## Demand for Dollars: Evidence from Survey Expectations (Working Paper No. WP/2026/186)

## Source details

**Canonical URL:** [Demand for Dollars: Evidence from Survey Expectations (Working Paper No. WP/2026/186)](https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026186-source-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2026/english/wpiea2026186-source-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2026/english/wpiea2026186-source-pdf.pdf.json)

---

### The US dollar’s role and data scope
- Dollar is "the linchpin of the global monetary and financial system," dominant for investment, funding, and hedging.
- Key statistics:
  - Foreign investors hold 30% of US Treasuries outstanding.
  - Foreign banks borrow around 230 billion US dollars via FX swaps each day.
  - More than 90 percent of bond mutual funds use currency forwards for hedging.
  - CLS operates the world’s largest currency cash settlement system, handling over 50% of the global spot, forward, and swap FX transaction volume.
  - The G10 currencies against the US dollar comprise more than 60% of global FX trading volume.
- Data and sample:
  - CLS granular hourly data covering spot, forward, and swap flows; sample period: September 2012 to May 2025.
  - Currency focus: USD versus AUD, CAD, CHF, EUR, GBP, JPY, NOK, NZD, SEK (G10).
  - FX swaps analysis focuses on maturities of less than 35 days to capture synthetic USD funding.
  - Expectations: Consensus Economics monthly forecasts (3-, 12-, 24-month FX; 3-month and 10-year interest rate projections; current account, CPI, real GDP forecasts).

### Theoretical framework and core mechanisms
- Investor types and instruments:
  - Directional traders take unhedged dollar positions in spot and forward markets.
  - Dollar borrowers use FX swaps to finance dollar asset holdings; care about synthetic dollar funding price.
  - Single dealer sector intermediates spot (SPT), forward (FWD), and swap (SWP) markets using a single balance sheet.
- Key equilibrium insights:
  - Dealers jointly determine the cost of warehousing directional risk and the cost of supplying synthetic dollar funding.
  - Cross-currency basis x equals the shadow cost of dealer balance-sheet capacity.
  - Expected dollar appreciation increases dealers’ directional book and crowds out capacity to supply synthetic dollars, widening the basis (x becomes more negative) and reducing synthetic dollar borrowing.
- Factor structure and attention to public forecasts:
  - Expected currency excess returns follow E_j[r_f^x] = β' E_j[λ]; directional traders rebalance along factor loadings β (dollar, carry, UMVE).
  - Directional traders combine private signals and public survey forecasts: E_j[λ(h)] = (1 − ω_j) λ_j(h) + ω_j λ_F(h).
  - Weight ω_j falls with forecaster dispersion σ(λ_F): ∂ω_j/∂σ(λ_F) < 0; response to macro uncertainty ξ depends on risk aversion γ_j.

### Three core model predictions (summary)
- Prediction 1 — Expected dollar appreciation:
  - Directional traders buy more dollars in spot and forward markets when expecting USD appreciation.
  - Dollar borrowers reduce synthetic USD borrowing via FX swaps when the dollar is expected to appreciate.
  - Dealer banks provide liquidity and take the opposite side across markets.
- Prediction 2 — Cross-sectional rebalancing along currency risk factors:
  - Investors rebalance from riskier foreign currencies into USD more than from safer currencies when USD appreciation is expected; analogous patterns for FX-funding flows conditional on expected funding-cost changes.
- Prediction 3 — Information value and uncertainty:
  - Investors place less weight on professional forecasters when forecaster dispersion increases and/or in times of higher uncertainty; survey expectations' predictive power for flows weakens with higher disagreement or VIX.

### Empirical findings — directional flows (spot and forward)
- Spot flows (monthly predictive regressions of next-month flows on 24-month FX forecasts):
  - Funds: FX Change coefficient 0.323*** (buy USD when USD expected to appreciate).
  - Non-Dealer Banks: FX Change coefficient 0.539*** (example magnitude: a consensus forecast of a 1% USD appreciation associated with non-dealer banks buying, on average, 0.54 billion USD over the following month).
  - Dealer Banks: FX Change coefficient −0.744*** (dealers supply USD; take opposite side).
  - Corporates: −0.029*; NBFIs: −0.085***.
  - Overall R^2 by column reported (examples): 0.335, 0.240, 0.319, 0.182, 0.266.
- Forward flows:
  - Excess FX Change (exchange-rate-change minus 24-month interest differential) predicts forward purchases:
    - Funds: 0.279***; Non-Dealer Banks: 0.105*; Dealer Banks: −0.305***.
  - Interest Rate Diff. coefficients differ by investor type (e.g., Funds: 0.265***; Dealer Banks: −0.196***).
- Short-term (3-month) vs long-term forecasts:
  - Raw 3-month FX forecasts have limited predictive power for spot flows.
  - The component of short-term forecasts that is common with long-term forecasts predicts flows strongly; long-horizon expectations explain short-run quantities better.

### Empirical findings — synthetic dollar funding (FX swaps and CIP)
- CIP and swap-flow predictability:
  - CIP formulation: CIP = i_us − (i_foreign − FD_annualized); CIP>0 ⇒ synthetic USD funding cheaper; CIP<0 ⇒ synthetic funding more expensive.
  - Swap regression key coefficients (swap flows in billions of USD; CLS records second leg so negative swap flow = net USD borrowing):
    - Dealer Banks: 130.899*** (supply dollars in anticipation of CIP easing).
    - Funds: −55.198*** (additional 55 billion USD synthetic borrowing for a 1 percentage point expected tightening of the CIP basis—i.e., basis becoming less negative / synthetic funding cheaper; negative coefficient interpreted as funds borrow synthetically when CIP forecast tightens).
    - Non-Dealer Banks: −85.192** (increase synthetic USD borrowing when synthetic funding becomes cheaper).
  - Decomposition:
    - FX Change coefficients in swap regressions: Dealer Banks −3.449***; Funds 0.429**; Non-Dealer Banks 3.037***.
    - Interest Rate responses differ across investor types (e.g., Funds: 12.669***; Non-Dealer Banks: −22.627*** in panel decomposition), indicating heterogenous funding-driven behavior.

