---
title: Impermanent Loss in DeFi and ML
url: https://www.emergentmind.com/topics/impermanent
type: topic
---

# Impermanent Loss in DeFi and ML

Impermanent

Impermanent loss, often abbreviated as IL, is a central concept in the study of decentralized finance (DeFi), automated market makers (AMMs), and time-series modeling benchmarks. The term is used in both financial engineering—describing a liquidity provider’s opportunity cost due to pool rebalancing—and, more recently, as a brand for dynamic temporal benchmarks in machine learning. This article focuses on the technical foundations and state-of-the-art research on impermanent loss, its statistical characterization, mitigation techniques, hedging approaches, and the emergence of adaptive evaluation protocols under the "impermanent" paradigm.

## 1. Mathematical Definition and General Framework

Impermanent loss measures the shortfall of a liquidity provider’s returns relative to a passive holding strategy under price divergence in AMMs. In the standard two-asset constant-product AMM (e.g., Uniswap V2), the reserves at time $t$ are $(R_A^t, R_B^t)$. If an LP deposits a portfolio worth $W_1 = R_A^1 + p_m^1 R_B^1$ at time 1 and withdraws at time $T$ when the reference price is $p_m^T$, the withdrawal value is $W_T^{\rm LP} = R_A^T + p_m^T R_B^T$, while passively holding yields $W_T^{\rm Ref} = R_A^1 + p_m^T R_B^1$. Ignoring trading fees, impermanent loss is defined as

\[
IL = \frac{W_T^{\rm LP}}{W_T^{\rm Ref}} - 1 = \frac{2 \sqrt{p_m^T / p_m^1}}{1 + (p_m^T / p_m^1)} - 1,
\]

in which a positive IL implies underperformance versus the hold benchmark [2401.07689].

Key generalizations are provided for $N$-asset pools governed by a constant function market maker (CFMM) invariant $F(q_1, ..., q_N; \zeta) = K$, yielding

\[
IL_j = \sum_{i=1}^N \left(q_i^t - q_i^T\right)Z_{i,j}^T,
\]

where $q^T$ are terminal reserves and $Z_{i,j}^T$ are terminal shadow prices [2301.06831]. Weighted-mean and general convex invariants yield more complex, often non-univariate IL surfaces, with geometric mean market makers (G3Ms) providing minimal parameterizations via exchange-rate ratios [2203.11352].

## 2. Statistical Properties, Path Dependency, and the IL–LVR Relationship

Impermanent loss is fundamentally path-independent, depending solely on initial and terminal prices, which can be contrasted with the path-dependent loss-versus-rebalancing (LVR) metric. For small price movements (very short times), the increment in IL and LVR coincide:

\[
\Delta IL = \Delta LVR = \frac{L}{\sqrt{p}}(1-\sqrt{p/(p+\Delta p)})^2.
\]

Over intermediate times ($0 \ll \sigma^2 T < 1$), both metrics have equal expectations (driven by the central limit theorem):

\[
\mathbb{E}[IL(T)] = \mathbb{E}[LVR(T)] = \frac{L}{4\sqrt{p_0}}\sigma^2 T.
\]

However, their distributions diverge: IL is highly skewed with most mass near zero (many price paths return close to start), while LVR accumulation yields a Gaussian-like density [2502.04097, 2410.00854]. At long time scales or high volatility, the equality breaks down, and the full distribution must be considered for risk management.

## 3. Effect of Fees, Arbitrage, and Dynamic AMM Mechanisms

Fee income is the only mechanism to counteract impermanent loss in AMMs. Under realistic on-chain turnover (e.g., Uniswap V2 WETH/USDC at $2\%$ per day), accrued trading and arbitrage fees regularly surpass rebalancing losses for all but extreme price moves (e.g., $p_m^T/p_m^1 \in [0.25, 4]$ corresponding to $-75\%$ to $+300\%$) [2401.07689]. The paper demonstrates that arbitrage-friendly environments, with low transaction costs for arbitrageurs, maximize fee income and stabilize LP returns. Block-adaptive and deal-adaptive dynamic fee schedules further reduce IL by 1–4% under normal conditions and up to 35% in high-volatility regimes, outperforming fixed-fee structures [2506.03001]. Alternative AMM architectures, such as those based on power-law invariants ($X^n Y = K$ for $n=4$), achieve up to 36% lower IL compared to constant-product models in simulations, especially when paired with dynamic rebate mechanisms [2502.20001].

