💰 quantitative finance

Algorithmic Intermediation and the International Transmission of U.S. Monetary Policy

This paper argues that the international transmission of U.S. monetary policy to emerging markets is destabilized not by algorithmic intermediation itself, but by the similarity of models across funds which causes correlated errors and herding, suggesting that policy should prioritize preserving model diversity over limiting non-bank intermediation.

Original authors: Fernando Toledo, Luis Dimotta Bré, Gabriel Montes-Rojas

Published 2026-07-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Fernando Toledo, Luis Dimotta Bré, Gabriel Montes-Rojas

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the global economy as a giant, bustling marketplace where money flows like water between countries. Sometimes, the water flows smoothly, filling up the fields of emerging markets (countries that are growing fast but aren't fully developed yet). Other times, the water suddenly drains away, leaving those fields dry and struggling. For a long time, economists thought this draining was caused by big, slow-moving giants: traditional banks and the slow, heavy decisions of human managers. But recently, a new type of player has entered the market: algorithms. These are computer programs, often powered by artificial intelligence, that make investment decisions in the blink of an eye.

The big question everyone is asking is: Are these super-fast computers saving the market by reacting instantly, or are they making it more dangerous by panicking together? To understand this, we need to know a few things. First, "monetary policy" is basically the central bank of a rich country (like the U.S.) turning a faucet on or off to control how much money is in the system. When they tighten the faucet, money gets expensive, and it often flows out of other countries. Second, "herding" is what happens when a flock of birds all turn at the exact same time; if they all make the same mistake, the whole flock crashes. This paper asks: when the U.S. tightens its money faucet, do these computer programs act like a calm, diverse flock that spreads out the risk, or do they act like a panicked stampede that makes the crisis worse?

The authors of this paper, Fernando Toledo, Luis Dimotta Bré, and Gabriel Montes-Rojas, decided to investigate this using a mix of computer simulations and real-world data. They built a mathematical model of the world with two main characters: the "Center" (the United States) and the "Periphery" (emerging markets). In their story, they introduced a special character called "Algorithmic Homogeneity." Think of this not as the speed of the computers, but as how much they are wearing the same "thinking hat." If every computer program is trained on the exact same data and uses the exact same logic, they are highly homogeneous. If they all use different data and different logic, they are diverse.

The paper's main discovery is a twist that flips our intuition. It suggests that having fast, AI-driven money managers isn't the problem in itself. The problem is when they all think exactly the same way. The authors found that when these algorithms are diverse, they actually help stabilize the market. It's like a classroom where every student has a different way of solving a math problem; if one student makes a mistake, another might get it right, and the class average stays accurate. In this scenario, the algorithms act as a stabilizer, smoothing out the flow of money even when the U.S. changes its policy.

However, the story changes dramatically when the market gets stressed. The paper shows that the algorithms don't necessarily change how similar they are; instead, the "safety line" for disaster drops. Even if the algorithms keep wearing the same "thinking hats" (maintaining the same level of similarity), the point at which that similarity becomes dangerous falls as the economy worsens. In calm times, the safety line is high, so even similar algorithms stay safe. But once stress hits, the safety line drops below the level of similarity, and the algorithms start herding. Instead of canceling each other out, their mistakes pile up. The authors found that this "herding" behavior amplifies the shock. When the U.S. tightens its money, these similar algorithms all decide to pull their money out of emerging markets at the exact same time, turning a small leak into a massive flood.

Crucially, the paper argues that this isn't just about how many algorithms there are or how fast they are. It's about their similarity. The authors ran simulations and looked at real data from 19 emerging markets between 2000 and 2024. They found that the amplification effect only happens during high-stress times, specifically when global fear is high (measured by something called the VIX, which is like a "fear gauge" for the stock market). In calm times, even if the algorithms are similar, they don't cause much trouble. But once the stress hits, the threshold for disaster drops, and the similar algorithms trigger a synchronized panic.

