed the reservoir. It seems to be leading the others. So first there was discussion of Covid 19 in the reservoir and then in corporate. Then it showed up in general and only lastly in in politics.
And look what happened after that. As you remember, in the US, it became a very divisive topic, right? You're going to wear a mask, not wear a mask. Yes. Vaccine. No vaccine. Right. So there's a lot of discussion. If you continue this graph on moving on, you're going to see that during the presidential election, during the the presidential debates, again, Covid 19 went up for in the in the political reservoir. So that's just to show you again how narratives look like. We're able to quantify them and. I'm going to give you just a few more examples just to illustrate that they might be important. So here I took an example from from State Street Insight. So this is something we call the narrative map. And what we try to do here is to highlight how which narratives are important in any given point of time. This is not cross sectional pricing yet. I'm going to get there, but this is just showing you that narratives are important. So what do we do here? We take a narrative that we quantify. Take. Take, for example, manufacturing. So we take manufacturing. We look every week. The percent of articles talk about manufacturing, and we look at the changes in that we take for this. This part here, this whole chart is about the S&P 500. So we take the S&P 500 weekly returns, run a regression on the changes in each of these narratives.
Okay. And we record our squared Hi r squared means there's a significant contemporaneous correlation between the movement of the S&P or the spy and the narrative. Okay. So that's the R squared measure. The R squared measure is the is the horizontal axis. Okay. It's high R squared and low square. High R square means the narrative is important and low r squared means the narrative is unimportant. There's also a Y variable here. The y variable is just the amount of discussion of the narrative. Okay, narratives are highly discussed versus that are less less discussed in the media. Everything is normalized. Cross-sectionally and and time series. Et cetera. So it's just to show you that it's a way to put all the narratives on the same map. So narratives that show up on the top here are heavily discussed. These on the bottom are less discussed. So narratives in this quadrant here are narratives that both have high discussion and the R-squared of a regression of weekly spy on the narrative is high. So we call this important hype. So these are narratives that are important and they're also discussed versus important silence versus unimportant silence and important hype. So what I want to show here is this is as of August 28th, so this is about a couple of weeks ago. And you can see that US growth, manufacturing are important. And they they they seem to be heavily discussed. When you look at the yield curve, recession, housing, they're also important, but they are less discussed.
Okay. All right.
Let's look a little bit about on the Time series. We can take each narrative and plot the R Square and the amount of discussion. So, for example, the recession narrative, you can see that the solid line is the amount of discussion, discussion of recession over the last year has gone down, but the R squared very recently has gone up. So it seems to be correlated with spy returns much more than before. That's why it came up on our monitors. Same thing for inflation. You can see the discussion went up, kind of went down recently. R squared has gone up suggesting that it recently it's becoming more important. Manufacturing discussion is flat. But over the last several months you see that the r squared is increasing. And finally I put here shortages. You can see low discussion, significant variation in the importance for the markets. You can do this the same map for DXY for any kind of asset this is just an example for S&P. Okay. So what I try to do so far is just to give you some color on the narrative that we're going to be using for this study. Okay. So I showed you we talked about the data. We talked about how we quantify narratives. What I want to do now is to move to the cross section of firms. So I grew up when I grew up in academia. People were talking about, I guess around 2000. Late in the 90s, people were talking about how the beta might price securities, whether it's beta or characteristic.
So it used to be this literature about is the firm earning high return because they are value firm or because they have a high beta. You guys remember something like that. Beta and beta. Beta is high minus low. It's a factor of fama-french factors. I think at this point everyone heard has heard a little bit about that. So the entire literature of cross-sectional asset pricing is trying to understand why some firms have high expected returns relative to other firms. Is it because they're small versus large? Is it because their value versus growth is because their quality versus low quality? Or is it the profitability? There's a profitability factor. Is it momentum firms? The gain momentum have high expected return? Et cetera. So people were just trying to understand cross-sectional pricing. What we're going to show you here is we can explain cross sectional pricing of assets, stock returns, the cross section using their sensitivity to narratives. We're going to show you that narrative betas can explain cross sectional differences in average stock returns. To do that, we're going to construct narrative mimicking portfolios. We're going to move from the narrative, which is this time series. We're going to move to a tradable security, a long, short portfolio. We're going to take the high beta stocks versus the low beta stocks. This minus that long short portfolio for each narrative.
