# Who Restores the Peg? A Mean-Field Game Approach to Model Stablecoin Market Dynamics

Crypto.world

> Fiat-collateralized stablecoins such as USDC and USDT are meant to trade at $1, and two channels are supposed to keep them there: a primary market where authorized arbitrageurs mint and redeem…

**arXiv**: 2601.18991

**Published**: 2026-01-26

**Authors**: Hardhik Mohanty, Bhaskar Krishnamachari

## AI summary

## Overview

**Research area:** Trading and market microstructure — specifically the economics of fiat-collateralized stablecoins (USDC, USDT), peg-restoring arbitrage, and systemic risk in digital asset markets.

**Technical level:** Intermediate. The paper builds a mean-field game with linear-quadratic control, a GARCH volatility process, and a numerical fixed-point solver, but the underlying economic story — arbitrageurs redeem at par in the primary market while retail traders buy and sell on exchanges — is explained in accessible terms.

**Scope:** The paper develops and empirically calibrates a dynamic mean-field game model of retail traders and arbitrageurs interacting across primary and secondary stablecoin markets, then uses it to attribute peg-reverting flows by channel and to locate a non-linear threshold in primary-market friction beyond which recovery slows sharply.

## What This Paper Is About

Fiat-collateralized stablecoins such as USDC and USDT are meant to trade at $1, and two channels are supposed to keep them there: a primary market where authorized arbitrageurs mint and redeem directly with the issuer at par, and secondary markets (centralized and decentralized exchanges) where anyone can trade. During stress episodes these coins de-peg, and it has not been clear which participants and which channel actually do the work of pulling the price back. The paper asks directly "who restores the peg?" and builds an equilibrium model that reproduces three historical de-peg episodes, decomposes the restoring flow into primary-redemption versus secondary-buying components, and identifies where the mechanism breaks down.

## Key Contributions

1. **A dynamic, equilibrium-based agent model for stablecoin de-pegs.** The authors formulate a mean-field game in which a continuum of retail traders and arbitrageurs solve forward-looking linear-quadratic control problems over mispricing, inventory, execution costs and congestion. They state this is the first study to analyze de-peg episodes through an equilibrium representation of dynamic strategic interaction between these populations across both primary and secondary markets.
2. **Calibration to three historical de-peg events.** The framework is fitted to the March 2023 USDC de-peg, the May 2022 USDT de-peg linked to the TerraUSD collapse, and the July 2023 localized USDT dislocation on Binance, using one-minute Binance Kline data. The authors report that the fitted model reproduces observed price trajectories and recovery half-lives.
3. **An endogenous order-flow decomposition by channel.** Because flows are derived from optimizing behavior rather than assumed, the model attributes peg-reverting pressure to primary-market mint/redeem flow versus secondary-market buying, event by event.
4. **A quantified non-linear stability threshold for primary-market access.** Sensitivity analysis identifies a critical range of primary redemption friction beyond which de-peg half-life rises sharply, with secondary-market liquidity acting mainly as an amplifier once primary bottlenecks are crossed.