### Empirical findings — cross-sectional heterogeneity along currency risk factors
- Directional flows interact with currency betas (spot flows, FX Change × beta interactions):
  - FX Change × Carry Beta:
    - Funds: 0.362*** (funds sell riskier carry currencies more into USD when USD appreciation expected).
    - Non-Dealer Banks: 0.406***; Dealer Banks: −0.819*** (dealers supply USD, with negative interaction).
  - FX Change × Dollar Beta and × UMVE Beta show similar heterogenous patterns consistent with rebalancing from high-beta (riskier) to low-beta currencies.
- Funding flows interact with betas (swap flows, CIP Deviation × beta interactions):
  - CIP Deviation × Carry Beta:
    - Funds: 74.283*** (when synthetic USD funding expected to tighten, funds reduce synthetic USD borrowing less via high-carry-beta currencies than via low-carry-beta currencies).
    - Dealer Banks: −136.356**; Non-Dealer Banks: 59.929 (heterogenous responses).
  - CIP Deviation × UMVE Beta:
    - Funds: 96.858***; Non-Dealer Banks: −140.712*.
  - Overall: funding responses propagate through factor structure (carry, UMVE) rather than consistently through dollar beta.

### Empirical findings — macro forecasts and forecaster disagreement
- Macro forecast predictors (spot flows, controlling for FX Change and macro differentials):
  - CA / GDP differential: strong predictive power (example coefficients):
    - Dealer Banks: 0.611***; Funds: −0.157*; Non-Dealer Banks: −0.343**.
    - Economic interpretation: a one percentage point expected decrease in the CA / GDP differential associated with non-dealer banks purchasing an additional 0.34 billion USD over the next month.
  - Interest Rates, Inflation, GDP differentials exhibit heterogeneous and investor-type-specific effects (see Table 7).
- Forecaster disagreement and state-dependent attention (High Disagr interactions):
  - Construct High Disagr via first principal component of forecaster dispersion (US and foreign) > sample median; interact with forecast differentials.
  - Key patterns (Table 8):
    - Interest Rates × High Disagr: coefficients vary by investor group; e.g., Funds: −0.927** (reduced sensitivity under high disagreement).
    - CA / GDP × High Disagr: Dealer Banks coefficient −0.543*** (muting of predictive power in high disagreement regimes).
    - Inflation × High Disagr and GDP × High Disagr show attenuated or reversed effects for many investor groups.
  - Complementary result: high global VIX yields qualitatively similar weakening of predictability.
  - Forecaster-level evidence: forecasters with larger past errors have less predictive power for investor currency flows.

### Robustness and causality checks
- Forecast revisions and causality:
  - Forecast revisions add incremental predictive power for some participants, but the level of survey forecasts is the dominant predictor of flows.
  - Granger-causality tests: survey expectations predict subsequent currency flows; past flows do not predict subsequent survey revisions.
- Alternative specifications and controls:
  - Results robust to alternative fixed effects, sample currency restrictions, controls for countries’ invoicing shares, and univariate specifications.
- Key robustness takeaways:
  - Prevailing level of survey forecasts is primary predictor of currency flows across instruments and participants.
  - Forecast revisions matter incrementally for some groups.
  - Directional causality runs from survey expectations to currency flows, not vice versa.

### Contributions and policy-relevant implications
- Contributions:
  - Demonstrates that survey-based exchange rate and macro expectations predict global dollar demand across spot, swap, and forward markets for granular market participants.
  - Links expected increases in the cost of synthetic dollar funding to reduced synthetic dollar borrowing by funds and smaller non-dealer banks; documents cross-sectional differences tied to currency risk exposures.
  - Provides evidence that funds buy dollars in forward contracts when professional forecasters expect USD appreciation over the next two years, complementing literature on partial hedging by mutual funds.
- Policy-relevant mechanisms:
  - Dealers’ balance-sheet constraints and the endogenous cross-currency basis amplify the effect of expected USD appreciation on funding costs and global dollar availability.
  - Heterogeneous investor responses and factor-structured rebalancing imply that shocks to USD expectations can propagate unevenly across currencies and market segments, with FX swaps playing a dominant role in international dollar funding.

*Source: wpiea2026186-source-pdf - Demand for Dollars: Evidence from Survey Expectations (Working Paper No. WP/2026/186)*

### 1.  Introduction

### 1.  Introduction

### The US dollar’s central role and heterogeneous demand
- The US dollar is described as "the linchpin of the global monetary and financial system," serving as the dominant currency for investment, funding, and hedging.
- Illustrative statistics reported:
  - Foreign investors hold 30% of US Treasuries outstanding according to the Federal Reserve System.
  - Foreign banks borrow around 230 billion US dollars via foreign exchange (FX) swaps each day (Kloks, Mattille, and Ranaldo,2024).
  - More than 90 percent of bond mutual funds use currency forwards for hedging (Sialm and Zhu,2024).
  - CLS operates the world’s largest currency cash settlement system, handling over 50% of the global spot, forward, and swap FX transaction volume.
  - The G10 currencies against the US dollar comprise more than 60% of global FX trading volume (BIS,2025).
- Dollar demand is heterogeneous across traded instruments (spot, forwards, swaps) and market participants (banks, corporations, investment funds, non-bank financials), making the composition and determinants of that demand critical for policymakers and academics.

### Theoretical framework and investor types
- The paper builds on canonical international finance models (e.g., Gabaix and Maggiori,2015; Itskhoki and Mukhin,2021) and augments them with:
  - A strong factor structure in exchange rates (Lustig, Roussanov, and Verdelhan,2011; Verdelhan,2017).
  - The idea that investors incorporate publicly available predictions by professional forecasters when forming expectations about future currency returns.
- Investor types in the model:
  - Directional traders: take unhedged dollar positions in spot and forward markets.
  - Dollar borrowers: finance acquisition of dollar assets via FX swaps and are primarily concerned with the price of synthetic dollar funding.
  - A single dealer sector intermediates all three markets using a single balance sheet.
- Equilibrium mechanism: the cost of warehousing directional risk and the cost of supplying synthetic dollar funding are jointly determined by dealers and the investor demand structure.

### Model predictions (three core predictions)
- Prediction 1: Responses to expected dollar appreciation
  - Directional traders buy more dollars via spot and forward markets if they expect a dollar appreciation.
  - Dollar borrowers engage in less synthetic dollar borrowing via FX swaps when the dollar is expected to appreciate.
  - Dealer banks provide liquidity and take the opposite side of these trades across all three markets.
- Prediction 2: Cross-sectional rebalancing along currency risk factors
  - Investors rebalance from foreign currencies into USD along the factor structure of currency risk.
  - Following an expected appreciation of the dollar, investors sell more high risk foreign currencies (e.g., Australian dollar) and fewer low risk foreign currencies (e.g., Japanese yen).
- Prediction 3: Information value and uncertainty
  - Investors place less weight on professional forecasters’ predictions about exchange rates and macroeconomic fundamentals when forecaster dispersion increases and/or in times of higher uncertainty.