And then I'm going to show you that depending on the discussion, the past narrative intensity can predict the narrative portfolio returns. So narratives that have been recently discussed, they're mimicking portfolios are going to outperform in the next six months. That's what I'm going to show you and I'm going to show you the relation of that in different time Windows. I'm going to show the relation of that to just regular price momentum and we control for all that. You can still see the narrative pricing is strong. So this is the first but also the main result. Some of the things I'm going to do later is just looking at robustness. What do we do here? We look at 347 narratives. I'll get to soon. I'll get I'll explain how we get these narratives. But overall, it's through the all our work with statehood and last almost decade. Analysts ask us or strategists, what do you what can you tell us about Brexit? What can you tell us about shortage? What can you tell us about armed conflict? We code these narratives and look at the Time series. So we have 347 of them. I'm going to show you a little bit more and exactly what these narratives are. But think about it. There's many, many narratives and we look at the aggregate reservoir. So these are all the media coverage that we have. We aggregate across all our reservoirs. Sample period is the last ten years.
This is something a little bit, I guess, more of a weakness in the paper. We don't have 30 years of data here. It's going to be ten years. The flip side of it. If you do find significance, then maybe it's really there. We're going to look at a universe of 3000, the the Russell 3000, just the US stocks, almost 3000 for each narrative. We're going to build a 25 long 25 short portfolio using the 52 week narrative beta. What does that mean? Every month I'm going to look at the last 52 weeks, run a regression of each stock return on the changes in the narrative controlling for the S&P record, the beta those that have a high beta at the most the highest 25. I'm going to put them in the log, the lowest 25. I'm going to put them in the short and create this long, short portfolio rebalance every single month with the last 52 week rolling betas. It's a long, short portfolio rebalanced monthly. What's going to be the test? I have 347 narratives, so that means I have 347 portfolios. Rebalance monthly. Every month I'm going to look back and I'm going to calculate the intensity changes of each of these narratives. I have 347 narrative portfolios, but I also have the underlying narrative intensity. I can check let's say the last six months was the overall discussion of armed conflict. Was it higher than the previous six months? Okay.
So for example, this portfolio here that I that I highlighted. It's A66 strategy. What does that mean? I look at the month intensity changes. Six months? What does that mean? I look at the six months relative to the previous six months, the changes. And I hold it for K months. I hold it for six months. What does that mean? I'm going to do what we call the Jagadish Titman ladder portfolios, so I'm really going to have six different portfolios, those that I sorted last month and those that are sorted the month before and the month before the month before. So it's really going to be kind of six portfolios. The way to think about it is I'm really running portfolios here that I'm holding the stocks, I'm holding the narratives for six months. That's how you need to think about it. Okay, Standard ladder portfolios, there's nothing new here. The result. Is that if you run this six six strategy, you get an annualized return of almost 6%. T statistic of two. If you run it over relative to the six factors of fama-french, you get an alpha of almost seven t stat of 253. Again, let me repeat 347 narratives. So we have 347 portfolios. We're done with the 3000 stocks. We transformed everything to the space of narrative portfolios. Now we have 300, almost 350 portfolios. We're going to price the cross section of these portfolios.
So we have each of these portfolio has an intensity of six months relative to the previous six months. We sort all of the narrative portfolios based on the narrative intensity. We pick the top ten narratives, those that increased in discussion the last six months versus those that have decreased in discussion last six months. And we do long short of that. So we're doing like a long short of narrative portfolios that each one of them is a long, short, by the way, that gives you an exposure. It's not like one minus one. It's like 200%, it's 400%. So all these numbers are already divided again by two. So you need to think about it as really like 100 long, 100 short. So that suggests that these portfolios that we sorted based on based on increased discussion the last six months, they end up outperforming in the next six months. When you look at very short term horizon, it's either insignificant or negative. When you start looking at longer horizons, you see this outperformance. Nine months, you see very similar results. The strongest results are on this six month period, which is similar to what Jagadeesh and Titman found when they just did straight momentum. So that begs the question of whether this is really momentum. When I'm picking the narrative portfolios that have that, their narrative has exhibited the most, the increase in narrative discussion. Am I just picking momentum stocks? Okay.