## Main Findings

- **Primary-market arbitrage is the dominant stabilizer in most episodes.** Across the three calibrated events the authors report that system-wide stress is predominantly stabilized by primary-market arbitrage, though the exact split is event-specific.
- **The restoring channel differs by event type.** In the May 2022 USDT episode, stabilization is overwhelmingly driven by primary-market redemptions, consistent with a functional redemption rail. In the March 2023 USDC episode, the primary channel initially contributes negatively because minting and redemptions were impaired, and recovery reflects a joint role of primary and secondary forces once the rails were restored. In the July 2023 USDT venue-level failure, both channels contribute little, consistent with an outage that simultaneously blocks primary arbitrage and fragments secondary liquidity.
- **The model reproduces recovery half-lives.** For May 2022, both data and model agree on a rapid, sub-3-hour recovery. For March 2023, the model estimates a half-day recovery. For July 2023, the model predicts a 45.9-hour half-life versus a 30.7-hour empirical value, correctly identifying a slow recovery.
- **Lower average RMSE than reduced-form baselines.** Averaged over the three events, the dynamic MFG achieves an average RMSE of 0.0128, against 0.0158 for AR(1), 0.0162 for ARMA-GARCH, and 0.0217 for ARX. Its average half-life error of 0.41 is lower than ARX (1.15) and below the reduced-form baselines taken collectively, though AR(1) at 0.36 and ARMA-GARCH at 0.24 have smaller half-life errors in isolation.
- **Out-of-sample performance is stable across train/test splits.** On the March 2023 USDC event, test RMSE stays in a narrow band around 7×10⁻³ and test MAE around 6×10⁻³ for 70/30, 80/20 and 90/10 splits, while in-sample errors decline monotonically as the training share grows (train RMSE 0.002809, 0.002287 and 0.002029 respectively).
- **Primary friction is the binding constraint.** When primary friction is inexpensive (roughly κ_P ≤ 10), de-peg half-life stays short at about 2–4 model steps across a wide range of secondary frictions κ_R and retail shares π_R. Once κ_P rises past a critical range of roughly 15–18, half-life increases sharply.
- **Secondary illiquidity matters mainly in the impaired-primary region.** Holding primary frictions at a moderately stressed level, recovery slows monotonically as the secondary liquidity-depth scale λ_scale increases (shallower books), with half-life exceeding 30 steps in the thinnest regimes. The slowdown is most pronounced at intermediate retail shares, roughly π_R in [0.4, 0.7].
- **Equilibria are tight ε-Nash.** Final arbitrageur exploitability is around 10⁻⁴ in the Stable regime, 10⁻³ in the De-peg regime, and a few 10⁻² in the Recovery regime. Maximum converged exploitability stays below a few 10⁻² even at the largest congestion coupling ξ and decays toward 10⁻⁴ as coupling weakens.
- **Calibrated frictions move in an interpretable pattern.** Calibrated κ_P is highest in stressed regimes and declines toward the stable regime; secondary frictions κ_R and κ_A tend to peak during recovery when order books are thinnest; venue-impact terms λ_s rise in stressed regimes.

## Methodology in Plain English

The authors treat the stablecoin market as two large groups of rational traders whose individual actions are too numerous to simulate one by one. Instead of modeling each trader, they model the aggregate market state — how far the price is from $1, how backed up the redemption queues are on each blockchain, and how much flow is going through each exchange and each mint/redeem channel — and let representative agents of each type choose trading strategies that are optimal given that aggregate state. This is the mean-field game trick: replacing a hugely complex many-player game with a tractable fixed-point problem.

Retail traders, who make up 85% of the population, can only trade on secondary venues and face higher execution frictions and inventory aversion. Arbitrageurs, the remaining 15%, face lower frictions and additionally have access to primary mint/redeem channels, paying a routing cost that depends on how congested each channel is. Both types minimize a discounted sum of costs covering trading profit and loss from mispricing, inventory risk, quadratic execution costs, and congestion costs that grow with aggregate flow.

Market prices then evolve mechanically from the aggregate flows, with each venue having its own price-impact coefficient and each blockchain channel its own coupling to the secondary price. Redemption backlogs build up from arbitrageur submissions and decay at a per-channel processing rate. To capture time-varying stress, volatility follows a GARCH process and feeds back into execution frictions, inventory aversion and venue price impact, so trading gets more expensive exactly when markets are turbulent.