### Empirical strategy and data
- Primary quantity data: granular and comprehensive dataset on global FX spot, forward, and swap trading activity from CLS Group (CLS).
  - CLS handles over 50% of global spot, forward, and swap FX transaction volume.
- Currency coverage: G10 currencies against the US dollar (comprising more than 60% of global FX trading volume).
- Expectations data: monthly surveys of finance professionals conducted by Consensus Economics (global banks, research institutes, and investors active in the currency market).
- Measurement choices:
  - Dollar demand measured as currency transactions that either involve the purchase or borrowing of US dollars by investors (clients) from dealer banks.
  - Forecast-based expectations of covered interest parity (CIP) deviations constructed to assess synthetic dollar funding dynamics.

### Empirical findings (three empirical dimensions supporting theoretical predictions)
- Evidence for Prediction 1 (forward-looking expectations matter for flows)
  - Directional trades: An expected appreciation of the dollar over the horizon of the next two years predicts that investment funds and non-dealer banks buy dollars and sell foreign currency over the next month using both spot and forward transactions.
  - Forward contracts: Purchasing dollars via forward contracts is consistent with both speculative motives and hedging motives.
  - Dealers and NBFIs are predicted to take the opposite side by providing currency liquidity.
  - Forecast horizon comparison:
    - Long-horizon exchange rate expectations explain short-run currency flows much better than short-horizon expectations.
    - Short-term (3-month) FX forecasts have only limited predictive power for spot currency flows; however, the common component of short-term and long-term FX forecasts accurately predicts currency flows.
- Evidence on synthetic dollar funding (FX swaps and CIP deviations)
  - When synthetic dollar funding becomes cheaper, investment funds and non-dealer banks engage in more synthetic dollar borrowing.
  - Borrowing decisions of both funds and non-dealer banks are driven by exchange rate expectations: they reduce USD borrowing when the dollar is expected to appreciate, consistent with a widening of the basis.
  - Smaller non-dealer banks increase synthetic dollar borrowing when US interest rates rise relative to foreign rates (consistent with cheaper synthetic dollar funding).
  - Funds exhibit the opposite pattern to small banks, increasing USD lending following a rise in US interest rates, suggesting investment funds’ FX swap trading is not purely funding-driven.
  - Dealer banks again take the opposite side of the market by providing liquidity.
- Evidence for Prediction 2 (cross-sectional heterogeneity along currency factors)
  - Two currency factor models considered:
    - The two-factor model by Lustig et al.(2011) and Verdelhan(2017) (dollar and carry factor).
    - The one-factor model by Chernov, Dahlquist, and Lochstoer(2023) (unconditional mean variance efficient portfolio).
  - Findings:
    - Conditional on expecting the USD to appreciate, investors exchange riskier currencies more heavily for USD than safe ones.
    - For funding USD via FX swaps, conditional on an expected increase in the cost of synthetic dollar funding, funds reduce dollar funding less via high-carry-beta currencies (e.g., Australian dollar) than via low-carry-beta currencies (e.g., Japanese yen).
- Evidence for Prediction 3 (forecaster disagreement and uncertainty)
  - Predictive power of survey expectations for currency flows declines significantly as forecaster disagreement increases, consistent with investors placing less weight on public signals when their informational content is more uncertain.
  - Complementary analysis using the VIX yields qualitatively similar evidence.
  - Overall, the influence of survey expectations on currency demand weakens in periods of elevated disagreement and uncertainty.

### Contributions relative to existing literature
- Extends literature on order flow and exchange rates by showing that exchange rate expectations and expectations about macroeconomic fundamentals predict global dollar demand across spot, swap, and forward markets for granular market participants.
- Adds to literature on synthetic dollar funding and CIP deviations by linking expected increases in the cost of synthetic dollar funding to reduced synthetic dollar borrowing by investment funds and smaller non-dealer banks, and by documenting cross-sectional differences tied to currency risk exposures.
- Contributes to literature on currency hedging by providing evidence that funds buy dollars in forward contracts if professional forecasters expect the dollar will appreciate over the next two years, complementing studies that document partial hedging by mutual funds (e.g., mutual funds hedge on average only 21% of foreign currency denominated asset holdings in Du and Huber,2023).

*Source: wpiea2026186-source-pdf - 1.  Introduction*

### 2.  Theoretical Framework

### 2.  Theoretical Framework

### Environment
- Model features heterogeneous investors and intermediaries; builds on Gabaix and Maggiori(2015), Itskhoki and Mukhin(2021), Vayanos and Vila(2021), and Kekre and Lenel(2024), augmented by factor structure in FX returns.
- Investor groups:
  - Directional traders (D): J types indexed by j ∈ {1, . . . , J}, market shares θ_j with ∑_J_j=1 θ_j = 1. They allocate wealth across N currencies; US dollar is home currency for all. A position at horizon h views currency excess return over [t, t+h], r_f^x(h) ≡ ∑_h_{k=1} r_f^x_{t+k}, with covariance Σ(h).
  - Dollar borrowers (B): unit mass, hold dollar assets and use FX swaps to finance them; concerned with synthetic dollar funding price rather than FX risk.
  - Dealer sector: single dealer intermediates spot (SPT), forward (FWD), and swap (SWP) markets using a single balance sheet. Cost of warehousing directional risk and supplying synthetic dollar funding jointly determined.
- Directional trader objective (Eq. (1)):
  - max_{W_j} E_j[−1/γ_j e^{−γ_j W'_j r_f^x(h)}], where W_j is vector of dollar positions, γ_j > 0 is constant absolute risk aversion, r_f^x(h) is vector of currency excess returns.
- Instruments:
  - Spot valuation change: r_SPT(h) = ∆s_{t,t+h}. (Eq. (2))
  - Forward return: r_FWD(h) = ∆s_{t,t+h} + (i_t − i^*_t) + x_i. (Eq. (3))
  - Swap nets out currency exposure and yields interest differential up to cross-currency basis x_i.
- Dollar borrower problem (Eq. (4)):
  - max_{B_i} (x_i + c) B_i − κ/2 B_i^2, where B_i is amount of synthetic dollar borrowing, c ≥ 0 is funding-access premium, κ captures cost of running larger synthetic dollar position.
- Dealer sector objective (Eq. (5)):
  - max_{Q_fx, Q_funding} Q'_fx E_D[r_f^x] + Q'_funding (−x) − γ_D/2 Q'_fx Σ Q_fx − τ/2 (∥Q_fx∥ + ∥Q_funding∥)^2, with dealer risk aversion γ_D and balance-sheet cost τ > 0.

