Liquidity Withdrawal Index (LWI)

• Liquidity Withdrawal Index is a family of indicators that quantify liquidity erosion by capturing the rate at which order-book depth is removed or decays.
• Different formulations of LWI are employed across electronic limit order books, automated market makers, and payment systems, using metrics like cancellation ratios, mechanical withdrawal probabilities, and behavioral propensities.
• Empirical analyses show LWI’s effectiveness in detecting transient liquidity stress, forecasting market movements, and guiding liquidity provider exit decisions in complex trading environments.

Searching arXiv for the cited LWI-related works to ground the article in current papers. arxiv_search(query="Liquidity Withdrawal Index limit order book crumbling quotes Intrinsic Network liquidity withdraw machine learning", max_results=10) arxiv_search(query="Liquidity Withdrawal Index limit order book crumbling quotes Intrinsic Network liquidity withdraw machine learning", max_results=10) Liquidity Withdrawal Index (LWI) denotes a family of domain-specific indicators intended to measure the erosion, withdrawal, or stress of available liquidity. The term is not attached to a single canonical formula across the literature. In electronic limit order books, LWI has been defined either as a calibrated probability of mechanically driven quote erosion or as a bounded ratio of top-of-book cancellations to available and newly added depth (Xu et al., 23 Apr 2026, Haochuan et al., 26 Sep 2025). In event-based foreign-exchange microstructure, it operationalizes an information-theoretic Liquidity metric over an Intrinsic Network (Golub et al., 2014). In automated and concentrated-liquidity market makers, it quantifies either the incentive for a liquidity provider to exit or the systemic consequence of that provider’s withdrawal (Bergault et al., 8 Sep 2025, RajabiNekoo et al., 25 Jul 2025). In payment-system stress modeling, it aggregates withdrawal propensity and pressure induced by trust, scar, rumor, and threshold-gated behavior (Amouzgar, 18 Feb 2026). A plausible implication is that LWI is best treated as a class of withdrawal-pressure indices whose mathematical form depends on market design, observability, and the mechanism being isolated.

1. Scope of the term and principal formulations

The cited literature uses “Liquidity Withdrawal Index” to represent distinct but related constructs. Some formulations are directly observable flow ratios, some are model-implied probabilities, and some are normalized control or stress functionals. The common element is an attempt to summarize when liquidity is being removed faster than it is replenished, revealed, or behaviorally retained.

Domain	LWI definition	Primary inputs
Electronic LOB crumbling	$\mathrm{LWI}_t = 100 \cdot \hat{p}_t$	message-level LOB, adds, cancels, executions, mechanical gate
Nasdaq MBO top-of-book stress	$$\mathrm{LWI}_t = \dfrac{\text{Cancels}_t}{\mathrm{MA}_{1\text{s}}(\text{DepthL1})_{t-1} + \max(\text{Adds}_t,\epsilon)}$$	level-1 adds, cancels, depth
Intrinsic Network stress	$$\mathrm{LWI}_t = 1 - \Phi\!\left(\dfrac{\gamma_t^{(w)} - K_w H^{(1)}}{\sqrt{K_w H^{(2)}}}\right)$$	directional-change transitions, transition matrix
AMM LP exit	normalized $-v(t,y,S)$ or boundary proximity	QVI solution, oracle price, inventory, fees
CLMM LP systemic risk	$$\mathrm{LWI}_i(t) = \mathcal{N}(\mathrm{LSIS}_i(t)) \cdot \pi_i(t)$$	ETWL, counterfactual price impact, withdrawal likelihood
Payment-system stress	average propensity or balance-weighted pressure	outflows, rumor, scar, trust, broadcasts

An important source of ambiguity is orientation. In the Intrinsic Network formulation, low LWI flags stressed, illiquid conditions and high LWI indicates ample liquidity (Golub et al., 2014). In the other cited formulations, high LWI typically denotes elevated withdrawal risk, elevated cancellation pressure, or stronger incentive to exit (Xu et al., 23 Apr 2026, Haochuan et al., 26 Sep 2025, Bergault et al., 8 Sep 2025, RajabiNekoo et al., 25 Jul 2025, Amouzgar, 18 Feb 2026).