So the first robustness check that we're going to show you after this is going to control for momentum. So this is just the six six strategy. These are the monthly returns you can see stronger in the second half, actually more recently. This is an event time study. You can see that if you hold these narrative portfolios from zero, you hold them for 12 months. You can see in the next six months there's outperformance and then there's flat. It doesn't really come down. The fact that it doesn't come down is interesting because maybe what's happening is I'm going to talk about it soon. There's investors underreact to information, but it's not just an under-reaction and then there's overreaction. It's just the underreact information. And the information is valuable. It's true information about fundamentals. Otherwise, you would see a reversal. So they just react to a narrative discussion late. Okay. What does that mean? It means let's say people have been talking about AI for six months now, more than the previous six months. If you build an AI basket based on exposure to AI narrative, that basket is expected to outperform in the next six months because people then get it. There's discussion, but they don't act on it. Maybe they don't know the discussion. Maybe they don't know relatively speaking, to the 350 narratives, which ones are more discussed versus less discussed? Okay. What if we control for price momentum? When we control for momentum.
You see, the results are still there. In fact, they become even stronger for the nine months portfolios. Okay, so if you control for narrative that have gone up in return in price over the last six months. So the way we do that is we first divide all the 347 narrative portfolios into those that have gone up over the last six months versus gone down the last six months. So we sort them into 30, 40 and 30%. And within each group we do, again, the high narrative versus the low narrative. And you can see if you control if you look at the spread in the narrative portfolio spread in each of these momentum groups and you average them, you still get these results. So that suggests that it's not just the narratives that have gone up in price, continue to go up. That's not what's going on. Is there a relation to media coverage level? What do I mean by that? So far what we've done is we looked at narrative betas. Now let me take a step back and explain why this is important. A narrative beta can be calculated for each firm regardless of the coverage of their firm in the media. The only thing you need to calculate an exposure of a firm to a narrative is just the weekly returns. If you have a weekly returns of a stock, you run a regression on the narrative and you can get the beta right.
You don't really need to see whether the stock was mentioned in the media. That's the beauty of this. We can extract the exposure just by looking at the aggregate. Discussion on the narrative. Not that it's discussed particular in particular to that firm with respect to that firm. So what we're going to do here is we're going to ask whether the discussion of a firm. In the context of this particular narrative. Is that important or is just the beta? Okay, So that makes sense to give you some more perspective. Each firm in the S&P 500. I mean, when you look at all firms, the S&P 500, less than 10% are covered in the media every day. It's not that you see all this coverage. If you're just relying on actual coverage to calculate the exposure of a firm to a narrative by just looking at articles that co mentioned the narrative and the firm, you're not going to find a lot. You're not going to find a lot. But with this narrative based approach, you can actually have a score for each firm. So what we do in this test, in this test is we actually look at the Co mentioning of a firm and a narrative each of the 347 narratives. We look at all the 3000 firms and check every month which of them are mentioned in the media in the context of each narrative. These are firms that have no media coverage with respect to the narrative.
And these are firms that have low media coverage and high media coverage. And we build the narrative beta portfolios within each of these groups. In other words, the results here are the six six strategy that I've shown you, but only including stocks that are not covered at all in the media. You still get a positive return? Not that significant. Okay. Again, our argue for ten year period. It's not bad. You still get a return. When you look at the firms that are highly mentioned in the media with respect to the narratives. You see a far stronger performance. So it does seem to help if you mention them in the media. But the point is we can even calculate exposure of a firm, even if it's not mentioned in the media for the particular asset, for the particular narrative. We also ran portfolios where instead of looking in narrative betas, we do a long, short portfolio of just firms that are mentioned in the context of the narrative. And when you do that, you don't really see a lot of results here. So this really suggests that the way we calculate betas is important. Okay. Is there a sector effect? In other words, when we chose when we choose these firms that have high beta versus low beta to a narrative, are we just picking. All the firms in the same sector is.
That's what we're doing. And the answer is no. So how do we do this? Each firm return that is used to calculate the narrative beta portfolios. And then we have a portfolio of portfolios, right? Each of stock return, we decompose into two parts the return of the sector of the stock and the return excess of the sector of the stock. And then we just calculate all the numbers. But replacing instead of the stock return, we replace either the excess return versus the sector and the sector. And when you do that, you kind of decompose the sector effect. And you can see when you look at the long short narrative portfolio, the excess sector shows up stronger, not maybe as significant as the average. The average sector in terms of magnitude is smaller, significant is higher. So what you can say here is that it's certainly not just driven by a sector effect. In other words, if you take the same sector. Some firms are going to be exposed to a narrative and some are not going to be exposed to a narrative. So you do have significant differences even within sector. That's what it means. Okay. Let's talk a little bit about the narratives. What I want to make sure is that there wasn't any selection bias in the narratives that I chose. I have 350 narratives, but you can ask, Well, how do he got up? How do you come up with these narratives? One criticism could be, Oh, you thought about these narratives because State Street strategists told you you should look at them.