### Optimal Choices
- Directional traders (mean-variance): first-order condition (Eq. (7))
  - W^*_j = 1/γ_j Σ^{−1} E_j[r_f^x].
  - Implication: directional traders tilt toward currencies expected to appreciate against the dollar; applies to both spot and forward markets.
- Proposition 1a (Directional flows: spot and forward):
  - Directional traders buy more US dollars through spot and forward instruments when they expect the dollar to appreciate, since W^* is increasing in E[r_f^x] and both the spot and the forward returns increase with the value of the dollar.
- Dollar borrowers: first-order condition yields (Eq. (8))
  - B^*_i(x_i) = (x_i + c)/κ, increasing in the cross-currency basis.
- Intermediaries: first-order conditions share same shadow cost τ(∥Q_fx∥ + ∥Q_funding∥):
  - directional: E_D[r_f^x] − γ_D Σ Q_fx = τ(∥Q_fx∥ + ∥Q_funding∥). (Eq. (9))
  - funding: −x_i = τ(∥Q_fx∥ + ∥Q_funding∥). (Eq. (10))
  - Eq. (10) interprets cross-currency basis as equal to shadow cost of dealer balance-sheet capacity.

### Market Clearing and Endogenous Cross-Currency Basis
- Directional market clearing:
  - Aggregate directional demand W_fx = Γ Σ^{−1} ̄E[r_f^x], with Γ ≡ ∑_j θ_j/γ_j and ̄E[r_f^x] ≡ (1/Γ) ∑_j (θ_j/γ_j) E_j[r_f^x].
  - Dealer directional book (from Eqs. (9) and (10)) (Eq. (11)):
    - Q_fx = (γ_D Σ)^{−1} (E_D[r_f^x] + x).
  - Market clearing W_fx + Q_fx = 0 implies (Eq. (12)):
    - E_D[r_f^x] = −(γ_D Γ ̄E[r_f^x] + x).
  - Interpretation: dealers’ beliefs must be negatively related to aggregate investor beliefs; magnitude governed by γ_D Γ; basis x enters additively, loosening balance-sheet constraint when funding revenue −x is earned.
- Proposition 1b (Directional flows: liquidity provision):
  - When dealers act as pure liquidity providers they take positions opposite to aggregate directional demand. If directional traders expect the dollar to appreciate and purchase dollars, dealers sell dollars to accommodate this demand.
- Funding market and basis determination:
  - Directional book size ∥Q_fx∥ proportional to Γ ̄E[r_f^x]; increases with expected dollar appreciation.
  - Dealer funding first-order condition implies (Eq. (13)):
    - ∥Q_funding∥(x) = −x/τ − ∥Q_fx∥, decreasing in the directional book.
  - Clearing funding market ( (x + c)/κ = −x/τ − ∥Q_fx∥ ) pins down basis (Eq. (14)):
    - x^* = − τ (κ ∥Q_fx∥ + c) / (τ + κ).
  - Comparative statics:
    - Basis widens as balance sheet costs increase since ∂x^*/∂τ < 0.
    - Relation to expected dollar appreciation (Eq. (15)):
      - ∂x^*/∂̄E[r_f^x] = (∂x^*/∂∥Q_fx∥)_{<0} (∂∥Q_fx∥/∂̄E[r_f^x])_{>0} < 0.
    - Intuition: expected dollar appreciation enlarges dealer directional book, crowding out capacity to supply synthetic dollars, widening the basis; W_funding = (x^* + c)/κ declines as x^* becomes more negative.
- Proposition 1c (Funding flows: swaps):
  - The cross-currency basis equals the shadow cost of dealers’ balance sheet capacity. An expected dollar appreciation increases the dealer’s directional book and thereby widens the basis, which in turn reduces synthetic dollar borrowing via FX swaps. Dealer banks take the opposite side of this trade and supply dollars synthetically.

### Factor Structure and Cross-Sectional Predictions
- Expected currency excess returns follow factor structure: E_j[r_f^x] = β' E_j[λ], where β is N×K matrix of factor loadings (e.g., dollar beta, carry beta, UMVE beta) and E_j[λ] is investor j’s expectation of factor risk premia.
- Implication for directional traders:
  - Since E_j[r_f^x] is increasing in β, Eq. (7) implies ∂W^*/∂β > 0: directional traders rebalance from foreign currencies into dollars along FX risk factor structure.
- FX swap market: assume factor structure enters through cross-currency basis rather than expected return; basis inherits carry-factor structure (low-interest-rate currencies exhibit most negative bases; high-interest-rate currencies much less negative).
- Proposition 2 (Cross-section of flows):
  - In the cross-section, currency flows propagate through the factor structure of currency risk. For directional flows, following an expected dollar appreciation investors sell more out of higher-risk (large β) and less out of safer (small β) foreign currencies. For FX funding flows, following an expected increase in the cost of synthetic funding investors reduce synthetic dollar borrowing more in low- relative to high-interest rate currencies.

### State-Dependent Attention to Public Forecasts
- Directional traders form expectations combining private information and public survey forecasts (Eq. (6)):
  - E_j[λ(h)] = (1 − ω_j) λ_j(h) + ω_j λ_F(h), with ω_j ∈ [0, 1] weight on public information.
- Weight ω_j depends on macroeconomic uncertainty ξ, investor risk aversion γ_j, and dispersion across professional forecasters σ(λ_F): ω_j = ω_j(ξ, γ_j, σ(λ_F)).
  - Restrictions:
    - ∂ω_j/∂σ(λ_F) < 0: greater forecaster dispersion reduces weight on public forecasts.
    - Sign of ∂ω_j/∂ξ depends on γ_j: ∂ω_j/∂ξ < 0 for sufficiently risk-averse investors and ∂ω_j/∂ξ > 0 otherwise.
- Marginal effect of public forecasts on currency demand (Eq. (16)):
  - ∂W^*_{ji}/∂λ_F = ω_j(ξ, γ_j, σ(λ_F)) / (γ_j σ_i^2) β'_i.
- Cross-partials:
  - ∂^2 W^*_{ji} / ∂λ_F ∂σ(λ_F) = (1 / (γ_j σ_i^2)) (∂ω_j/∂σ(λ_F)) β'_i < 0. (Eq. (17))
  - ∂^2 W^*_{ji} / ∂λ_F ∂ξ = (1 / (γ_j σ_i^2)) (∂ω_j/∂ξ) β'_i ≶ 0. (Eq. (18))
    - First cross-partial unambiguously negative; second inherits sign of ∂ω_j/∂ξ.
- Proposition 3 (Disagreement and uncertainty):
  - Investors place a lower weight on professional forecasters’ expectations when forecaster disagreement rises. The response to macroeconomic uncertainty depends on investors’ risk aversion. Consequently, the response of currency flows to public forecasts weakens with forecaster disagreement and, conditional on risk aversion, with macroeconomic uncertainty.

