The fox knows many things, but the hedgehog knows one big thing. Well, today's presentation from us is going to be a bit of a foxy hedgehog. We have one big idea, which is studying investor behaviour, but we're going to take a number of different approaches to that. So first we're going to review a bit some ideas that were first introduced at this conference last year, inferred information and the trade typology. Now, the first idea is the trade typology. So any trade that you observe can be classified as either a buy or a sell, and every trade has to come from either an overweight or an underweight position. So we utilise this idea to sort trades out into individual quadrants, and then we look at aggregates across these quadrants to identify new ways of looking at patterns of trading. In particular, we sort things into what we call conviction trades, and these are trades where you're building on positions, if you're buying into existing overweights, selling out of existing underweights further, versus what we call unwinding trades. Unwinding trades are cases where you're exiting an existing position, so you're selling out of an overweight or buying into an underweight. Now, in the future we plan on doing this matching at the position level and building out separate indicators for each of these quadrants and for these various aggregations. Today, however, we're going to show you how even utilising the indicators that you currently have in your hand, you can apply these ideas to aggregated flows and look at aggregated conviction trades and aggregated unwinding trades. Now, one quick note on how this might work in the context of what we call active flows. So within the equities space, and these are particularly referring to global equities sectors, we have a concept called active flow, and this captures manager rebalances within portfolios. Now, these active flows net to zero within any given portfolio, and instantly net to zero across all portfolios. When we apply this conviction unwinding decomposition to our active flows, it's necessarily the case that conviction plus unwinding flow is also going to be equal to zero, which means that conviction and unwinding flows become mirror images of one another. The same logic applies to any partition of this set of flows, for instance overweights and underweights. When we look at FX examples today, we are not going to be using active flows, we are going to be using total flows, so the same sort of mirror image would not apply there. So the first of several applications. Here the idea is to try to time the equity value minus growth factor using in this case unwinding flows. We're going to be looking at the net unwinding flow. The idea is that when net unwinding flow is relatively high, that means that investors are buying back their underweight positions. We think that this may correspond to buying back beaten down value stocks, and hence may be a positive signal for the value factor on the whole. So what we will do is we're going to go long value when winding flows are relatively high, and short where it's relatively low. So these are the results obtained by implementing this idea. Now, we've implemented this in two different ways. On the one hand we've done a long-short strategy, so either go long value, short growth or vice versa and applying that idea does seem to work. Timing that long-short strategy using that unwinding flow does seem to work on average through time. We've also looked at this in what we call a time strategy where we're switching between going along the value factor versus just holding the AQUI market portfolio, and again, that time strategy also outperforms just passively holding the AQUI and, of course, outperforms the value minus growth factor, which itself has not done terribly well over the past couple of decades. So again, this is a very crude instrument, these aggregated unwinding flows, but they still seem to offer some value in the context of timing and factor. So now we're going to apply the conviction flow idea, and here we're going to try to time the market itself. Here the idea is that if investors are in a net weight building on their overweight positions, this may indicate a generally bullish view in the cross section of sectors in this case, which we expect may propagate to the market itself. It may ultimately indicate relative market outperformance at the overall equity level. So the idea is that we are going to go along the market whenever we see conviction flows are relatively high, short otherwise. Again, we implement this two ways; one is a long-short strategy, either go long or short the AQUI based upon net conviction flow. Conversely we also implement a timing strategy. Here we're swapping between the AQUI and just holding cash. In either case we do seem to see that there's a very simple idea, does hold some promise, and on average we do earn positive relative returns through time. So that's all well and good for equities, but let's try and apply these ideas into a different asset glass. G10 FX, in particular the timing of a G10 momentum factor. Now, here we're using all the crosses against the dollar, so we're not using the dollar flows and holdings themselves, but the other G9 crosses against. In this case, the idea is if we see investors buying into conviction positions, so if they are buying into existing overweights let's say, the yens, the euros, Swiss, etc., we think that that may be conducive to the further outperformance of the momentum factor. So when net conviction flows across G10 currencies are relatively high, we're going to say go long momentum; when they are low we're going to