2. Central limit order books: quote erosion, cancellations, and effective liquidity

In the paper "When Quotes Crumble: Detecting Transient Mechanical Liquidity Erosion in Limit Order Books" (Xu et al., 23 Apr 2026), LWI is built from a mechanics-constrained detector for “crumbling quotes,” defined as episodes in which the best bid or ask deteriorates by multiple ticks because displayed top-of-book depth is removed faster than it is replenished. The paper distinguishes transient, mechanically induced liquidity withdrawal from informational repricing by enforcing efficient-price stability, opposite-side stability, and transience. Ground truth is created in ABIDES through stochastic regime switches in an adaptive market maker’s side-specific quoting skew, with side labels

$$Z_t^{(\mathrm{bid})} = \mathbf{1}\{\beta_t > 0.5 + \xi\}, \qquad Z_t^{(\mathrm{ask})} = \mathbf{1}\{\beta_t < 0.5 - \xi\}.$$

A detection pipeline then identifies deterioration steps, clusters them into candidate events, filters them using book consistency and price-stability constraints, computes severity and recovery features, and outputs a calibrated crumbling probability via a three-layer MLP with hidden dims $[64,32]$, LayerNorm, GELU, and dropout $0.1$. The event-level probability is

$$\hat{p}_t = \mathcal{F}_\eta(e(t)) \cdot \sigma\!\bigl(w^\top f_\phi(\tilde{x}(e(t))) + b\bigr),$$

and the resulting index is

$\mathrm{LWI}_t = 100 \cdot \hat{p}_t.$

Optional EWMA smoothing is

$$\mathrm{LWI}_t = \dfrac{\text{Cancels}_t}{\mathrm{MA}_{1\text{s}}(\text{DepthL1})_{t-1} + \max(\text{Adds}_t,\epsilon)}$$0

Experiments report Binary Rule AUC $$\mathrm{LWI}_t = \dfrac{\text{Cancels}_t}{\mathrm{MA}_{1\text{s}}(\text{DepthL1})_{t-1} + \max(\text{Adds}_t,\epsilon)}$$1, Logistic Regression AUC $$\mathrm{LWI}_t = \dfrac{\text{Cancels}_t}{\mathrm{MA}_{1\text{s}}(\text{DepthL1})_{t-1} + \max(\text{Adds}_t,\epsilon)}$$2, and MLP AUC $$\mathrm{LWI}_t = \dfrac{\text{Cancels}_t}{\mathrm{MA}_{1\text{s}}(\text{DepthL1})_{t-1} + \max(\text{Adds}_t,\epsilon)}$$3, with the abstract reporting a $$\mathrm{LWI}_t = \dfrac{\text{Cancels}_t}{\mathrm{MA}_{1\text{s}}(\text{DepthL1})_{t-1} + \max(\text{Adds}_t,\epsilon)}$$4 AUC improvement over rule-based baselines and robustness across normal, bull, bear, and high-volatility conditions (Xu et al., 23 Apr 2026).

A second LOB-specific formulation appears in "Forecasting Liquidity Withdraw with Machine Learning Models" (Haochuan et al., 26 Sep 2025), where LWI is defined as a bounded, interpretable measure of transient liquidity removal at the best quotes. The generic ratio over horizon $$\mathrm{LWI}_t = \dfrac{\text{Cancels}_t}{\mathrm{MA}_{1\text{s}}(\text{DepthL1})_{t-1} + \max(\text{Adds}_t,\epsilon)}$$5 is

$$\mathrm{LWI}_t = \dfrac{\text{Cancels}_t}{\mathrm{MA}_{1\text{s}}(\text{DepthL1})_{t-1} + \max(\text{Adds}_t,\epsilon)}$$6

The paper then uses a stabilized 250 ms implementation,

$$\mathrm{LWI}_t = \dfrac{\text{Cancels}_t}{\mathrm{MA}_{1\text{s}}(\text{DepthL1})_{t-1} + \max(\text{Adds}_t,\epsilon)}$$7