*Source: 2.  Theoretical Framework (wpiea2026186-source-pdf).*

### 3.  Data

### 3.  Data

### Currency flow data (CLS)
- Data source: CLS Group (CLS), the world’s largest FX settlement system.
- Coverage and participants:
  - Majority of the 72 settlement members are large multinational banks.
  - More than 25,000 “third-party” clients of these settlement members, including other banks, funds, non-bank financial firms, and corporations.
  - CLS covers different groups of currency market participants: corporates, funds, non-bank financial firms (NBFIs), and banks.
- Frequency and instruments:
  - Records buy and sell trading volume in the base currency and number of transactions on an hourly basis, spanning Sunday 9 pm to Friday 9 pm (London time, GMT).
  - Data cover spot, swap, and forward currency flows.
  - Account for approximately 50% of global FX trading volume as reported in the BIS Triennial Surveys.
- CLS data construction:
  - Step 1: Reports transactions between corporates/banks, funds/banks, and non-bank financial firms/banks based on counterparty identity (no information on initiator).
  - Step 2: Proprietary network analysis to categorize all market participants into price takers (buyside) and market makers (sellside).
  - Indirect inference of transactions between price taker banks and market maker banks by computing the difference between total buyside trading activity and reported transactions between banks and corporates/funds/NBFIs; this approximation will generally understate trading activity between price taker banks.
- Measurement conventions:
  - In each hourly interval, buy volume = amount of base currency acquired by price takers from market makers; sell volume = amount of base currency sold by price takers to market makers.
  - Net currency flow (main variable of interest) = total buy volume less total sell volume, denominated in billions of USD (positive values indicate net purchases of USD).
  - When studying flows between aggregate buyside and sellside, perspective is that of liquidity providers: net flows computed as sellside minus buyside flows.
- Sample and currencies:
  - Sample period: September 2012 to May 2025.
  - Focus: G10 currency pairs — the USD versus AUD, CAD, CHF, EUR, GBP, JPY, NOK, NZD, and SEK.
- Instruments and maturities:
  - Spot and forward transactions represent directional flows and are reported across all maturities.
  - FX swaps: analysis focuses on maturities of less than 35 days to capture synthetic USD funding flows.
- Empirical patterns (from Figure 1 and text):
  - FX swaps exhibit the largest net flow imbalances (dominant role in international dollar funding).
  - Spot FX transactions have greater average net flows than forwards but are an order of magnitude smaller than FX swaps.
  - Forward transactions comprise the smallest part of the market.
  - Directional currency flows are dominated by dealer and non-dealer banks and investment funds; corporates and NBFIs account for a relatively modest share of aggregate net currency flows.
- Additional details:
  - For FX swaps, reported flows correspond to the second leg of the transaction.
  - Data are monthly and span the period from September 2012 to May 2025.

### FX and economic forecasts (Consensus Economics)
- Data source: Consensus Economics monthly forecasts.
- Coverage and variables:
  - Exchange rate forecasts for 3-, 12-, and 24-month horizons.
  - Interest rate projections for three-month interest rates and ten-year government bond yields covering 3- and 12-month forecast periods.
  - Macroeconomic fundamentals: current account balances, consumer prices, and real GDP growth (forecasts for current and subsequent year-end).
- Sample and scope:
  - Sample period: September 2012 to May 2025.
  - Focus on G10 currencies and corresponding economies.
- Construction of measures:
  - Forecasts are available at individual forecaster-level; consensus forecasts constructed by aggregating individual forecaster data (average forecast).
  - Measures of forecaster disagreement include highest and lowest forecasts and standard deviation of predictions.
  - Forecast measures are expressed as differentials (US minus foreign) because FX transactions involve two currencies.
    - FX rate change = percentage change in the exchange rate over 24 months, using the USD as the base currency.
    - Interest rate differential = difference between 3-month (annualized) interest rates in the US and the corresponding foreign rates projected 12 months ahead.
    - Current account to GDP (CA to GDP) differential, consumer price differential, and real GDP growth differential = differences between US and foreign forecasts at the end of the subsequent year.
  - Same methodology applied to shorter horizons (e.g., three months) to assess shorter-term forecast importance.
- Forecasted CIP basis (forward-looking measure):
  - CIP basis defined as CIP = i_us − (i_foreign − FD_annualized), where FD is the forward discount annualized according to the maturity between the forward rates.
  - CIP>0 implies direct USD interest rate is higher than synthetic dollar interest rate (synthetic dollar funding cheaper); CIP<0 implies synthetic dollar funding is more expensive.
  - Forward-looking CIP: the CIP deviation implied for a 3-month investment beginning three months ahead (expected arbitrage opportunity over period t+3 to t+6, funded at t).
  - Construction uses Consensus Economics forecasts of 3-month interest rates three months ahead and the implied 3-month forward discount for period t+3 to t+6.
  - Forward discount computed as (log(F0,6) − log(F0,3)) × 360/90 × 100.
- Notes on conversions and proxies:
  - Survey forecasts for CHF, NOK, and SEK are expressed relative to the EUR rather than the USD; conversion to dollar expectations uses the prevailing USD/EUR spot rate and accounts for covariance via expanding historical covariances starting in January 2000.
  - German interest rate forecasts used as a proxy for euro area interest rate predictions (no specific euro area forecasts available).
  - Current account forecasts are scaled by the previous year’s GDP to ensure comparability across countries and time.
- Interpretation and caveats:
  - Consensus forecasts widely used by policymakers, investors, and businesses.
  - Survey forecasts may be biased by strategic incentives; the authors emphasize that their results speak to the predictive power of consensus forecasts for trading flows of the marginal investor, not to the rationality of the forecasts themselves.
  - Option-implied expectations are an alternative but reflect risk-neutral beliefs; the study employs survey forecasts to capture subjective beliefs under the physical measure.
- Empirical summary from Table 1 (average values of forecast variables across currencies):
  - Table 1 columns: Currency | FX Rate Change in PP | Interest Rate Diff. in PP | CA to GDP Diff. in PP | Cons. Prices Diff. in PP | GDP Diff. in PP | CIP Basis in PP
  - Table 1 rows (exact values preserved):
    - AUD −0.909 −0.398 −1.526 −0.199 −0.277 −0.100
    - CAD −2.366 0.141 −1.163 0.197 0.142 −0.065
    - CHF 1.686 1.915 −12.592 1.594 0.590 −0.454
    - EUR −1.286 1.230 −5.425 0.603 0.599 −0.393
    - GBP −1.983 0.233 0.220 −1.042 0.374 −0.207
    - JPY −1.608 1.600 −6.053 1.146 1.070 −0.637
    - NOK −7.090 −0.092 −12.834 0.183 0.238 −0.319
    - NZD −0.056 −1.002 1.502 0.189 −0.313 −0.135
    - SEK −6.368 0.912 −8.228 0.423 0.134 −0.378
  - Notes on Table 1:
    - FX Rate Change defined as (FX_24mo − FX)/FX × 100, with USD as base currency.
    - Interest Rates, CA to GDP, Consumer Prices, and real GDP growth are differentials between the US and the quote country, forecasted 12 months ahead (for interest rates) or at the end of the subsequent year (for the other variables), measured in percentage points.
    - CIP deviation = difference between the annualized forward discount and the annualized interest rate differential between the quote country and the US, with USD as base currency; forward discount computed as (log(F0,6) − log(F0,3)) × 360/90 × 100; interest rate differential embedded in CIP is based on 3-month interest rates forecasted three months ahead.
    - Sample: G10 currencies quoted against the USD; monthly data from September 2012 to May 2025.
  - Aggregate patterns from Table 1 and text:
    - On average, all foreign currencies in the sample are expected to appreciate against the USD over the subsequent 24 months except the Swiss franc.
    - Predicted 3-month interest rate differentials are mostly small and positive (slightly higher US interest rates); notable exceptions: Australia and New Zealand (high-interest-rate currencies).
    - CA to GDP differentials negative for most currencies (US projected to run negative CA / GDP balances).
    - Consumer-price and GDP differentials slightly positive on average (higher real growth and inflation forecasts for the US).
    - CIP basis typically negative, implying synthetic dollar funding is more expensive relative to the cash market; the basis is negative in more than 90% of sample observations, with heterogeneity across currencies.