say, go short momentum. These are the results that we obtained; again, applying this very simple idea to timing a G10 momentum factor does seem to work on average, so there's some promise in this approach. Now we're going to take a slightly different view, do the sort of equivalent of timing the market in FX. Here we're going to say, let's time DXY. Now, here we're actually going to be looking at conviction selling, so where conviction selling of the other G10 crosses is relatively high, where investors are selling in to say their euro underweights and are doing that across several currencies, then we're going to bet that that may be a positive for the dollar, which is likely at the other end of many of those trades. So when conviction selling or minus one times conviction net flow is relatively high, then we're going to go long the DXY, otherwise short the DXY. Again, we see that this simple idea does seem to be able to time the dollar factor, if you will, in aggregate through time over the last couple of decades. All right. So now we're going to switch gears; this is the first foxy turn, and we're going to switch to the idea of risk, behavioural risk, and Ken alluded to this a bit in his talk as well. Now, what we did last year is we borrowed an old idea from grinold and kahn called the Marginal Contribution to Risk. Now, what this is, is this idea tells you for a given position in your portfolio whether it be a stock, a sector, a currency, whatever, how much of that particular asset contributes to the overall risk in your portfolio. So we're going to apply that idea to our behavioural portfolio, to our aggregate real money positioning that we observe. As a holdings measure we also have our own idea, which is a little bit new, that we're calling the marginal contribution to incremental risk. The concept is very similar, except rather than multiplying through a holdings vector through a covariance matrix of returns, we're going to multiply it to a flow vector through the covariance matrix of returns and tell you how risk is being added to or subtracted from the aggregate portfolio by recent amounts of trading in various assets. So you can think of the marginal contribution to risk as being risk positioning, and the marginal contribution to incremental risk as being risk flow. By the way, risk flow and risk positioning can also be classified within the same unwinding conviction typology that we've applied to the direct observation of flow. So just to remind you of how these calculations work. If you want to compute the marginal contribution to active or incremental risk, you start with your vector of let's say flows or holdings. So if we're thinking about equity sectors, there would be 11 numbers in here. If it's G10 currencies, given that we're excluding the dollar here, we would have nine items in that vector. You're essentially filtering through the positions or flows through a covariance matrix. So you take your vector positions or flows, you whack it against the covariance matrix, particularly the first entry would be whacked against the first column of the covariance matrix. So if we were talking about energy, that would be the covariance of energy with all the other sectors. You multiply through, you add everything up, you scale this result by the overall portfolio risk at a point in time, and boom, you have your marginal contribution to risk, whether it's active or incremental. Then you rinse and repeat for the other 11 sectors, nine currencies, thousand stocks, whatever you want, and then you get your overall vectors of marginal contributions to risk. So let's see how these things look at a point in time. Now, these are recent data from November 9, and on the left we have a scatter of flows versus holdings, so that's just straight behavioural data. You can see there's sort of a scattering; some things are being bought, some things are being sold, some things are overweight, some things are underweight. On the right, however, we look at applying this same classification to risk flows and risk holdings. Broadly speaking the distribution looks similar, but everything has sort of been shifted over to the right. In particular, some sectors like consumer discretionary, where Amazon lives, have migrated a great deal along the risk spectrum in terms of positioning. So looking at things through a risk lens tells us slightly different things than simply observing the pattern of flow and positioning directly. We can draw different inferences about where investors are adding to risk, which is not the same at where they're just adding to their positions. Next we do this same thing in the context of G10 FX. On the left again we just have flows and holdings, and on the right we're looking at risk flows and risk holdings. Again, we see a bit of a change. When we go from left to right, we move from a reasonably even scatter with a positive correlation between flows and holdings to everyone living in a conviction-buying quadrant looking through the risk view. On the chart on the right, everyone is in the top right quadrant, everything is conviction risk buying. There's also a much tighter correlation between risk flow and risk holdings, there's a 90 per cent correlation, than there was between just flows and holdings. We also see some things cluster and other things separate. The yen, appropriately, is very lonely and very far away from everyone else as a relative safe haven currency. So that concludes this section of the talk. Now, continuing our next foxy turn I'm going to pass things on to Scott and Travis to talk about a new idea to combine equity sentiment and equity investor behaviour.