So maybe statewide strategies because they know what's going on all the time. They're telling you to look at narratives that are important and that would end up driving security prices. So maybe there's some selection going on here and we start tracking these narratives when they're important. So maybe there's something going on and we didn't really realize, but actually there's some kind of mechanical or selection issue in the way we select narratives. So what I want to do here is explain to you how we pick these narratives. So we actually have in our data sets, we have plenty of narratives. I mean, we have more than a thousand different narratives, but each of the narratives that I discussed here, they actually come from. One of 14 narrative groups. So each narrative we tag into, we tag to a group, either a commodity related commodity, corporate fundamentals, cryptos, ESG, etcetera. And each of these 347 narratives, some of them could also have sub narratives. So, for example, armed conflict, which would be somewhere in society. Can have a sub narrative. Iran, Ukraine Covid 19 could have a sub narrative supply-chain US economy. Fed could have inflation. Quantitative easing list goes on and on. So you can have on the one hand narrative groups and you can have sub narratives. In addition, there are 53 what we call evergreen narratives.
So these are narratives that were not defined by us. We didn't choose them and neither as State Street strategists. These are narratives that are based on Jel classification, Jel Journal of Economic Literature, and they have a classification they've been looking at for years, for decades, of how to classify different articles in economics. So that's a predefined set of narratives, and all of them are on our system. The summary is when you run this type of strategy for each group, either this group or that group, this group or this group, you always get the same result. The strategy is significant. Okay. So it gives you some more comfort that there is no really selection bias in the type of narratives we look at. That's the point. So if you look at the actual numbers, if you try to if you try to decompose each of the 347 narratives into the return of the narrative group versus the specific return of the narrative, if you do that exercise, you get that both of them are important. It's not really coming from one thing. If you take the 682 sub narratives, you can see you run the same strategy compared to the baseline similar results. There's a bunch of robustness tests. I'm not going to go over all of them. You take out, you know, stocks for this study. We only include stocks with more than $1 price. If you take that restriction off, you still get the results.
If you only look at the subset of S&P 500 firms, you still get the results. If you only value if you evaluate the strategies, you get the same result. If you look at the Evergreen Strategies top ten versus top five narratives in the portfolio, you still get. Alphas. There are statistically significant when you compare the first half versus the second half. You see that most of the effect is really coming from the second half. So if there's something to say here is that the the narrative pricing that we find is even becoming stronger and stronger over time. So what does this mean? Again, the result is the basic result here. The main result is narratives. That have exhibited a significant change in discussion. A positive change in discussion over the last six months. The narrative portfolio is associated with these narratives end up outperforming the next six months. That's the basic result. Now, what's happening here? So we started investigating. Is it because of earnings announcements? Is all the information coming around? Earnings announcement turns out take earnings announcement month. Out of the analysis, you get the same result. So it's not earnings announcements per se. Are the narratives, the narrative betas and the narrative intensity. Is that capturing something about earnings? Maybe what's going on is that investors are underreacting to information about future cash flows. So maybe what's going on, the fact that people are now talking more about AI, maybe what's what's going to happen is that earnings of AI exposed firms are going to go up, but people are just underreacting to it is that's what's happening.
Indeed, that's what we find. If you look at the earnings changes from date zero, that's portfolio construction. And this is an event time. You look at the firms in the portfolio, look at their earnings change over the next eight quarters. You find that there is an increase in earnings over the next the peaks over the next 18 months. So to me, this suggests that people are underreacting to actual earnings information. There's a new narrative happening. People are talking about it. The prices don't reflect it yet, but over on average, in a year and a half, it's going to start showing up in earnings. Do analysts understand that? We don't find evidence that they do. We look at earnings. We look at earnings revisions, the revisions of earnings expectations over the time. When you see that the intensity of the narrative increases. We did a few controls and you see that if you double sort the narratives by earnings changes versus the intensity of the narrative, you find that in each revision group you still have the same the same result. So it's not the case that analysts understand that. So our interpretation is that investors seem to underreact to long run information on cash flows that are reflected in the narrative.