Here cancellations include Cancel messages and size-decreasing Modify messages at level 1, while executions are excluded from cancellation volume. The forecasting framework uses Nasdaq Market-by-Order data on a 250 ms grid, targets at 250 ms, 1 s, 2 s, and 5 s, and compares AR, HAR, and XGBoost. The reported pattern is that 250 ms forecasts perform poorly, HAR improves on AR at 1–2 s, and XGBoost dominates at 5 s, with $$\mathrm{LWI}_t = \dfrac{\text{Cancels}_t}{\mathrm{MA}_{1\text{s}}(\text{DepthL1})_{t-1} + \max(\text{Adds}_t,\epsilon)}$$8 above $$\mathrm{LWI}_t = \dfrac{\text{Cancels}_t}{\mathrm{MA}_{1\text{s}}(\text{DepthL1})_{t-1} + \max(\text{Adds}_t,\epsilon)}$$9 for most tickers, including AAPL $$\mathrm{LWI}_t = 1 - \Phi\!\left(\dfrac{\gamma_t^{(w)} - K_w H^{(1)}}{\sqrt{K_w H^{(2)}}}\right)$$0, NVDA $$\mathrm{LWI}_t = 1 - \Phi\!\left(\dfrac{\gamma_t^{(w)} - K_w H^{(1)}}{\sqrt{K_w H^{(2)}}}\right)$$1, TSLA $$\mathrm{LWI}_t = 1 - \Phi\!\left(\dfrac{\gamma_t^{(w)} - K_w H^{(1)}}{\sqrt{K_w H^{(2)}}}\right)$$2, NBIS $$\mathrm{LWI}_t = 1 - \Phi\!\left(\dfrac{\gamma_t^{(w)} - K_w H^{(1)}}{\sqrt{K_w H^{(2)}}}\right)$$3, RKLB $$\mathrm{LWI}_t = 1 - \Phi\!\left(\dfrac{\gamma_t^{(w)} - K_w H^{(1)}}{\sqrt{K_w H^{(2)}}}\right)$$4, and SNAP $$\mathrm{LWI}_t = 1 - \Phi\!\left(\dfrac{\gamma_t^{(w)} - K_w H^{(1)}}{\sqrt{K_w H^{(2)}}}\right)$$5 (Haochuan et al., 26 Sep 2025).

A related precursor appears in "Liquidity crises on different time scales" (Corradi et al., 2015). That paper does not define an LWI by name, but it introduces the two ingredients from which an LWI can be constructed: static depletion of the LOB at short horizons and failure of dynamic compensation between market orders and limit orders at longer horizons. Its side-specific exponential liquidity is

$$\mathrm{LWI}_t = 1 - \Phi\!\left(\dfrac{\gamma_t^{(w)} - K_w H^{(1)}}{\sqrt{K_w H^{(2)}}}\right)$$6

and liquidity imbalance is

$$\mathrm{LWI}_t = 1 - \Phi\!\left(\dfrac{\gamma_t^{(w)} - K_w H^{(1)}}{\sqrt{K_w H^{(2)}}}\right)$$7

At $$\mathrm{LWI}_t = 1 - \Phi\!\left(\dfrac{\gamma_t^{(w)} - K_w H^{(1)}}{\sqrt{K_w H^{(2)}}}\right)$$8 seconds, the paper reports

$$\mathrm{LWI}_t = 1 - \Phi\!\left(\dfrac{\gamma_t^{(w)} - K_w H^{(1)}}{\sqrt{K_w H^{(2)}}}\right)$$9

with fitted parameters $-v(t,y,S)$0 and $-v(t,y,S)$1, and the strongest correlation near $-v(t,y,S)$2–$-v(t,y,S)$3 ticks. At $-v(t,y,S)$4 minutes, effective liquidity is instead characterized by resilience, summarized by

$-v(t,y,S)$5

The resulting conception of LWI is time-scale dependent: static depletion dominates at short horizons, whereas failure of compensation dominates at longer horizons (Corradi et al., 2015).

3. Event-based and multi-scale formulations in intrinsic time

In "Multi-scale Representation of High Frequency Market Liquidity" (Golub et al., 2014), liquidity is defined through an event-based framework of directional changes and overshoots mapped into an Intrinsic Network. For thresholds $-v(t,y,S)$6, each threshold carries a binary overshoot direction, forming a state vector $-v(t,y,S)$7 and corresponding network state