The paper also rules out a few common ideas. It explicitly states that the danger isn't the size of the non-bank financial sector (how much money these funds manage) or the speed of the trading. You could have a huge sector of fast traders, and if they are all thinking differently, they might actually be safe. The danger is specifically the "correlation of errors"—when they all get the same thing wrong at the same time.

So, what does this mean for the future? The authors suggest that regulators shouldn't just try to limit how big these funds can get or how fast they can trade. Instead, they should focus on keeping the "thinking hats" diverse. They propose that we need to measure how similar these algorithms are and encourage them to use different data and models. It's like telling a flock of birds to look in different directions so they don't all crash into the same tree. The paper concludes that while AI and algorithms are here to stay, their impact depends entirely on whether we let them become a monolithic, thinking-in-unison crowd or keep them as a diverse group of independent thinkers. The evidence suggests that without this diversity, the next time the U.S. turns the faucet, the algorithms might just make the water drain a lot faster than it should.

Technical Summary: Algorithmic Intermediation and the International Transmission of U.S. Monetary Policy

Problem and Motivation
This paper investigates whether the rise of algorithmic and AI-driven fund management alters the international transmission of U.S. monetary policy to emerging markets. While existing literature on the Global Financial Cycle (e.g., Rey, 2015; Miranda-Agrippino & Rey, 2020) documents that U.S. shocks drive a common factor in global asset prices, the specific micro-founded mechanism through which intermediation amplifies or dampens these shocks remains a "black box."

The authors argue that the source of instability is not the presence of algorithms per se, but the similarity of models across funds. When algorithms rely on identical signals and architectures, they produce correlated forecast errors. During periods of stress, these errors reinforce one another, intensifying capital-flow responses. Conversely, diverse models allow errors to offset, potentially stabilizing flows. This distinction is critical given the rapid growth of non-bank financial intermediation (NBFI), which reached 51% of global financial assets by 2024, with the fastest growth occurring in the hedge fund and investment vehicle sectors most likely to employ algorithmic strategies.

Methodology
The paper employs a two-pronged approach: a structural theoretical model and an empirical panel analysis.

Theoretical Framework (DSGE Model):
The authors develop a two-region (U.S. center and emerging market periphery) Dynamic Stochastic General Equilibrium (DSGE) model. The core innovation is the intermediation sector, which consists of:
- Traditional Funds: Slow-adjusting lenders subject to quadratic adjustment costs.
- Algorithmic Funds: A continuum of funds ( $j \in [0,1]$ ) that lend based on noisy signals regarding U.S. policy shocks ( $MP_t$ ) and local fundamentals ( $F_t$ ).
The model introduces a single parameter, algorithmic homogeneity ( $\phi \in [0,1]$ ), defined as the fraction of forecast-error variance shared across funds. Signals are decomposed into a common component (driven by shared training data/architecture) and an idiosyncratic component.

Crucially, the model incorporates state-dependent signal precision. Following Orlik & Veldkamp (2024), the precision of local signals deteriorates as fundamentals move away from their steady state (i.e., during deep recessions or overheating). This creates a mechanism where the threshold for amplification is endogenous to the economic state.
Empirical Strategy:
The authors test the model's central prediction using a panel of equity portfolio flows to 19 emerging markets from 2000Q1 to 2024Q4 ( $N=1,636$ ).
- Dependent Variable: Quarterly equity portfolio flows (IMF BPM6).
- Key Independent Variables:
  - Monetary Shock ( $r_t$ ): High-frequency U.S. monetary policy surprises (San Francisco Fed USMPD).
  - Algorithmic Behavior ( $A_{herding}$ ): A proxy for $\phi$ , constructed as the rolling 8-quarter correlation between a country's position changes and a global benchmark (BIS Locational Banking Statistics).
  - Stress Regime ( $Crisis_t$ ): A dummy variable equal to 1 when the VIX exceeds 30.
- Specification: A triple-interaction fixed-effects model:
  $Y_{it} = \alpha r_t + \beta A_{i,t-1} + \delta Crisis_t + \beta_4 (A \times Crisis) + \beta_5 (r \times A) + \beta_6 (r \times Crisis) + \gamma (r \times A \times Crisis) + \mu_i + \varepsilon_{it}$
  The coefficient of interest, $\gamma$ , captures the additional amplification of capital outflows during stress regimes driven by algorithmic herding.