Great, thank you Alex. Like you mentioned, in this section of the presentation we'll be discussing a product that we're hoping to watch next year called the Country Equity Scorecard in Developed Markets with the idea of capturing sentiment for country indices in the developed market across a suite of indicators that we have here at State Street Associates, including investor behaviour and media stats. So diving into this a bit deeper, a little more detail here. At a high level, if you're familiar with our behaviour risk scorecard, otherwise known as the BRS, where we aggregate risk appetite across many different indicators and asset classes to generate a holistic view of market risk, we do something in a similar vein for countries, where we aggregate a sentiment score across different indicators at the country level. We're hoping to quantify and measure country sentiment in order to identify risks where there might be underperformance for a given country indices, as well as opportunities for outperformance in the relative sets. So now that we have a high level understanding of what the scorecard is going to be, we can dive into a bit more in terms of the construction as well as robustness of the signals that we generate. Alex really went in-depth on investor behaviour already so we won't spend too much time here, but we're really going to use three core indicators from our country equity series from our investor behaviour, the proprietary custody information that we have. Those are going to be our flows, holdings and borrowings. Flows, as you'd expect, we say go with the flow. These are highly persistent months on months of buying or selling from a homogenous group of investors. So when we see positive flows for a country, we expect positive future performance. Holdings we view as a bit of a trade indicator. The higher-level holdings, the trade may be crowded and you may see an exit out of that trade as well in a meaner version event. As investors exit that trade, we expect future negative returns. Borrowings, which Alex actually spoke about last year is a relatively new indicator that we released at the country level. It measures the amount of short selling for a given country index, and similar to holdings we view this as a crowded trade. So if borrowings is relatively high, we expect those shorts to start to close out, become more expensive, and as they close we expect future positive returns. So before we go any further and talk about the back tests and the information that these signals may possess, it's really helpful to understand how we're normalising them and also constructing the back tests themselves. For investor behaviour indicators, we normalise these in a very standard framework, in a framework that we've used for many years now across our indicators, where we cross to mean our flows, holdings and borrowings to get a general sense of the relativeness of those indicators, and then we use a time series normalisation or Z score for each country to understand if those indicators are high or low relative to themselves. We can then assign these scores to the countries and we sort based on the positive and negative country sentiment there, and we can sort into turnstile buckets. We'll go along the top turnstile to the most positive scores and short at the bottom turnstile, to the most negative scores. We can create an overlapping ladder trade strategy, so overlapping sub-portfolios, and that's how we'll test our robustness in our back tests. It's also worth noting that all of our indicators have a publication lag included, so two days for our media stats, and we get to that later on in this presentation, and three days for investor behaviour. With that, I'll hand it off to Scott to talk about some of our base case back tests with investor behaviour.
Thanks, Travis. First we just want to start with a pretty straightforward example, so we're going to leverage our flows driven strategy in which we avoid crowded trades and seek underweight assets. As you can see, with this straightforward approach it does pretty well. The next example we have we want to improve our risk to reward ratio by including our EBI to identify countries that have crowded short positions, thus we may see positive future performance as positions are closed or vice versa. I do want to note that the threshold, the equities, the borrowings and holdings indicators for Z scores greater than one or negative one to identify exchanges where a mean reversion is expected. As you can see, this performs quite better than their initial straightforward approach with risk-adjusted returns improving quite significantly. Our T statistic are great than two for statistical significance. I do want to note that as investor behaviour does really well in this construct, we believe there's missed signals from our institutional money that we believe digital media can capture. So I'll hand it back to Travis and he can introduce the digital media indicators.
Thank you, Scott. Now we'll discuss, how can we start to capture an additional level of sentiment here that Scott alluded to in the sense of media driven events, and to do that we'll be using MKT media stats, which Ronnie will be discussing later today. He'll go more in depth here. At a high level, State Street media stats leverages advanced natural language processing to scrape the web for hundreds of thousands of digital media sources to generate a multitude of indicators. For our scorecard we're going to be using three core country scores; the sentiment, intensity and disagreement. Sentiment, as it implies, is generally the tone of a digital media and how it's covering a given country index. Intensity is the volume of that country, so how much coverage is a given country getting? We expect that to be positive going forwards. It's a little different than that stock level; at the stock level we expect somewhat of a short-term meaner version there. You can think of GameStop as a really prolific example of that. When the price increased substantially there was high coverage, there was an overvaluation. At the country level, we see this meaner version is drawn out to a much longer time period; 40, 60 trade days. Our back tests are signalling the scorecard is in a 20-day time range. Disagreement we view as a negative indicator of future returns. This is essentially the amount of positive and the amount of negative coverage, hence the disagreement. The higher level of disagreement that we see, we expect more negative future returns. This is due to the constraints on institutions in terms of their short sales. The higher level of disagreement usually means the optimists point of views are incorporated in the prices more, hence there may be a reversion going forward. The next question here is, how do we encapsulate all this coverage in the media to identify extreme events that might be happening, like high levels of disagreement or potentially where there's an event with a high degree of sentiment as well as intensity. To do that, we create what's called a composite media score. Here we normalise all of our media indicators in the cross section, so the comparable across developed market countries, and we add our intensity to the sentiment score and subtract out disagreement. When we look at these extreme thresholds, though, when media score is above one, we see significant positive future returns. When it's below negative one, we see negative future returns. So we have a pretty good spread there in terms of these significant events that we see in our composite media score. So now we use this information to overlay with our investor behaviour, what real money institutions are actually doing, to kind of create a more robust and encapsulating sentiment score for our countries. I'll hand it back to Scott to discuss what those results look like.