### Exchange rates and forward contracts (Bloomberg)
- Data source: Bloomberg hourly spot mid quotes and daily forward prices for various maturities.
- Notation and conventions:
  - Spot price for country i at time t: S_i,t.
  - Forward prices: F_i,t.
  - An increase in F or S corresponds to an appreciation of the US dollar relative to the foreign currency (F and S quoted with USD as the base currency).
  - Mid prices used when calculating exchange rate changes ∆logS_i,t.
- Covered interest rate parity and approximations:
  - In normal times, forward rates satisfy covered interest rate parity.
  - Use 2-year forward rates and spot prices to compute interest rate differentials: r_i,t − r_US,t ≈ log F_i,t − log S_i,t.

### Measures of currency risk (currency betas)
- Objective: Study cross-sectional differences across currencies using currency betas from factor models.
- Factor models used:
  - Bivariate model (Lustig et al., 2011; Verdelhan, 2017): includes the dollar and carry factors.
  - One-factor model (Chernov et al., 2023): unconditional mean-variance efficient (UMVE) portfolio.
- Construction of factors:
  - Currency excess returns defined as log S_t − log F_t−1,t; a higher factor return corresponds to an appreciation of the dollar.
  - Dollar factor: average change in all currencies with respect to the US dollar (akin to market portfolio in CAPM).
  - Carry factor: returns of a self-financing long-short strategy that borrows in low-interest-rate currencies to invest in high-interest-rate ones; constructed by sorting currencies monthly into five portfolios based on the one-month forward discount relative to the USD prevailing over the previous month; “long” = highest interest rates relative to the US; “short” = low-interest-rate currencies.
- Sample for factor construction:
  - Start with the same cross-section of 39 countries as in Verdelhan (2017) but exclude euro area countries, leaving 28 countries to construct risk factors.
- Estimation of betas:
  - Dollar beta (β_DOL_i) and carry beta (β_CAR_i) estimated using 60-month rolling window regressions of ∆log S_i,t on the respective factor:
    - ∆log S_i,t = a_i + β_DOL_i Dollar_t + ε_i,t
    - ∆log S_i,t = a_i + β_CAR_i Carry_t + ε_i,t
  - UMVE beta (β_UMVE_i) estimated similarly:
    - ∆log S_i,t = a_i + β_UMVE_i UMVE_t + ε_i,t
- Interpretation:
  - Dollar and carry betas summarize country characteristics that determine a currency’s risk exposure.
  - Countries with large carry betas tend to be on the periphery of the global trade network and typically export commodities and import finished goods; carry betas align with interest rate differentials.
  - Countries with large dollar betas typically face large trade costs in goods.
  - UMVE captures solely the priced risk component of currency returns.
- Additional materials:
  - Table D.1 in the Online Appendix summarizes betas by currency pair.

*Source: 3. Data (Content unit: wpiea2026186-source-pdf - 3.  Data)*

### 4.  Empirical Analysis

### 4. Empirical Analysis

### Currency flow predictability
- Purpose and setup:
  - Estimate predictive regressions of currency flows (spot, forward, swap) on survey forecasts formed at time t to predict flows at t+1.
  - Spot regression specification: SPT_i,t+1 = μ_i + α_t + β′X_i,t + ε_i,t (Eq. (22)); key regressor X_i,t contains forecasts of FX rate changes over the following 24 months (“FX Change”).
  - Forward regression specification: FWD_i,t+1 = μ_i + α_t + β′X_i,t + γ′IRD_i,t,24m + ε_i,t (Eq. (23)); X_i,t contains “Excess FX Change” (forecasted FX change minus 24-month interest rate differential); control for contemporaneous 24-month interest rate differential.
  - Swap regression specification: SWP_i,t+1 = μ_i + α_t + β CIP_i,t + ε_i,t (Eq. (24)); CIP forecasts constructed using three-month-ahead interest rate forecasts and forward discounts.
  - All regressions include currency μ_i and time α_t fixed effects. Standard errors computed using Driscoll and Kraay (1998).

- Spot flows (Table 2, panel (a)):
  - FX Change coefficients (spot flows, in billions of USD):
    - Dealer Banks: −0.744***
    - Corporates: −0.029*
    - Funds: 0.323***
    - NBFIs: −0.085***
    - Non-Dealer Banks: 0.539***
  - Overall R^2: 0.335, 0.240, 0.319, 0.182, 0.266 (by column)
  - Interpretation:
    - Funds and non-dealer banks: positive and significant coefficients → purchase USD when USD expected to appreciate.
    - Dealer banks and NBFIs: negative and significant → supply USD (liquidity providers).
    - Economic magnitude example: a consensus forecast of a 1% USD appreciation is associated with non-dealer banks buying, on average, 0.54 billion USD over the following month.

