To understand why this shift in perspective toward dispersed information is truly a qualitative shift, consider the following idea, one of the deepest in financial economics : transaction quantities play two distinct roles – they clear markets and they convey information (Grossman 1976 ; Kyle 1985 ; Glosten and Milgrom, 1985). Only the first of these roles is reflected in the “more buyers than sellers” view of why flows affect prices. The second is the more insightful (and too often missed). It arises in contexts where information is dispersed because transaction flows affect people’s expectations (about future fundamentals and prices). That this second role is empirically relevant to foreign exchange is clear from, for example : (1) findings that transactions have different effects on price, dollar for dollar, depending on the institution type behind them and (2) findings that transaction flows in one currency market have price effects in other currency markets, despite not occurring in those other markets. In addition, not only is transaction flow not “just demand” in the sense above that it plays more than the first role, it is also not just demand in that it reflects transactions. Shifts in demand that move prices need not involve any transactions whatsoever. For example, in traditional macro models of exchange rates, demand shifts move price before any transactions can occur at the old price, and at the new price people are indifferent again between buying and selling. In these models, then, there is no relation whatsoever between transaction flows and price movements.

This presentation summarizes key lessons from the New Micro approach with respect to explaining and forecasting exchange rates. [3] Section 1 provides an executive summary. Section 2 addresses three overarching topics that regularly receive attention in this literature. These include : (1) differences between various transaction flow measures, (2) how to think about causality between transaction flows and returns, and (3) strategies for pinning down underlying flow drivers. Section 3 addresses main findings.

Futures flow data is the second main flow measure. Advantages of the futures data include : they are available publicly (with delays), they are available in long samples (> 10 years), and they are rather highly correlated with currency returns in concurrent data. Shortcomings of futures data include the following. First, futures data are not publicly available at the same level of disaggregation as end-user flows (and, again, the disaggregated data are more powerful). The segment breakdown in futures data is limited to two basic categories, “hedgers” and “speculators”. Second, not only do historical futures data have lower leading correlations with currency returns than end-user flows, we can expect that their public availability will soon eliminate any forecasting power through competition effects, if this has not already occurred. Third, futures market participants are not representative of the full universe of currency market end-users. For example, futures transactions do not include the subset of currency demands requiring immediate delivery.

The third competing data type is that from end-users. Sources include commercial and custodial banks. These data sets have the advantages of covering long sample periods (> 10 years) and many currencies (permitting cross-sectional as well as time-series analysis). Moreover, these datasets can typically be disaggregated. Commercial bank datasets generally have strong coverage of all end-user segments, both financial and non-financial. Which types of end-user flows convey the most information is, of course, an empirical question, one that we address below.

From my own experience with the causality issue, it is best to first open conceptual room for understanding why flows should cause price changes, and then address why alternatives to this flow-to-price causality are often faulty. That is the approach taken here. The theory of trading is quite helpful for opening conceptual room for why flows cause price changes. Models of trading show that flow causes price adjustment because it conveys information. [6] What kind of information could currency flows possibly convey? There are two basic kinds : (1) risk-neutral fundamentals and (2) risk fundamentals. [7] The term “risk-neutral fundamentals” refers to macro variables that would drive exchange rates even in a risk-neutral world (i.e., a world where the risk of taking on currency positions has no effect on prices). These variables are money supplies, GDPs, interest rates, and inflation – the variables at the center of macro exchange rate models.

The category of “risk fundamentals” is more subtle. It includes variables that take effect as one moves from a risk-neutral world to a risk-averse one in which risk affects position-taking and prices. Examples of risk fundamentals include risk preferences, hedging demands, and aggregate liquidity demands, all of which vary over time in ways that are not publicly observable. Even when independent of traditional macro fundamentals, these variables drive trades that can and should affect prices : trades motivated by shifts in risk fundamentals have to be absorbed by willing counterparties, and if these counterparties are risk averse then price changes in their favor is what makes them willing. The resulting price changes persist and can even be permanent (see, e.g., the model in Evans and Lyons 2002a). [8] That flows actually do affect exchange rates via this risk-fundamentals channel is obvious to most practitioners.