$-v(t,y,S)$8

The paper defines per-transition surprisal $-v(t,y,S)$9 and path surprisal

$$\mathrm{LWI}_i(t) = \mathcal{N}(\mathrm{LSIS}_i(t)) \cdot \pi_i(t)$$0

Under ergodicity and a central limit theorem, the normalized activity-free liquidity metric is

$$\mathrm{LWI}_i(t) = \mathcal{N}(\mathrm{LSIS}_i(t)) \cdot \pi_i(t)$$1

The LWI operationalization over a rolling window $$\mathrm{LWI}_i(t) = \mathcal{N}(\mathrm{LSIS}_i(t)) \cdot \pi_i(t)$$2 is

$$\mathrm{LWI}_i(t) = \mathcal{N}(\mathrm{LSIS}_i(t)) \cdot \pi_i(t)$$3

with an optional exponentially weighted variant

$$\mathrm{LWI}_i(t) = \mathcal{N}(\mathrm{LSIS}_i(t)) \cdot \pi_i(t)$$4

In the 12-threshold setup, the paper reports $$\mathrm{LWI}_i(t) = \mathcal{N}(\mathrm{LSIS}_i(t)) \cdot \pi_i(t)$$5 and $$\mathrm{LWI}_i(t) = \mathcal{N}(\mathrm{LSIS}_i(t)) \cdot \pi_i(t)$$6, with one transition on average approximately every 5 seconds. The empirical evidence is case-study based: in USD/JPY during August 2007, LWI approached near-zero in the three weeks preceding the roughly $$\mathrm{LWI}_i(t) = \mathcal{N}(\mathrm{LSIS}_i(t)) \cdot \pi_i(t)$$7 drop on Aug 16, 2007; in EUR/CHF during August–September 2011, severe illiquidity was observed in the week preceding the roughly $$\mathrm{LWI}_i(t) = \mathcal{N}(\mathrm{LSIS}_i(t)) \cdot \pi_i(t)$$8 move around Aug 9, 2011, and liquidity was restored after the SNB set the 1.20 floor on Sep 6, 2011 (Golub et al., 2014).

This event-based formulation differs sharply from the LOB formulations. It requires only midprice ticks and a transition matrix rather than message-level order-book reconstruction, and it measures the unlikeliness of multi-scale price trajectories rather than direct cancellation pressure. A plausible implication is that it is especially suitable where order books are opaque, whereas top-of-book or message-level LWIs are better suited to electronically transparent venues.

4. Automated and concentrated-liquidity market makers

In "Optimal Exit Time for Liquidity Providers in Automated Market Makers" (Bergault et al., 8 Sep 2025), LWI is derived from an optimal stopping problem for a representative LP in a constant-product AMM with trading function $$\mathrm{LWI}_i(t) = \mathcal{N}(\mathrm{LSIS}_i(t)) \cdot \pi_i(t)$$9 and pool marginal price $$Z_t^{(\mathrm{bid})} = \mathbf{1}\{\beta_t > 0.5 + \xi\}, \qquad Z_t^{(\mathrm{ask})} = \mathbf{1}\{\beta_t < 0.5 - \xi\}.$$0. The oracle price evolves as

$$Z_t^{(\mathrm{bid})} = \mathbf{1}\{\beta_t > 0.5 + \xi\}, \qquad Z_t^{(\mathrm{ask})} = \mathbf{1}\{\beta_t < 0.5 - \xi\}.$$1

while buy and sell arrival intensities depend on the misalignment between $$Z_t^{(\mathrm{bid})} = \mathbf{1}\{\beta_t > 0.5 + \xi\}, \qquad Z_t^{(\mathrm{ask})} = \mathbf{1}\{\beta_t < 0.5 - \xi\}.$$2 and $$Z_t^{(\mathrm{bid})} = \mathbf{1}\{\beta_t > 0.5 + \xi\}, \qquad Z_t^{(\mathrm{ask})} = \mathbf{1}\{\beta_t < 0.5 - \xi\}.$$3:

$$Z_t^{(\mathrm{bid})} = \mathbf{1}\{\beta_t > 0.5 + \xi\}, \qquad Z_t^{(\mathrm{ask})} = \mathbf{1}\{\beta_t < 0.5 - \xi\}.$$4