Thanks Travis. So as Travis laid out, we can see this relationship between our composite MKT score in future returns is quite positive. So when overlaying that or adding that with our investor behaviour, we do see that it enhances our performance in our work to risk ratio. Our risk just returns and continues to have statistical significance. So this is capturing additional country level insights with our digital media. I do want to note that we've only been focusing on a single holding period; ten day. I do want to note and show here in the next slide that our composite strategy improves all base case strategies across holding periods in USD local denominated indices. I want to highlight that combining digital media coverage with institutional behaviour results is robust across country sentiment signals. You're probably curious about how this might look over time or what our country scorecard might be saying today, so I'll hand it back to Travis to tell us what that looks like.
Great. I just want to give a quick preview of what this might look like on insights. This top left corner we have a heat map that looks at the sentiment scores through time for the countries that we cover. As you can see, we have long buckets and the short buckets, green and red respectively. For those countries in those buckets, so it's the chart on the right, we see that when we're in the long bucket we do see future access positive returns, and the short buckets on average see future negative returns. Currently we're looking at United States is in our short bucket, and that's driven by heavy overweights by institutional investors as well as outflows in what's most likely a profit taking selling. Officially the Central Bank came out with talking about tapering and raising rates, so that's also decreasing the amount of support we're seeing in our monetary policy. We're also long a couple of European countries. The ECB seems to be endlessly supportive, and so that's not surprising there as well. So with that, I'll hand it off to Alex to wrap it up.
All right, thanks Travis. I'll very briefly summarise what we just said. One, if we look at these typological ideas and decompositions we get value, even when we're starting with aggregated flows and positioning. So doing it imprecisely and crudely still works to a degree. Second, there can be stark differences in what we see by looking at the pattern of behaviour as it is directly versus filtering it through a risk filter, or looking at it from the vantage or standpoint of risk positioning and risk flows. Finally, we can successfully combine, yet again, we've done this before and now we're doing it yet again, media information and investor behaviour information to produce even more predictive forecasts for future returns. With that, that's basically the end. We are now going to open the floor to any questions that you may have.
I'll start. Question one. How does the typology produced by matching at the position level differ from what is derived at the aggregate level? One word: netting. There is a whole lot of netting. As there's disagreement amongst individual portfolios, currently if we take an existing indicator and say energy is in the top right quadrant, it's conviction buying, yes, all right, fine. I'm sure it is on net, but there could be a whole lot of action going on behind the scenes. You can think of that one dot. We've placed energy in this top quadrant. Really, it's sort of factored into four little dots, each one of which is comprised of a set of trades where some energy investors are conviction buyers, some are conviction sellers, some are unwinding buyers, some are unwinding sellers. That netting can matter a lot. So what we intend to do is in the future, by releasing all those indicators we'll be able to understand the net comes out in one place, but a net coming from a place of severe disagreement where there's a lot of fighting amongst the different contributors to that aggregate flow could mean something very different to a net coming from a place were generally speaking, yes, the investors are aligned. There's also imprecision in that when we look at the net position where we say energy is seeing conviction buying, all right, but we have a set of flows that are buying and a set of holdings that are underweight. They might not be the same investors. There's a very big difference between having a net underweight and a net buying position versus is it the case that individual agents that are overweight are the ones that are buying? So that's a very big difference, and that's something we'll be able to identify and understand the implications of in the future when we build out these indicators in a more precise way. Second question, I think this is more for Travis and Scott. How persistent are the signals in the country scorecard?
That's a good question. If you're familiar with our investor behaviour back tests, they're quite persistent so the turnover rations can tend to be quite low there. However, something that's interesting is that when we overlay the media stats in our digital media composite score there, we do see that the turnover increases to around 17 per cent or so, but with that said, you're getting a signal boost there given that those signals are usually indicative of an event that's being covered by digital media. So it is persistent, especially driven by investor behaviour, and then media adds in another layer of turnover where it's identifying recent trends and recent events.
I have another one for Travis and Scott. How did the composite strategy behave during the COVID crash in Feb/March 2020?
If you slide in a back test there, it performed quite nicely. That was, here we go.
Maybe this is hard to see, maybe we should, there you go.
We actually had a boost in performance during COVID, and this is again because of the relative trade were long-short within the country equity indices, but we actually get a nice boost given that we're going long in some safe haven countries, and short not so much, some of the countries that might have seen a bigger hit on their equity indices.
Yet another one for Scott and Travis, and maybe a little bit of me. Will you be generating an EM scorecard too, because obviously we're starting with developed markets?
We're hoping to take a very similar framework and apply it to emerging markets. I'll let you add in any more.
Yes, it's ongoing research. So we did initially try to look at everyone together, and there were differences in terms of how EM versus DM behave. So we did decide, as we have in many other contexts, to treat development markets separately to emerging markets. So yes, we do intent to study emerging markets, though the formulation will be largely similar. It may not be identical, because different indicators do sometimes behave a bit differently in these different market segments. I think that is all of our questions. Yes, so that's it.