In contrast, most people are skeptical about the empirical relevance of flow conveying information about risk-neutral (macro) fundamentals. This skepticism arises in large part because inside information about macro fundamentals in the hands of one or a few traders is rather unlikely, at least in major markets. However, this skeptical view is being revised in the face of recent research (reviewed later) that shifts attention from inside information in the hands of a few to small bits of information dispersed throughout the whole economy. Consider some specific examples, variables like the employment report, the trade balance, and the CPI index. Official announcements of these variables occur much later than the underlying economic activity that they measure. In the meantime, that same underlying activity is affecting currency transactions (e.g., an import generating an FX trade). Importantly, it is not necessary for end-users to view themselves as trading strategically on these bits of macro-relevant information. What is necessary is that their FX trades be correlated with these bits. At higher levels of aggregation, like that in end-user flows, the macro information conveyed by these trades becomes significant. In contrast, the traditional view on how exchange rates impound this dispersed macro information is that the bits are first aggregated by official institutions, then a summary is conveyed to the whole market simultaneously in the form of a macro announcement. But as noted, research shows that public macro announcements account for disappointingly little exchange rate variation in total (despite being important within brief time windows). The modern view – backed by recent results using end-user flows – is that much macro information is in fact aggregated by markets rather than official institutions, with flow playing a central role. Under this modern view, the impounding of macro information in price is a continuous process of dispersed information being “symmetrized” rather than a lumpy process of announcement arrivals.

We move now from opening conceptual room for flow-to-price causality to considering some alternatives to flow-to-price causality. The principal alternative is that price changes cause flows, so-called reverse causality. Surely some reverse causality is present. Nevertheless, often the evidence offered to support reverse causality is misinterpreted. For example, some have argued that when price movements precede flows this implies reverse causality (typically called feedback trading). Not necessarily true. Here is a counter-example. Suppose flows are positively correlated over time (net purchases followed by net purchases, on average, and vice versa) and are not in any way caused by price. (We know that major-bank end-user flows are positively correlated over time ; not being caused by price is simply an assumption to make the example clear.) Now remember that exchange rates are forward-looking asset prices, and so effects from anticipatable flows should be fully discounted immediately. So, a positive shock to dollar buying today will increase the value of the dollar today and will be followed by further positive flow into dollars (i.e., price is forecasting these flows). Though the price increase occurs before the subsequent dollar purchases, there is no positive feedback trading : by construction causality is running wholly from flow to price. Bottom line : because the exchange rate is a forward-looking asset price, one cannot conclude that prices are causing flow just because prices sometimes precede flow.

In addition, reverse causality is often asserted without supporting empirical evidence. For example, some people argue that trading strategies like “the trend is your friend” or “momentum” produce positive feedback that can explain the strong positive correlation found in aggregate data between flows and price. But it is not enough for some people to follow these strategies ; it must be true that the market on average follows these strategies (on average because we are speaking of aggregate flows here). This is a higher hurdle. In fact, empirical work on this topic in foreign exchange finds the opposite result : the direction of feedback trading in aggregate interbank data is, if anything, negative (Evans and Lyons, 2002b). The empirical case that feedback from price to flow is causing more than a small share of the positive correlation is, at present, not compelling.

The last main alternative to causality running from flow to price is that there is no causality between flow and price and that a third factor is driving both. Consider the following seemingly plausible example. Suppose public macro news arrives that is good for the dollar. Such news would cause the dollar to appreciate and at the same time cause positive flow into the dollar, generating the positive correlation without any causal link. But this view is flawed : it is inconsistent with rational markets (or, in the jargon, inconsistent with “rational expectations”). Information that is truly public, in the traditional sense that everyone observes it simultaneously and interprets it the same way, should induce price adjustment that is instantaneous, i.e., before any transactions can occur at old prices. At the new price, which impounds all the new information, the incentive to buy is no greater than the incentive to sell, so on average positive transaction flow into dollars does not result from the good dollar news (i.e., transaction flow does not systematically correlate with the price adjustment). Now, if one allows instead for public news whose exchange rate implications are not agreed upon by all – a realistic departure from traditional specifications – then the causal link running from flow to price is restored : flow is how price-setting marketmakers learn about the resulting market interpretations. (More precisely, marketmakers learn from flow insofar as their own observation of the public news does not inform them fully about all participants’ views). Flow data provide some nice lessons along these lines (reviewed in the subsection below on Flow Links to Macro Fundamentals).

The remainder of this subsection focuses on this recent work on end-user flows. Specifically, to what extent do end-user flows account for exchange rate movements and how is this affected by extending the horizon? The issue of extending the horizon is important because it bears on critical questions like the persistence of price effects. For addressing longer-horizon effects, end-user flows provide a critical advantage over extant interbank data.