The LP’s continuation value $$Z_t^{(\mathrm{bid})} = \mathbf{1}\{\beta_t > 0.5 + \xi\}, \qquad Z_t^{(\mathrm{ask})} = \mathbf{1}\{\beta_t < 0.5 - \xi\}.$$5 satisfies the HJB quasi-variational inequality

$$Z_t^{(\mathrm{bid})} = \mathbf{1}\{\beta_t > 0.5 + \xi\}, \qquad Z_t^{(\mathrm{ask})} = \mathbf{1}\{\beta_t < 0.5 - \xi\}.$$6

with terminal condition $$Z_t^{(\mathrm{bid})} = \mathbf{1}\{\beta_t > 0.5 + \xi\}, \qquad Z_t^{(\mathrm{ask})} = \mathbf{1}\{\beta_t < 0.5 - \xi\}.$$7. Since the stopping payoff is $$Z_t^{(\mathrm{bid})} = \mathbf{1}\{\beta_t > 0.5 + \xi\}, \qquad Z_t^{(\mathrm{ask})} = \mathbf{1}\{\beta_t < 0.5 - \xi\}.$$8, immediate withdrawal is optimal when $$Z_t^{(\mathrm{bid})} = \mathbf{1}\{\beta_t > 0.5 + \xi\}, \qquad Z_t^{(\mathrm{ask})} = \mathbf{1}\{\beta_t < 0.5 - \xi\}.$$9. The paper proposes several LWI constructions. The value-gap version is

$[64,32]$0

with normalization

$[64,32]$1

A boundary-proximity index and a QVI-residual index are also defined. Numerical solution is obtained by an Euler operator-splitting scheme or a Longstaff–Schwartz regression method. Comparative statics show that volatility, fees, arbitrage intensity, noise trader intensity, and misalignment $[64,32]$2 determine optimal exit. In the ETH–USDC baseline example, the paper computes $[64,32]$3 at a state with strong misalignment, interpreted as a strong incentive to withdraw immediately (Bergault et al., 8 Sep 2025).

In concentrated-liquidity DEXs, "SILS: Strategic Influence on Liquidity Stability and Whale Detection in Concentrated-Liquidity DEXs" (RajabiNekoo et al., 25 Jul 2025) defines LWI as a forward-looking LP risk score. The framework reconstructs per-tick liquidity from Uniswap v3-like Mint and Burn events, computes Exponential Time-Weighted Liquidity,

$[64,32]$4

or recursively,

$[64,32]$5

and measures each LP’s systemic importance through the Liquidity Stability Impact Score,

$[64,32]$6

The resulting LWI is

$[64,32]$7

where $[64,32]$8 is the estimated probability of near-future withdrawal, informed by ETWL trends, recent burns or collects, and anomaly scores. The paper also defines an impact-only $[64,32]$9, a time-averaged LWI, and a pool-level index $0.1$0. Unsupervised anomaly detection is performed with Isolation Forest, Local Outlier Factor, Robust Mahalanobis distance, or Random Cut Forest. In the toy example, removing LP1 raises price impact enough to produce $0.1$1, and with $0.1$2 the corresponding $0.1$3 is $0.1$4 (RajabiNekoo et al., 25 Jul 2025).

The AMM and CLMM formulations shift the meaning of LWI from order-book stress to LP withdrawal incentives and systemic dependency. The former is endogenous to a solved control problem; the latter is counterfactual and networked, combining impact magnitude with withdrawal likelihood.

5. Payment-system stress and behavioral withdrawal pressure

In "Modeling Trust and Liquidity Under Payment System Stress: A Multi-Agent Approach" (Amouzgar, 18 Feb 2026), liquidity withdrawal is modeled behaviorally rather than microstructurally. Customers carry trust $0.1$5, scar $0.1$6, rumor $0.1$7, behavioral mode $0.1$8, balance $0.1$9, and a withdrawal indicator $$\hat{p}_t = \mathcal{F}_\eta(e(t)) \cdot \sigma\!\bigl(w^\top f_\phi(\tilde{x}(e(t))) + b\bigr),$$0. Merchants contribute operational and broadcast states, with broadcast persistence inducing lag after technical recovery. Scar evolves according to

$$\hat{p}_t = \mathcal{F}_\eta(e(t)) \cdot \sigma\!\bigl(w^\top f_\phi(\tilde{x}(e(t))) + b\bigr),$$1

trust according to

$$\hat{p}_t = \mathcal{F}_\eta(e(t)) \cdot \sigma\!\bigl(w^\top f_\phi(\tilde{x}(e(t))) + b\bigr),$$2

and rumor according to

$$\hat{p}_t = \mathcal{F}_\eta(e(t)) \cdot \sigma\!\bigl(w^\top f_\phi(\tilde{x}(e(t))) + b\bigr),$$3

with

$$\hat{p}_t = \mathcal{F}_\eta(e(t)) \cdot \sigma\!\bigl(w^\top f_\phi(\tilde{x}(e(t))) + b\bigr),$$4