Key Results

Theoretical Propositions:
- Proposition 1 (Amplification under Homogeneity): Homogeneity amplifies the variance of the aggregate lending response to a monetary shock, not the mean. Correlated errors survive aggregation, scaling the response volatility by $\sqrt{\phi}$ .
- Proposition 2 (Attenuation under Heterogeneity): As $\phi \to 0$ (diverse models), forecast errors cancel out, rendering the algorithmic sector a perfectly informed, stabilizing intermediary.
- Proposition 3 (State-Dependent Threshold): There exists a threshold $\phi^*$ separating attenuation from amplification. This threshold is state-dependent: it falls as local fundamentals deteriorate ( $\Delta F$ rises). Consequently, an algorithmic sector that stabilizes flows in tranquil times can amplify them in crises without any change in $\phi$ itself, simply because the threshold drops below the prevailing level of homogeneity.
Empirical Findings:
- State-Dependent Amplification: The triple-interaction coefficient $\gamma$ is negative and statistically significant (at the 5% level in the baseline). This confirms that algorithmic herding amplifies capital outflows following contractionary U.S. shocks, but only during high-volatility (stress) regimes.
- No Effect in Tranquil Times: The two-way interaction ( $r_t \times A_{herding}$ ) is insignificant in normal times, supporting the model's prediction that homogeneity does not amplify transmission when fundamentals are stable.
- Robustness: The negative sign of $\gamma$ holds across a battery of 12 specifications, including alternative clustering, outlier removal, and excluding specific countries (e.g., China) or crisis episodes (e.g., 2020 COVID).
- Limitations: The statistical significance of the result is sensitive to two design choices: (1) defining the stress regime via the VIX (continuous measure) rather than calendar years (step function), and (2) including the 2008–09 Global Financial Crisis. When the 2008–09 crisis is excluded, the coefficient remains negative and large but loses statistical significance, a result the authors attribute to the loss of the most informative stress episode for a state-dependent mechanism.

Significance and Contributions
The paper makes three primary contributions to the literature:

Micro-Founding the Global Financial Cycle: Unlike prior work that treats intermediation as a black box, this paper provides a micro-founded mechanism (the aggregation of correlated forecast errors) explaining why the strength of U.S. monetary transmission varies over time and across regimes.
Behavioral vs. Institutional Risk: The authors demonstrate that the risk of amplification lies in behavioral correlation (model homogeneity) rather than the institutional size of non-bank intermediaries. A large NBFI sector is not inherently fragile; it becomes so only when models converge.
State-Dependent Sign Reversal: Departing from literature that posits a monotonic relationship between algorithmic adoption and risk (e.g., Meng & Chen, 2026; Qiu & Han, 2026), this paper identifies a qualitative sign reversal. The same algorithmic sector stabilizes flows when fundamentals are healthy and amplifies them only when fundamentals deteriorate.

Policy Implications
The authors argue that current macroprudential frameworks, which focus on the size and concentration of institutions, are misaligned with the risks posed by algorithmic intermediation.

Regulation of Diversity: Policy should focus on preserving model diversity rather than limiting the size of non-bank intermediation. A regulator capping the size of the sector while leaving models correlated may inadvertently raise fragility by forcing consolidation onto common platforms.
Avoiding Standardization: Mandating common risk models or stress scenarios could be self-defeating, as it increases $\phi$ and the amplification factor.
Counter-Cyclical Management: Because the amplification threshold falls during crises, supervisors must evaluate model homogeneity against a state-contingent threshold. The optimal time to encourage heterogeneous models is during calm periods, before the incentive to imitate erodes diversity.

The paper concludes that the international financial system need not become more fragile due to AI; the outcome is conditional on model diversity. The policy priority is to make this diversity observable and defend it counter-cyclically.

More like this