Table 1 presents regression results for the relation between realized excess currency returns in $/€ and concurrent end-user flow (from Evans and Lyons, 2003b ; Citibank data from 1993-2000 ; see also the published tables in my book, The Microstructure Approach to Exchange Rates, MIT Press, 2001, pages 254-255). The regressions include flows from all 6 of the main end-user segments : “Corporations” refers to non-financial corporations, “Traders” refers to hedge funds and other leveraged traders, and “Investors” refers to unleveraged asset managers such as mutual funds and pension funds (these three are broken into US and Non-US parts). (table 1 reports only those end-user segments that were significant at that 1 percent level for all three horizons.) Two points emerge. First, the price impact of trades (dollar for dollar) from different sources is quite different, implying different information content. Second, the explanatory power of flows for concurrent returns rises substantially with horizon : at the daily horizon the regression with all 6 flow segments produces an R² of only 8 percent, whereas the same regression at the 1-month horizon produces an R² of 30 percent. [10] Using end-user flow data from a custodial bank (State Street), Froot and Ramadorai (2002) also find higher correlations between end-user flows and returns as the horizon is extended to 1 month ; their flow measure is institutional investors, which consolidates the “Traders” and “Investors” segments used here.

Table 1
Do different end-user flows have different information potency, dollar for dollar?

Among the many questions that the table might motivate, four are particularly interesting. First, why does explanatory power rise with horizon? This is indicative of non-concurrent correlation between flows and returns, so that when data are time-aggregated, more of this correlation shows up as concurrent. I document non-concurrent correlation in the forecasting analysis below. Second, why should trades’ information content depend on the transacting party’s location? The discussion above regarding risk-neutral fundamentals versus risk fundamentals provides some insight into this question. Think, for example, of the idea that flow helps aggregate dispersed information about aggregate real output or the aggregate risk preferences of financial institutions. We might expect that variations in these variables are in many instances country specific. Thus, the trades originating in a particular country/region are likely to be more highly correlated with changes in these variables from that same country/region, implying different information content than trades from other countries/regions. Third, why are some of these coefficients negative, which seems to have the weird implication that the purchase of euros is bad news for the euro? [11] This is less weird than it might seem. To understand why, consider the negative coefficient from a forecasting perspective. When the market at large learns from a given bank’s interbank trades that that bank has received end-user orders, the market cannot separate underlying segment identities. In contrast, an econometrician using the underlying bank data can separate the identities. Ex post, then, the initial price response in the market to that bank’s interbank trades is too large for some underlying segments and too small for others. When the initial response is too large, the relation of that segment’s flows to subsequent returns will be negative, as the market corrects its unavoidable mistake. (Note that this explanation accounts for negative forecasting coefficients, not negative concurrent coefficients.) The fourth interesting question that arises is whether end-user flows convey information beyond that in interbank flows? The answer is yes : when both types of flow are included in a regression of daily excess returns, the null that the coefficients on end-user flows are zero is rejected at the 1 percent level (based on the 4-month sample of daily Reuters 2000-1 interbank flows used in Evans and Lyons, 2002a).

If indeed dispersed information about future fundamentals is being aggregated through flow, then flow should help forecast those fundamentals. To address this one can test whether flows Granger cause macro variables. Consider a set of variables considered fundamental across a wide range of macro modeling traditions : money growth, output growth, and inflation. Evans and Lyons (2003b) show that flows have more forecasting power than the exchange rate itself, a rather striking result. Flows Granger cause US money growth, US output growth and German inflation, whereas the exchange rate Granger causes only US output growth. (That only inflation is Granger caused for the German variables is perhaps less surprising, given the inflation focus of the Bundesbank over the bulk of the 1993-2000 sample.) There does indeed appear to be information in end-user flows useful for forecasting future fundamentals. This result highlights the qualitatively new issues that end-user data allow one to address. Note too that because the market at large does not have access to these disaggregated end-user flows, the finding that flows are an even more powerful predictor than the exchange rate is not a violation of any simple efficiency criterion.

We turn now to a second approach for examining whether flow conveys macro information. The basic idea of this approach is that if flow is conveying this information, then the information should be impounded in price. Accordingly, the exchange rate should forecast future macro variables. Effectively, this is what the Granger causality tests address, though they do so in a rather particular way. Our second approach does so in a completely different (complementary) way. The starting point is an influential paper by Mark (AER, 1995). Mark estimates regressions of the following form :

where p_t+k – p_t is the log change in the spot rate over k periods and f_t denotes fundamentals measured at time t (measured from money supply and GDP data – see paper for details). The idea is that exchange rates may be “catching up” to fundamentals over time, i.e., if price is below fundamental value today (f_t > p_t), then on average it should rise over time – a positive coefficient β. (This is essentially an error-correction model.) Mark shows that as the horizon is lengthened, the fit of the regression gets better and better. At horizons in the 3-5 year range, the model fits the data quite well, suggesting that ultimately fundamentals are getting into exchange rates, but it takes considerable time.

