Algorithmic fund managers can amplify U.S. monetary shocks in emerging markets, but only when they herd on similar models; preserving model diversity, not just shrinking non-bank intermediation, is the policy priority to reduce spillovers.

Main Finding

Algorithmic (AI-driven) fund management alters how U.S. monetary policy is transmitted to emerging markets primarily through the similarity of models across funds. When algorithmic funds produce correlated forecast errors (high algorithmic homogeneity, ϕ), their trades reinforce one another and amplify capital outflows following U.S. monetary tightening—but this amplification occurs only in high-volatility (stress) regimes. In tranquil regimes algorithmic funds can stabilize flows if their models are diverse. Empirically, the herding dimension of algorithmic behavior amplifies equity portfolio outflows to 19 emerging markets (2000–2024) conditional on a VIX-defined stress regime; the amplification is robust in sign across many specifications but statistical significance depends on key design choices (stress definition and inclusion of 2008–09).

Key Points

• Mechanism

  • The paper introduces algorithmic homogeneity ϕ ∈ [0,1]: the fraction of forecast-error variance that is common across algorithmic funds.
  • Correlated forecast errors survive aggregation (via an exact law of large numbers for a continuum) and produce synchronized lending changes across algorithmic intermediaries.
  • Synchronized withdrawal of lending raises the external finance premium (µ), tightening credit and amplifying the macroeconomic impact of U.S. monetary shocks.
  • Heterogeneity (low ϕ) causes idiosyncratic errors to cancel in the aggregate; in the limit of full diversity algorithmic funds act as perfectly informed, stabilizing intermediaries.
  • The threshold of ϕ at which amplification arises is state-dependent: when local fundamentals deteriorate (signals become noisier), the threshold falls and amplification becomes more likely—so the same algorithmic sector can stabilize in normal times but amplify in crises.

• Theoretical contribution

  • Develops a two-region DSGE (U.S. center and representative emerging-market periphery) that embeds a continuum of algorithmic funds and a traditional, slow-intermediary sector.
  • Pinpoints a single primitive (forecast-error correlation ϕ) as the driver of state-dependent amplification, distinct from balance-sheet or capital-constraint channels in prior macro-finance work.
  • Clarifies that homogeneity operates through aggregation of errors rather than through changing signal weights or strategic coordination.

• Empirical finding & caveats

  • Uses equity portfolio flows to 19 emerging markets (2000–2024) and a triple-interaction regression that lets the effect of U.S. monetary shocks vary with algorithmic herding and a stress-regime indicator.
  • Finds amplification of outflows associated with the herding dimension of algorithmic behavior only during high-volatility (VIX-defined) episodes.
  • The negative amplification coefficient is consistently present across an extensive robustness battery; statistical significance, however, relies on (i) defining the stress regime by the VIX (rather than calendar-year crisis dummies) and (ii) retaining the 2008–09 crisis—the most informative stress episode. Excluding that episode weakens significance.

Data & Methods

• Model

  • Two-region DSGE with:
    • U.S. center that sets monetary policy (monetary shocks drive a global factor).
    • Periphery with households, firms, and two types of intermediaries: traditional (slow) funds and a continuum of algorithmic funds.
  • Algorithmic funds observe noisy signals about U.S. policy and local fundamentals; noise decomposes into common and idiosyncratic components so that ϕ is the fraction of signal-error variance shared across funds.
  • External finance premium µ(L) rises with lending volume L (µt = μ̄ (Lt/ L̄)^ξ), so coordinated lending moves map into non-linear changes in loan costs and investment.
  • State dependence is introduced via signal precision: local-signal variance increases as fundamentals deviate from steady state, lowering the homogeneity threshold for amplification.

• Empirical strategy

  • Outcome: equity portfolio flows to 19 emerging markets (2000–2024).
  • Key regressors: U.S. monetary shocks; an empirical proxy for algorithmic herding (the paper’s “herding dimension” of algorithmic behavior); and a stress-regime indicator (VIX high vs low), interacted in a triple-interaction framework so the monetary-shock effect can vary with both herding and regime.
  • Identification notes: the design leverages variation in U.S. monetary surprises and cross-country/time variation in the herding proxy and stress indicator. Inference uses schemes appropriate for panel/time-series dependence.
  • Robustness: extensive battery reported; authors are explicit that significance depends on defining stress via VIX and on inclusion of the 2008–09 crisis quarter observations.

• Measurement caveats

  • The draft emphasizes a behavioral/herding dimension of algorithmic activity as the empirical proxy for ϕ. The summary here follows the paper’s framing that the empirical herding measure captures the cross-sectional correlation dimension of algorithmic funds’ behavior; details of construction are in the full paper. Results are conditionally sensitive to how stress episodes are defined and to the inclusion of extreme crisis episodes.

Implications for AI Economics

• For policy and macroprudential design

  • Focus policy on preserving model diversity and reducing informational/model overlap across large non-bank intermediaries, rather than only constraining the size of non-bank intermediation.
  • Supervisory attention should include monitoring of model-similarity indicators, overlap in training data pipelines, and common-signal exposures that could generate correlated forecast errors.
  • Stress-testing frameworks should incorporate state-dependent effects of algorithmic homogeneity: algorithms may be stabilizing in normal times but become amplifiers under stress.
  • Disclosure and microprudential tools (e.g., mandated reporting of model features, data sources, or common factor exposures; countercyclical buffers for correlated activity) could mitigate the systemic amplification channel.

• For research in AI and macro-finance

  • Empirically measuring algorithmic homogeneity (forecast-error correlation) is crucial; cross-sectional and time-varying measures of model overlap should be developed and validated.
  • Future work should integrate the correlation-of-errors channel with balance-sheet amplification channels (capital constraints, VaR-based procyclicality) to study potential compounding effects.
  • The state-dependent sign reversal (stabilizing → amplifying) highlights the need for regime-aware models and for inference strategies that leverage stress episodes while accounting for their limited number.
  • Comparing non-strategic homogeneity (correlated errors from shared training/inputs) with strategic coordination (e.g., RL-based collusion) is an important next step—both mechanisms can coexist and may reinforce systemic risk.

Summary judgment - The paper provides a clear micro-founded mechanism linking algorithmic model similarity to international monetary transmission, and it produces a sharp, testable hypothesis that algorithms’ systemic effect is regime-dependent. Empirical evidence supports the mechanism qualitatively and in sign; statistical significance is credible under the baseline design but depends on sensible (and documented) choices about regime definition and crisis inclusion. The policy message is actionable: reduce model monoculture and monitor behavioral correlations across large non-bank intermediaries.