So again, the narrative is being discussed. It goes up. We think that means cash flows are going to go up on average. Which firms are those that are going to be exposed? We use the data to identify that. Okay. So let me conclude. What did we try to do here? We talked about narratives, how we quantify narratives, and then we showed an approach of how to calculate exposures to narrative. This is the idea is with the betas. So that helps us, I think, define a whole new set. Of factors of portfolios that are mimicking the different narratives that are discussed in the media. And what I try to show you here, one application of it is that these portfolios of narratives can explain cross sectional differences in stock returns. The stocks that are exposed to narratives that are in discussion that exhibit a positive increase in discussion end up outperforming the next six months, significantly outperform. And the economic interpretation. Again, this is work in progress. And again, we're very excited to hear your comments. What we think is going on is people just underreact to the information about future cash flow. There could be some more information about maybe discount rate, but at least part of it seems to be not understanding yet that AI is a big deal and it's going to show up in cash flows. That's what we're that's what the results suggest. Thank you very much.
Speaker2: All righty. Thank you. I think we do have time for a couple of questions. So you're able to submit your questions on Slido or we can ask in the crowd. So it looks like we have a couple back there.
Speaker3: Hi. This is very informative. Actually, I have two questions. One is probably just a quick one. How did you account for Meme stock? Was that an outlier and you took them out of your subset?
Ronnie Sadka: Why should we take them out?
Speaker3: Because they were driving. I know. I'm just curious about the meme.
Ronnie Sadka: Stocks are going to show up as examples of of a narrative that went up very quickly. Those are going to show up. If you see index, I don't know. I can't go back in the slides, but in that table, the short run narrative, those that if you compare an increased discussion over a week versus the previous week, the narrative, the meme stocks are going to be there. And indeed, on average we find the results that the returns are negative. So there does seem to be some overreaction to a very short term increases in narrative. But over longer periods of three months, six months, nine months, there seems to be underreaction. So I think actually like the meme stock example a lot because I think that's a good example of an overreaction. It's very similar. If you think about stock returns, there's a significant reversal in the short term price go up, go down, go up, go down. And in the long run there's more momentum, very similar results we find in narratives, although we control for the regular reversal of momentum in stocks returns. And we don't when we find that that doesn't explain these effects. Is that helpful? Yeah.
Speaker3: The question the second one is just say the stock is for high narrative. How do you control for the next six months where it may be exposed to another narrative? Is that something that you control for?
Ronnie Sadka: Yes, it's a very good question. We we run in a sense, we run for every firm. We run for every firm we run all the betas. We run for each firm, we run all the exposures for each of the narratives. We don't do it in a multivariate way because we don't have enough data, right? So you just look at 52 weeks. So we just run two factors the the market and the change the narrative, but we do it for every narrative and every stock. You could have a situation that when you're doing a long, short for a given narrative, the firm can show up there and it can show up in another narrative portfolio. That could be the case. We're not excluding that. Okay. You can have the opposite. You can have in the same portfolio. The firm is going to cancel each other itself because it might be positive related to one and negative relate to another. Yes.
Speaker2: I think we have another question over there.
Speaker4: Thank you, Ronnie. I have a question. You mentioned the narrative data are very important. I wonder whether the narrative data have high autocorrelation. And also the second question is that I'm interested in the mimicking the narrative mimicking portfolio when you form your mimicking portfolio. Have you looked at each of the mimicking portfolio and make story make sense of it? Does it make sense to you as a quant?
Ronnie Sadka: Yes. Well, as a quant, I'm not sure whether it makes sense as a good question, but. But the answer is yes. We looked at. So let's two things you asked. First is the autocorrelation of the beta. The betas are as stable as a 52 week period window allows them to be in the sense they're going to be stable from month to month, but after six months or their half, life is probably going to be short. So it's going to be within after a year, you might have a totally different beta for the firm. How to think about it. I'm not sure. It could be that at some point of time, think about Covid as a narrative. In any given point of time, the last two years, a firm could have been exposed to, you know, sitting at home, people sitting at home and want to buy. I don't know. They want to buy Nike because they want to go run outside. And that firm could have a completely different exposure because of shortages in the next in the next cycle. So it's not clear to me that that's a wrong thing. Okay. But the result is indeed that the Beatles are not that stable. They're as stable as the window that we that we measure them.