Withdrawal is threshold-gated by

$$\hat{p}_t = \mathcal{F}_\eta(e(t)) \cdot \sigma\!\bigl(w^\top f_\phi(\tilde{x}(e(t))) + b\bigr),$$5

and then activated with

$$\hat{p}_t = \mathcal{F}_\eta(e(t)) \cdot \sigma\!\bigl(w^\top f_\phi(\tilde{x}(e(t))) + b\bigr),$$6

Aggregate outflow is

$$\hat{p}_t = \mathcal{F}_\eta(e(t)) \cdot \sigma\!\bigl(w^\top f_\phi(\tilde{x}(e(t))) + b\bigr),$$7

with normalized outflow $$\hat{p}_t = \mathcal{F}_\eta(e(t)) \cdot \sigma\!\bigl(w^\top f_\phi(\tilde{x}(e(t))) + b\bigr),$$8.

The paper then maps these dynamics into two primary LWIs. The propensity version is

$$\hat{p}_t = \mathcal{F}_\eta(e(t)) \cdot \sigma\!\bigl(w^\top f_\phi(\tilde{x}(e(t))) + b\bigr),$$9

and the pressure version is

$\mathrm{LWI}_t = 100 \cdot \hat{p}_t.$0

The model proves that, under bounded memory, threshold gating, sticky broadcasts, and exogenous recovery, there exist trajectories with

$\mathrm{LWI}_t = 100 \cdot \hat{p}_t.$1

where $\mathrm{LWI}_t = 100 \cdot \hat{p}_t.$2 is the outage nadir. Simulations with $\mathrm{LWI}_t = 100 \cdot \hat{p}_t.$3–$\mathrm{LWI}_t = 100 \cdot \hat{p}_t.$4, $\mathrm{LWI}_t = 100 \cdot \hat{p}_t.$5–$\mathrm{LWI}_t = 100 \cdot \hat{p}_t.$6, $\mathrm{LWI}_t = 100 \cdot \hat{p}_t.$7–$\mathrm{LWI}_t = 100 \cdot \hat{p}_t.$8, $\mathrm{LWI}_t = 100 \cdot \hat{p}_t.$9–$$\mathrm{LWI}_t = \dfrac{\text{Cancels}_t}{\mathrm{MA}_{1\text{s}}(\text{DepthL1})_{t-1} + \max(\text{Adds}_t,\epsilon)}$$00, $$\mathrm{LWI}_t = \dfrac{\text{Cancels}_t}{\mathrm{MA}_{1\text{s}}(\text{DepthL1})_{t-1} + \max(\text{Adds}_t,\epsilon)}$$01–$$\mathrm{LWI}_t = \dfrac{\text{Cancels}_t}{\mathrm{MA}_{1\text{s}}(\text{DepthL1})_{t-1} + \max(\text{Adds}_t,\epsilon)}$$02, and $$\mathrm{LWI}_t = \dfrac{\text{Cancels}_t}{\mathrm{MA}_{1\text{s}}(\text{DepthL1})_{t-1} + \max(\text{Adds}_t,\epsilon)}$$03–$$\mathrm{LWI}_t = \dfrac{\text{Cancels}_t}{\mathrm{MA}_{1\text{s}}(\text{DepthL1})_{t-1} + \max(\text{Adds}_t,\epsilon)}$$04 steps show delayed peaks and hysteresis. Instant transfer substitution consistently reduces peak avoidance, and the paper reports a reduction of 1–2 percentage points across seeds, but cumulative outflows can remain high under persistent broadcasts (Amouzgar, 18 Feb 2026).

This formulation generalizes LWI beyond trading venues. It measures behavioral run pressure generated by memory persistence, social reinforcement, merchant messaging, and delayed normalization after an operational shock.