- Forward flows (Table 2, panel (b)):
  - Excess FX Change coefficients (forward flows, in billions of USD):
    - Dealer Banks: −0.305***
    - Corporates: 0.026
    - Funds: 0.279***
    - NBFIs: 0.023
    - Non-Dealer Banks: 0.105*
  - Interest Rate Diff. coefficients:
    - Dealer Banks: −0.196***
    - Corporates: 0.106***
    - Funds: 0.265***
    - NBFIs: 0.086***
    - Non-Dealer Banks: −0.041
  - Overall R^2: 0.263, 0.478, 0.531, 0.345, 0.226
  - Interpretation:
    - Forward demand driven by expected currency excess returns (exchange rate change + interest rate differential).
    - Funds and non-dealer banks increase forward USD purchases when expecting higher future dollar excess returns; dealer banks sell dollars forward.

- Short-term forecasts (Table 3):
  - Panel (a) FX Change (3m) coefficients (spot flows, in billions of USD):
    - Dealer Banks: −0.136
    - Corporates: 0.049*
    - Funds: −0.052
    - NBFIs: −0.017
    - Non-Dealer Banks: 0.155
  - Panel (b) decomposition (3-month FX Change shared with 24m and residual):
    - FX Change (3m) shared with 24m:
      - Dealer Banks: −3.308***
      - Corporates: −0.159***
      - Funds: 1.493***
      - NBFIs: −0.376***
      - Non-Dealer Banks: 2.365***
    - FX Change (3m) residual:
      - Dealer Banks: 1.063***
      - Corporates: 0.127***
      - Funds: −0.636***
      - NBFIs: 0.118***
      - Non-Dealer Banks: −0.681**
  - Interpretation:
    - No net predictability of monthly spot flows from raw 3-month forecasts because the residual component predicts flows with opposite sign to the shared component; long-horizon forecasts carry predictive content for quantities.

- Funding flows (swap markets) and CIP (Table 4):
  - Definition: CIP basis = i_us − (i_foreign − FD_annualized); CIP>0 implies synthetic dollar funding cheaper than cash; basis mostly negative in sample.
  - CIP forecasts: use three-month-ahead forecasts of US and foreign three-month cash rates and three-to-six-month forward discount to compute expected CIP deviation.
  - Table 4, panel (a) CIP Deviation coefficients (swap flows, in billions of USD):
    - Dealer Banks: 130.899***
    - Corporates: 0.016
    - Funds: −55.198***
    - NBFIs: 1.259
    - Non-Dealer Banks: −85.192**
  - Interpretation:
    - Funds: negative coefficient −55.198*** → additional 55 billion (bn) USD synthetic borrowing for a 1 percentage point (pp) expected tightening of the CIP basis (i.e., basis becoming less negative / synthetic funding cheaper); CLS records second leg so negative swap flow = net USD borrowing.
    - Non-dealer banks: negative and large (−85.192**) → increased synthetic USD borrowing when synthetic funding becomes cheaper.
    - Dealer banks: positive 130.899*** → supply dollars in response to expected CIP easing, offsetting most investor demand (131 bn USD reported as near full accommodation).
  - Table 4, panel (b) decomposed regressors:
    - FX Change coefficients: Dealer Banks −3.449***; Corporates −0.0020; Funds 0.429**; NBFIs 0.003; Non-Dealer Banks 3.037***
    - Interest Rates coefficients: Dealer Banks 11.402; Corporates 0.123; Funds 12.669***; NBFIs 0.974**; Non-Dealer Banks −22.627***
  - Interpretation:
    - Expected USD appreciation (FX Change) associated with a widening (more negative) CIP basis → raises synthetic funding cost and reduces net USD borrowing by funds and non-dealer banks.
    - Responses to expected interest rate differentials vary across investor types (e.g., funds increase USD lending following US rate rises).

### Heterogeneity across currencies
- Purpose and specification:
  - Test whether predicted currency flows propagate through exchange rate factor structure (dollar, carry, UMVE betas).
  - Regression: y_i,t+1 = μ_i + α_t + β′X_i,t + φ(X_i,t × Z_i,t) + δ′Z_i,t + ε_i,t (Eq. (25)); interaction φ identifies cross-sectional differences in sensitivity.

- Directional flows and factor interactions (Table 5):
  - Panel (a) FX Change and Carry Beta (spot flows, in billions of USD):
    - FX Change coefficients: Dealer Banks −0.495**; Corporates −0.038**; Funds 0.288***; NBFIs −0.089***; Non-Dealer Banks 0.339*
    - FX Change × Carry Beta coefficients:
      - Dealer Banks: −0.819***
      - Corporates: −0.033***
      - Funds: 0.362***
      - NBFIs: 0.073***
      - Non-Dealer Banks: 0.406***
  - Panel (b) FX Change and Dollar Beta:
    - FX Change × Dollar Beta:
      - Dealer Banks: −0.477***
      - Corporates: −0.011*
      - Funds: 0.231***
      - NBFIs: 0.037***
      - Non-Dealer Banks: 0.216***
  - Panel (c) FX Change and UMVE Beta:
    - FX Change × UMVE Beta:
      - Dealer Banks: −0.109
      - Corporates: −0.061**
      - Funds: 0.198
      - NBFIs: 0.120***
      - Non-Dealer Banks: −0.148
  - Interpretation:
    - Funds, NBFIs, non-dealer banks: larger dollar purchases (or larger outflows from foreign currency) when expected USD appreciation is concentrated in currencies with higher carry, dollar, or UMVE betas — i.e., investors rebalance from riskier to safer currencies along traded-risk-factor exposures.
    - Dealer banks provide liquidity by selling dollars, with interactions often negative.