Ronnie Sadka: So that's a good point. Second point, whether they make sense. We actually spend a lot of time on that and it's in the paper, but I didn't have time to present it. What did we do? We did things like trying to look at in any given point, which is the narrative that is the most discussed. And then look at the portfolio return of that. Does it really correlate? We try to see which narratives and narrative portfolios explain, let's say the fama-french factors. So you can see things like value investing is going to be related to the HMO factor. I think it's kind of interesting. So when you have so many of these narratives, again as a quant, it's hard to make sense of every single result. But some of them really pass the smell test. So when you look at armed conflict, you see that I think with some of the stuff that Will showed also did a good job of showing which currencies are exposed to which narrative. So there's always going to be some errors, right, and estimation issues. But I do think it's capturing something real. Thank you for that.
Speaker2: Great. I think we had one more question in the back.
Speaker5: Hi. Thanks very much for your talk. So I was wondering two questions also, if I might. And so if one were to regress or use whatever fitting method you use to try to explain returns using your set of narratives or even sub narratives, then how what is the residuals look like? How much of the returns or performance is not explained or explainable by by the by the narratives? And my second question is that's an awful lot of especially at the sub narrative level. That's a lot of them. I think I saw a number in the six hundreds. And so if hypothetically, I'm guessing some of them are correlated. So if you were to make orthogonal. Vectors. How many dimensions would you in narrative space have?
Ronnie Sadka: Thank you. Very good. Very good questions. Okay. The first question about how much can we actually explain of the cross section? I would say probably a very low number. Now, if you just use existing models to explain the cross section of future returns, anything above 1% is fantastic, right? I mean, that's the reality. I mean, it's we don't really that's one of the puzzles. We don't really when you look at individual firm level, the R squares are going to be very low. When you start looking at portfolios, there's going to be higher. So I don't have a precise answer, although we did do fama-macbeth regressions at the narrative portfolio level. So I can calculate that and I'm sure we have. I just don't remember the average R squared, but my sense is it would be low. But even if you control for momentum and other factors, it does seem to be the most significant one in terms of t stat. Now your other question is about how many different narratives do we really have? So it's a good question. I don't have a good answer to that. Maybe, Gideon remembers, but I don't remember that we've done a principal component and try to understand how many. It's a very good question. How many principal components really explain the most overall, the variation of all the narratives? I would suspect that at least the 14 tags that we have are might not might be less correlated than just the individual ones. But I haven't I haven't just don't recall if we look at the average correlation between the narratives, but I don't know the answer to that. But it's a very good it's a very good question. It's a very good question.
Speaker2: Thank you. Wonderful. I think that's all we have time for. So thank you so much. That was very good.
Ronnie Sadka: Thank you very much, guys.
State Street LIVE: Research Retreat offers a wide range of academic expertise and timely market insights.
Just like social media fuels consumer trends, media narratives are influencing financial markets and investor behavior.
According to Ronnie Sadka, Haub family professor of finance, senior associate dean for faculty, finance department, chairperson at Boston College's Carroll School of Management, and academic partner at State Street Associates, media-derived narratives are driving markets over the long run, and serve as a good tool for predicting market returns beyond traditional indicators.
Trending narratives outperform
Using data sourced from State Street’s MediaStats Series – an artificial intelligence (AI) platform that transforms millions of unstructured media data points into actionable insights for institutional investors – Sadka quantified the narratives against portfolios he constructed based on recent popular narratives. In all, 150,000 digital media sources and roughly 350 narratives from MediaStats – a collaboration between Cambridge, Massachusetts-based MKT MediaStats and State Street Global Markets – were used in the analysis.
Sadka’s findings showed that portfolios containing assets that reflect trending narratives outperform those with themes of descending media coverage. The results suggest that investors respond better to short-run narratives than to long-run narrative trends.
Enhanced asset allocation
Media-derived narratives, whether true or not, help explain market-wide moves and may be used to enhance asset allocation strategies. As Sadka explained, narrative indicators can help measure exposures and stock returns in a meaningful way. “Stocks exposed to narratives with recent increased intensity exhibit positive earnings growth for up to a year,” he said.
As Sadka concluded, popular thinking and trends ultimately drive an individual’s decisions about how and where to invest. For this reason, it is increasingly important for investors to track economic media narratives and understand how they influence financial markets and how they can impact asset prices over time.
“Investors and analysts are underreacting to valuable information about future cash flow that are reflected in the narratives. They often don’t act or react to narrative discussions too late,” he said. “Artificial intelligence, which is currently a big deal, I would say will show up in future cash flows.”