An alternative approach turns the basic logic of the Mark regression on its head. The underlying idea motivating the Mark regression is that exchange rates do a poor job of tracking fundamentals over shorter horizons. Under this view the FX market is not working particularly well in impounding macro information. The view that order flow is important in aggregating macro information is in some sense quite the opposite : order flow is conveying information to the market today that is relevant to the paths of macro variables in the future. Thus, under this view, order flow acts to telescope the market’s forward-looking expectations and impound changes in those expectations into today’s price. This is a view that markets are instead working quite well, and that the problem with the earlier analysis in the literature is that the variable f_t (money supply and real output, but only their current values) is a poor measure of the expected future fundamentals that the market is discounting in price.

This opposite view to that underlying the work of Mark can be captured by the following variation on his regression :

The idea now is that if price today is forecasting strong future fundamentals (p_t > f_t), then those fundamental macro variables should tend to strengthen – a positive coefficient β. What one finds, in a nutshell, is that the supporting evidence from this version of the equation is as least as strong as that found by Mark. This suggests that exchange rates are indeed impounding forward-looking information as one would expect of an asset market.

I do not report those results here for three reasons. First, the lengthening of the horizon k together with overlapping observations makes the test statistics for evaluating the significance of β and the regression fit tenuous. Indeed, several papers were published in the years following the Mark (1995) paper that were dedicated to showing how tenuous inference really is in this particular setting. Second, a paper by Engel and West (2002) effectively addresses the same question. (They also found that exchange rates have some power to forecast macro variables, but still not so much as theory would predict.) Third, remember that this approach makes no use of end-user flows whatsoever. We embarked on this path because at the time we believed that flows could not possibly be forecasting macro aggregates if exchange rates were not forecasting macro aggregates. But I no longer believe this : there is no good reason why all the information in end-user flows is necessarily impounded in price, so that price makes end-user flows redundant as a forecasting variable. Indeed, the forecasting results below (forecasting returns using flows) imply that the current exchange rate does not already impound all the information in end-user flows. Given this, it is not a stretch to believe that current exchange rates do not already impound all the information in end-user flows that is specifically macro related.

The third type of evidence we produced regarding links between end-user flows and macro information is based more specifically on the arrival of macro news. This analysis borrows heavily from the methodology introduced in Evans and Lyons (2003a). The basic idea is the following. We can measure daily rates of macro news flow by measuring the macro news items that arrive each day over the Reuters Money Market Headline News screen. (The median number of arrivals per day over the end-user flows sample relating to US or German macroeconomics is 11, with considerable variation around this median.) We then ask the question : is the variance of segment flows higher on days with a lot of macro news flow? The answer to this question is very definitely yes, and it is yes for all 6 of the segments introduced in table 1. We then model flow determination in a more structured way to determine what fraction of flow volatility is accounted for by macro news arrival (in the jargon, we do a variance decomposition). The answer is that in the $/€ market (the only one for which the analysis has been done thus far), macro news arrival accounts for 30-40 percent of each of the six segment flow variances (monthly frequency). Though not direct evidence, in itself, that end-user flow segments convey macro information, these results indicate that end-user flows are linked to the process of macro information flow.

Let us shift from the explanatory regressions in table 1 (concurrent flow and return data) to the corresponding predictive regression (see table 2 notes, next page, for regression model). The forecasting results in Panel A of table 2 are really quite striking : the coefficients on US corporate and investor flows are highly statistically significant and the R² statistic is 19 percent. [12] (Only coefficients with forecasting power at the 1 percent level are reported, the same threshold used in table 1.) For comparison, the big forecasting puzzle in the empirical literature is the so-called forward bias puzzle, the well documented result that monthly forecasting using the beginning-of-period forward discount produces significant coefficients. But despite this being the best forecasting variable identified in academic research, in that case R² statistics are generally only 2-4 percent.