The equilibrium is computed by policy iteration: solve each population's best response to the current aggregate state, regenerate the aggregate path from those policies, and repeat until the path stops moving and no agent can profitably deviate. Exploitability — the fractional improvement available from a unilateral deviation — is computed explicitly to certify the ε-Nash property. For empirical work, each de-peg event is split into de-peg, recovery and stable phases, and parameters are fitted by minimizing mean squared error against observed mispricing using Differential Evolution, a population-based global optimizer chosen for robustness to local minima. Data are one-minute candlestick series from Binance's public spot API: USDC/USD from March 11–15, 2023; USDT/USD from May 12–16, 2022; and USDT/USD from July 15–31, 2023. Empirical half-lives are estimated with an AR(1) fit and the standard ln(2)/(−ln ρ) formula.

## Why This Matters

**Impact on research.** Prior work either infers reversion from reduced-form regressions on empirical data or compares stablecoin designs in static equilibrium. This paper shows that a dynamic equilibrium model with explicit frictions can reproduce observed recovery dynamics while also producing counterfactuals — the threshold and channel decomposition are not recoverable from a regression. It also brings mean-field game machinery, previously applied to bank runs and optimal execution, into stablecoin market microstructure.

**Real-world applications:**

- **Issuer design of mint/redeem rails.** The κ_P threshold gives issuers a concrete quantity to monitor: accessibility, fees, settlement speed and operational resilience of primary redemption matter more for peg robustness than reserve quality alone.
- **Regulatory stress testing.** Regulators could use the calibrated friction parameters as indicators of how quickly a stablecoin would recover under a given shock, and where secondary liquidity stops helping.
- **Risk management for DeFi protocols.** Lending and derivatives protocols accepting stablecoins as collateral can use the half-life estimates and the retail-share sensitivity to size haircuts during stress.
- **Distinguishing systemic from localized events.** The July 2023 case shows the model can identify an event where neither channel contributes meaningfully, which is diagnostically different from a run on reserves.

**Industry relevance.** The results reframe peg stability as an operational property of redemption infrastructure rather than purely a balance-sheet property, which is directly actionable for issuers, exchanges and market makers who control those rails.

## Future Directions

- **Richer agent heterogeneity.** The authors acknowledge that two representative populations in a linear-quadratic setting simplify heterogeneity among liquidity providers and can understate non-linear behavior under extreme stress, including binding constraints and discontinuous primary-rail access. Expanding the population types is a natural next step.
- **Endogenizing volatility and liquidity regimes.** The conclusion notes that the model treats volatility and liquidity as regime-based (the sentence is truncated in the provided content), pointing toward making these state variables endogenous to agent behavior rather than exogenous inputs.
- **More de-peg episodes and asset types.** The calibration covers three events on two coins; extending to other fiat-backed coins, algorithmic designs and additional historical episodes would test how general the κ_P threshold is.
- **Using the threshold as a policy and design target.** Since the model yields an operational criterion for peg robustness, an open question is how specific rail designs — settlement latency, fee schedules, access controls such as KYC or banking limits — map onto the calibrated κ_P scale.

## Target Audience

Researchers in market microstructure and digital asset economics; stablecoin issuers and treasury/operations teams responsible for mint and redeem infrastructure; risk managers at exchanges, market makers and DeFi lending protocols that hold or accept stablecoins as collateral; and regulators working on stablecoin reserve, redemption and systemic-risk standards. Readers without a control-theory background will find the economic narrative and empirical results accessible, while the formal model and solver details will mainly interest those comfortable with dynamic optimization and equilibrium computation.

## Authors’ abstract

USDC and USDT are the dominant stablecoins pegged to \$1 with a total market capitalization of over \$300B and rising. Stablecoins make dollar value globally accessible with secure transfer and settlement. Yet in practice, these stablecoins experience periods of stress and de-pegging from their \$1 target, posing significant systemic risks. The behavior of market participants during these stress events and the collective actions that either restore or break the peg are not…

[The full abstract on arXiv](https://arxiv.org/abs/2601.18991)

[Read the original paper](https://arxiv.org/abs/2601.18991)

---

Canonical: https://crypto.world/paper/who-restores-the-peg-a-mean-field-game-approach-to-model-stablecoin-2601.18991