- Funding flows and factor interactions (Table 6):
  - Panels (a)–(c) interact forecasted CIP deviations with carry, dollar, and UMVE betas (swap flows, in billions of USD).
  - Panel (a) CIP Deviation × Carry Beta:
    - CIP Deviation coefficients: Dealer Banks 135.286***; Corporates 0.065; Funds −57.435***; NBFIs 1.276; Non-Dealer Banks −87.299***
    - CIP Deviation × Carry Beta:
      - Dealer Banks: −136.356**
      - Corporates: −1.692***
      - Funds: 74.283***
      - NBFIs: −0.069
      - Non-Dealer Banks: 59.929
  - Panel (b) CIP Deviation × Dollar Beta:
    - CIP Deviation × Dollar Beta:
      - Dealer Banks: 42.106*
      - Corporates: 0.083
      - Funds: −3.079
      - NBFIs: −0.407
      - Non-Dealer Banks: −41.185*
  - Panel (c) CIP Deviation × UMVE Beta:
    - CIP Deviation × UMVE Beta:
      - Dealer Banks: 52.869
      - Corporates: −1.127
      - Funds: 96.858***
      - NBFIs: 0.829
      - Non-Dealer Banks: −140.712*
  - Interpretation:
    - Funds: positive and significant interaction coefficients with carry and UMVE betas → when synthetic USD funding is expected to become more expensive (CIP forecast more negative), funds reduce synthetic USD borrowing more strongly via low-risk currencies than via high-risk currencies.
    - Dealer banks accommodate demand and provide more synthetic USD funding in low-risk currencies where sensitivity strongest.
    - No consistent interaction with dollar beta for swap flows, consistent with dollar factor capturing common return variation rather than cross-sectional funding differences.

### Macroeconomic forecasts and forecaster disagreement
- Macroeconomic forecasts (Table 7):
  - Regress spot flows on FX Change (24m), forecasted interest rate differential in 12 months (“Interest Rates”), forecasted CA / GDP differential, Inflation differential, and GDP growth differential (all expressed US minus foreign).
  - Table 7 coefficients (spot flows, in billions of USD):
    - FX Change: Dealer Banks −0.747***; Corporates −0.0190; Funds 0.432***; NBFIs −0.065***; Non-Dealer Banks 0.385**
    - Interest Rates: 0.181; 0.138*; 0.781; 0.151*; −1.391*
    - CA / GDP: 0.611***; 0.017; −0.157*; −0.100***; −0.343**
    - Inflation: −0.442; 0.313***; 1.237*; −0.148; −0.846
    - GDP: 1.673**; 0.625***; −1.512***; −0.077; −0.644
  - Interpretation:
    - Exchange rate forecast retains significance.
    - CA / GDP differential has strong predictive power: an increase in CA / GDP differential (US deficit smaller relative to abroad) predicts more USD purchases by institutional investors (NBFIs, non-dealer banks). Example: a one percentage point expected decrease in the CA / GDP differential is associated with non-dealer banks purchasing an additional 0.34 billion USD over the next month.
    - Corporates respond to expected increases in US inflation and GDP growth by buying USD, consistent with trade-related hedging.

- Forecaster disagreement (Table 8):
  - Construct disagreement measures: compute difference between 90th and 10th percentiles of forecasters; extract first principal component separately for US and foreign country and define High Disagr dummy = 1 when first PC exceeds sample median. Interact macro forecast differentials with High Disagr.
  - Key coefficients and interactions (spot flows, in billions of USD):
    - FX Change: Dealer Banks −0.694***; Corporates −0.0230; Funds 0.415***; NBFIs −0.062***; Non-Dealer Banks 0.349*
    - Interest Rates: 0.008; −0.049; 1.509*; 0.142; −1.696*
    - Interest Rates × High Disagr (=1): 0.529; 0.205***; −0.927**; 0.020; 0.111
    - CA / GDP: 0.966***; 0.023; −0.243**; −0.108***; −0.609***
    - CA / GDP × High Disagr (=1): −0.543***; −0.005; 0.199***; 0.008; 0.339***
    - Inflation: 1.037; 0.340**; 1.900**; −0.344**; −2.769**
    - Inflation × High Disagr (=1): −1.972**; −0.037; −0.538; 0.228; 2.259***
    - GDP: −0.148; 0.813***; −2.844***; 0.052; 2.240**
    - GDP × High Disagr (=1): 2.135; −0.219; 1.689*; −0.147; −3.536***
  - Interpretation:
    - During periods of high forecaster disagreement, the predictive relations between macro forecasts and currency flows are substantially muted or reversed for many investor groups (funds, NBFIs, non-dealer banks), implying reliance on consensus forecasts in low-uncertainty regimes.
    - When pooling across regimes, opposing effects across low- and high-disagreement periods can offset, reducing average predictive power.
  - Additional results:
    - Forecaster-level regressions: forecasts by forecasters with larger past errors have less predictive power for investor currency flows (details in Online Appendix E).
    - Global uncertainty (High VIX) yields qualitatively similar results to forecaster disagreement: predictability of currency flows from financial and macro expectations is weaker during times of high VIX (results in Online Appendix F, Table F.1).

*Source: wpiea2026186-source-pdf - 4.  Empirical Analysis (PDF).*

### 5.  Robustness

### 5. Robustness

### Forecast revisions and reverse causality
- Two additional exercises assess whether survey expectations predict currency flows (see Section G and Section H).
- Investigate whether changes in survey expectations contain predictive information beyond the level of expectations:
  - For both exchange rate and forecasted CIP deviations, forecast revisions only exhibit incremental predictive power for the currency flows of some groups of market participants.
  - Interpretation: some investors respond not only to the prevailing level of expected returns and funding conditions but also to meaningful updates of these expectations.
  - Horse-racing forecast revisions against the level of survey forecasts: the level of the survey forecast is the dominant predictor of currency flows, suggesting trading decisions are primarily predicted by the expected future environment rather than by recent forecast updates.
- Direction of predictive causality using Granger-causality tests:
  - Survey expectations robustly predict subsequent currency flows.
  - Past currency flows do not predict subsequent revisions in survey forecasts, consistent with the limited availability of CLS data.

### Alternative empirical specifications
- A range of robustness checks confirm baseline results across alternative empirical specifications (see Section I):
  - Alternative fixed effects structures, including specifications that only include time fixed effects (which control only for variation in unobservable global factors), yield qualitatively similar results relative to the baseline.
  - Restricting the cross-section of currencies to align with prior work (e.g., Kremens et al., 2025) produces consistent results in terms of both statistical and economic significance.
  - Results are robust to controlling for countries’ invoicing shares in international trade.
  - Findings are also present in univariate regressions.
- Overall conclusion from these exercises: survey expectations contain predictive information for currency flows that is not driven by alternative controls or model specifications.

### Key takeaways (from robustness analyses)
- The prevailing level of survey forecasts is the primary predictor of currency flows across instruments and participants.
- Forecast revisions matter incrementally for some participant groups, indicating heterogeneity in responsiveness to updates versus levels.
- Directional causality runs from survey expectations to currency flows, not the reverse, per Granger-causality tests.
- Results are robust across fixed effects choices, sample restrictions, controls for invoicing shares, and univariate specifications.

*Demand for Dollars: Evidence from Survey Expectations Working Paper No. WP/2026/186*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026186-source-pdf.pdf_