Table 2
Forecasting excess returns 1 month ahead with end-user flows

Going a step further, recall that an important objective is to clarify the nature of the underlying dispersed information driving flows. For example, it is possible that the forecasting power in end-user flow data arises because flow forecasts risk-neutral fundamentals, i.e., the variables that enter a monetary model with no risk premium (money supplies, outputs, interest rates, price levels). That flows do indeed have some forecasting power in this narrow respect is indicated by the Granger causality results. But it is also possible that flows are instead forecasting risk fundamentals, e.g., determinants of foreign exchange risk premia. The most obvious example is the possibility that flows forecast future flows, which will have concurrent price impact when realized due to simple market absorption effects. (If the flows forecasting future flows are not public information, and end-user flows are not, then there is no market efficiency argument that requires all the price impact to occur with the initial flow realization.) This would produce a correlation between current flows and future returns, but it would have nothing to do with forecasting future macro fundamentals.

As a first cut on disentangling these two distinct views on why flows forecast returns (i.e., on the type of information flows convey), consider a decomposition of the excess return regressions. The basic idea is the following. We know that flows forecast future returns. We also know that flows are an important concurrent determinant of returns. If flow is forecasting future flow, then the part of the future returns that flow should forecast is the part that is determined by those future flows. Accordingly, one can regress returns on concurrent flows and use the fitted values from the regression to separate the return series into a flow explained part (the fitted values) and the flow-unrelated part (the regression residual). One can then determine whether flow tends to forecast the flow-explained part of the return or the flow-unrelated part. The former would suggest the “flow forecasting future flow” channel. The latter is consistent with the “flow forecasting narrow fundamentals” channel.

The punch-line of the Panel-B results is that flow has power to forecast both components. That said, roughly two-thirds of this forecasting power is unrelated to future flow effects on price (the power to forecast the future regression residual). This is consistent with the Granger causality results : flow appears to be forecasting price components that are impounded in price without any relation to flow, or at least without any relation to Citibank’s end-user flow. Whether this last statement generalizes to more encompassing definitions of flow is a question that requires richer flow data sets than are currently available.

There are two key elements for thinking about sustainability : segment structure and information structure. Consider segment structure first. By highlighting segment structure I mean to highlight the fact that the six segments used in the regressions above definitely do not share the same concurrent relation with price, dollar for dollar, nor the same forecast relation with price. This is good news for sustainability. If, in contrast, all flows had statistically indistinguishable links to price, then other players using even small subsets of end-user flows data would essentially have access to the same market signals. This would increase the competition in trading on these signals, and would rather rapidly render future prices independent of current flow information. This is not the case. Some end-user flow datasets do not include non-financial corporate flows, which are quite important (from table 2). Similarly, the fact that US clients appear more valuable for forecasting (both US corporates and US investors) may be bad news for end-user datasets with low penetration among US clients (if this property of the data is robust).

The second key element for sustainability is what I call information structure. With this I am referring back to the organizing framework provided by the separation of risk-neutral fundamentals and risk fundamentals. Now, the correlation of flow signals received by different market participants is clearly important for sustainability. But rather than think about this correlation in a one-dimensional way, it is useful to consider the two dimensions provided by these two distinct types of information. The question then becomes the following : in each of these two separate categories, how correlated is a given set of flow signals with the flow signals of other participants? What this does is open conceptual room for understanding why a given institution’s flow data might convey particular types of information well relative to data from other institutions, contributing to sustainability. Think, for example, of the category of risk fundamentals. Effectively, in that category what matters is the total net flow in the market and the effective time horizon over which the position changes underlying that flow are likely to be held. Suppose, in the extreme, that the flow received by each bank is an idiosyncratic realization of end-user demand shocks. In this setting, what matters for forecasting price is the sum of all these orders. And no matter how aggressively other banks capture and exploit their own flow data, the end-user flow received by a given bank cannot be inferred from the trades of others, and therefore cannot be impounded in price until the market gets some sense for them. (This is in fact the information structure in the model of Evans and Lyons 2002.) In this instance, a bank’s information advantage would be sustainable (as would the separate information advantages of the other banks). My point is not that the “idiosyncratic flows” assumption is realistic. It is rather that a more disaggregated view of the different components of fundamentals can only enhance one’s appreciation for the sustainability of individual banks’ flow-based information advantages. The one-dimensional view of signal correlation is too simplistic.