## 4. Hedging Impermanent Loss: Option Replication and Delta-Neutral Strategies

Multiple works establish that impermanent loss in constant-product AMMs can be statically hedged using a "strip" of European options. Explicitly, for concentrated positions on Uniswap v3 over $[P_\ell, P_u]$, the expected IL is

\[
\mathbb{E}^{\mathbb{Q}}\left[\mathrm{UIL}^R(P_t)\right] = -\frac{1}{2}\int_{P_\ell}^{P_u} K^{-3/2} C(t, K) dK,
\]

where $C(t, K)$ denotes Black–Scholes call option prices [2205.12043, 2503.21967]. Coverage within a finite range is guaranteed by a suitable purchase of a "strangle"—a small position in out-of-the-money calls and puts—funded by LP fees. Model-free hedges are possible for G3Ms charging proportional fees by dynamically rebalancing the difference of reserves in the external market, resulting in a perfect super-hedge of IL when reserve processes are of finite variation [2303.11118]. Delta-hedging via a portfolio of vanilla derivatives can further neutralize price risk in both uniform and concentrated-liquidity pools [2208.03318].

## 5. Empirical Evaluation, Profitability, and Fee–IL Tradeoffs

Empirical studies on major pools reveal that a majority of LPs, especially passive or long-horizon providers, have suffered net losses: e.g., $199.3$M in fees versus $260.1$M in IL for $43\%$ of Uniswap v3 TVL, with only transient "flash" LPs realizing significant positive net returns [2111.09192]. Net profitability is dictated by the ratio of fee APR to the impermanent-loss APY, with break-even occurring only when cumulative fees offset deterministic rebalancing drag [2108.06593]. Profitability zones can be quantified analytically: for a given one-sided fee $\phi$ and price ratio $r$, LPs are profitable only for $r\in[1/T,\,T]$ with $T = \frac{(2-Y)^2}{(2Y-1)^2},\, Y = 1-\phi$ [2604.28014]. Fee selection can thus be used as a design lever, tuning the width of sustainable zones where both LPs and arbitrageurs are net-positive.

## 6. Extensions: Multi-Asset Pools, Power-Law and Proactive Market Makers

Impermanent loss generalizes to $N$-asset constant function market makers with parameterizable invariants. Geometric-mean and power-law invariants ($X^n Y = K$) allow for univariate, tractable IL formulas and enable fine control over loss curves across market regimes [2203.11352, 2502.20001]. Multi-token proactive market makers (such as DODO's PMM and its $N$-asset extension, MPMM) empirically achieve an order-of-magnitude reduction in median IL (95% vs PMM, 99.7% vs CPMM) and better capital efficiency by distributing loss across more tokens and adopting a flatter cost surface [2309.00632]. These architectures require careful oracle integration and on-chain computation to maintain hedging efficacy.

## 7. "Impermanent" in Temporal ML Evaluation

Separately, "Impermanent" has become a technical term in time-series ML evaluation, denoting a live, continuously-updating benchmark designed to assess the temporal generalization and robustness of forecasting models under distributional shift [2603.08707]. The Impermanent protocol enforces rolling-origin, embargoed scoring, and standardized pipelines on a large-scale, highly non-stationary data stream (e.g., GitHub event counts). Metrics include MASE and scaled CRPS; leaderboard tracking and fair evaluation protocols prevent test contamination. This live-benchmarking framework reframes the evaluation from static accuracy to long-horizon, sustained performance and response to dynamic shocks, making temporal robustness measurable.

---

Impermanent loss is an inexorable corollary of convexity and arbitrage in all AMMs with nontrivial rebalancing structure. Theoretically tractable, empirically quantifiable, and practically hedged, IL constitutes both a design constraint and a fundamental source of frictions in DeFi. Its measurement and minimization—through dynamic fees, novel invariants, hedging portfolios, and protocol-layer adaptivity—drives contemporary research at the intersection of financial engineering, decentralized systems, and applied stochastic analysis [2401.07689, 2203.11352, 2410.00854, 2502.04097, 2301.06831, 2309.00632, 2503.21967, 2604.28014, 2205.12043, 2502.20001, 2506.03001, 2303.11118, 2108.06593, 2111.09192, 2106.14404, 2302.11942, 2603.08707]. The extension of "impermanent" to dynamic benchmarking in time-series forecasting underscores its semantic migration into the technical lexicon of robustness assessment beyond finance.

Source: https://www.emergentmind.com/topics/impermanent