We find that abrupt, short-horizon price moves (e.g., daily) are more closely tied to the trades of financial institutions. This is perhaps unsurprising, but there are some subtleties nonetheless. For example, we find that on that phenomenal day in the dollar-yen market in October 1998 (price fell from over 130 yen to the dollar to less than 120), it was the financial institutions (at least at Citibank) that were selling the dollar. The subtlety is that is was not the hedge funds who were selling dollars heavily. (The story in the popular press was that hedge funds were all trying to unwind their dollar-yen carry trades that day.) In fact, hedge funds bought dollars, net, that day. We find that the unleveraged (“real-money”) investors were the ones selling the dollar. Like hedge funds, non-financial corporates were also net buyers of dollars that day, though their net buying was a good deal smaller than that of hedge funds. Turning to a sample that includes other daily rate moves that were among the largest, we find the same result that financial institutions were much more likely on those days to be trading in the direction of the rate than non-financial corporations. Beyond the October 1998 yen-dollar event, though, it is not systematically true that the real-money investors were more likely than the hedge funds to be the big net traders on those days.

For longer horizon price moves, in contrast, we find that non-financial corporate flows are more likely to be the driver. We offer two pieces of evidence. First, if one looks at the major downward trend in the dollar-euro rate from 1999 to 2002, one finds that only the non-financial corporates track the long trend in the exchange rate in terms of cumulative euro selling. This is not a statistical test, but it certainly does make it difficult to understand how the financial institution flows could account for the cumulative net change in the rate. Second, and more statistically based, note that the only coefficient in the forecasting results (table 2) that is positive is for corporates (US corporates in particular). The financial segments in that table show coefficients that are, on balance, negative, meaning that if they are buying euros this month then the euro tends to depreciate next month. As noted above in the discussion of how to understand negative forecasting coefficients, it may simply be that the market initially over-reacted to these financial-institution flows, and that is why the subsequent effect is so strongly negative. In any event, in the process of time-aggregating the date underlying table 2, we found that as one extends from 1 week, to 2 weeks, to 3, and then 4, the coefficients on corporates remain positive and get progressively larger, whereas the coefficients on investors get progressively more negative. This too is consistent with positive corporate-flow effects on price that take longer to impound in price, but are ultimately stronger than financial-institution flow effects (which appear to decay after about a year ; see Froot and Ramadorai, 2002).

These lower frequency relationships are absolutely crucial to most practitioners so let me add a couple more comments. First, that corporate flows should be relatively important at longer horizons makes sense from a theoretical perspective. International macro models put great emphasis on the underlying real economy as a low-frequency driver of exchange rates. These models also highlight the longer-horizon importance of balance-of-payments sustainability, which is intimately linked to the corporate activity that produces the current account. Second, there remains much work to be done with respect to corporate flows, e.g., linking them to data sources for public flow such as those within official balance of payments statistics.

Of the two organizing ideas, the first is the more provocative, so let me clarify why I believe that the structure of currency markets is indeed deeply dependent on credit risk. An important factor in considering whether current FX customers would want to participate in an open, electronic auction market (with all participants on equal footing) is whether they would be willing to assume the counterparty credit risk arising from trading bilaterally with other non-bank customers. Banks themselves are careful not to do this without tight management of the attendant exposures. Non-financial institutions, or even financial institutions like pension funds, hedge funds, and mutual funds, are in general quite averse to assuming such risk. True, these institutions are familiar with trading currency futures, for which clearinghouses have been designed to mitigate this risk. But a clearinghouse system of sufficient size to handle the worldwide spot market and at the same time offer flexible access is unlikely to be economically practicable given the different national legal jurisdictions and concomitant legal risk.

Commercial banks, on the other hand, are particularly well suited to play this role. They have developed significant relationships with institutional customers. Indeed, there is a long line of research in corporate finance based on banks’ comparative advantage in monitoring the credit-worthiness and performance of their corporate customers. These relationships, together with banks’ expertise in managing credit risk, gives them an advantage that is, in my judgment, the key driver of the current bank-intermediated dealership structure. [13] It contributes substantially to bank dealers’ ability to compete with the new, auction-based alternatives recently offered to FX customers by new entrants. [14]

In the paper I offer three scenarios for the future of FX market institutions. These three scenarios are not only the most likely (in my judgment), when considered jointly they also highlight the competitive forces at work. The time frame I have in mind is 10 years.

But if banks are making little money, why are they so adamant about hanging on to the liquidity-provision part of the business? Why not just let Future 1 play itself out? Enter the indirect profitability of FX liquidity provision. Large FX banks now recognize an important source of indirect profits from intermediating flows, namely their ability to exploit information contained in those flows. Indeed, one of the factors that is depressing spreads so much for the largest, most coveted customers is the intense bidding for this flow information by dealing banks. The banks are, in a very real sense, investing in flow information by cross-subsidizing their comparative disadvantage in liquidity provision. [16]

R. K. L.