6. Validation, operational uses, and recurring limitations

Validation standards differ sharply across formulations. The ABIDES crumbling framework benefits from agent-level, side-specific, time-resolved ground truth and therefore supports ROC/AUC evaluation, temporal ablations, and tests under independent and Hawkes-driven autocorrelated withdrawal dynamics (Xu et al., 23 Apr 2026). The Nasdaq MBO forecasting framework emphasizes out-of-sample $$\mathrm{LWI}_t = \dfrac{\text{Cancels}_t}{\mathrm{MA}_{1\text{s}}(\text{DepthL1})_{t-1} + \max(\text{Adds}_t,\epsilon)}$$05 and RMSE under expanding-window walk-forward validation with embargo, together with stationarity checks and cross-ticker feature consensus (Haochuan et al., 26 Sep 2025). The Intrinsic Network paper is case-study based and does not report ROC/AUC or precision-recall metrics (Golub et al., 2014). The effective-liquidity analysis at multiple time scales is empirical and structural rather than framed as a predictive benchmark competition (Corradi et al., 2015). AMM and CLMM formulations rely on solved control problems, counterfactual simulations, or on-chain replay; payment-system LWI relies on calibrated multi-agent dynamics and proxy measurements (Bergault et al., 8 Sep 2025, RajabiNekoo et al., 25 Jul 2025, Amouzgar, 18 Feb 2026).

Operational use also varies by domain. In order-book settings, elevated LWI is intended for real-time monitoring, market surveillance, dynamic spread widening, participation control, timing of quote pullbacks, refill decisions, and execution-policy adjustments (Xu et al., 23 Apr 2026, Haochuan et al., 26 Sep 2025). In AMMs and CLMMs, LWI is used to guide LP exit decisions, identify high-impact LPs, gate burns through a protective oracle layer, and monitor pool-level fragility (Bergault et al., 8 Sep 2025, RajabiNekoo et al., 25 Jul 2025). In payment systems, LWI is proposed for incident response, post-recovery communication, merchant messaging control, and explicit monitoring of delayed-peak risk rather than technical uptime alone (Amouzgar, 18 Feb 2026).

Several limitations recur. Real markets generally lack ground truth distinguishing mechanical withdrawal from informational repricing, so mechanics-constrained filters reduce false positives but cannot fully guarantee mechanism identification outside simulation (Xu et al., 23 Apr 2026). Top-of-book LWIs ignore deeper-book resilience unless explicitly extended to L2 or L3, and single-venue MBO may miss intermarket withdrawal (Haochuan et al., 26 Sep 2025). The Intrinsic Network metric depends on a Markov approximation and on threshold selection, and its evidence is case-study based (Golub et al., 2014). The time-scale-dependent LOB framework requires periodic recalibration of $$\mathrm{LWI}_t = \dfrac{\text{Cancels}_t}{\mathrm{MA}_{1\text{s}}(\text{DepthL1})_{t-1} + \max(\text{Adds}_t,\epsilon)}$$06, $$\mathrm{LWI}_t = \dfrac{\text{Cancels}_t}{\mathrm{MA}_{1\text{s}}(\text{DepthL1})_{t-1} + \max(\text{Adds}_t,\epsilon)}$$07, $$\mathrm{LWI}_t = \dfrac{\text{Cancels}_t}{\mathrm{MA}_{1\text{s}}(\text{DepthL1})_{t-1} + \max(\text{Adds}_t,\epsilon)}$$08, and horizon weights (Corradi et al., 2015). The AMM exit model assumes a Brownian oracle price without jumps, a constant-product mechanism, and full withdrawal only (Bergault et al., 8 Sep 2025). SILS is sensitive to swap-flow regime shifts, routing, oracle quality, and cross-pool interactions (RajabiNekoo et al., 25 Jul 2025). The payment-system formulation requires proxy measurement for rumor, broadcasts, and avoidance, which may be difficult to observe directly (Amouzgar, 18 Feb 2026).

A common misconception is that LWI is a single standardized index with a fixed interpretation. The cited literature shows the opposite. LWI may be a probability, a cancellation ratio, an activity-normalized surprisal quantile, a normalized negative continuation value, a product of systemic impact and withdrawal likelihood, or a balance-weighted behavioral propensity. Another misconception is that technical recovery ends liquidity risk. The payment-system model explicitly formalizes delayed peaks after recovery, while the LOB and DeFi papers show that persistent clustering, regime dependence, or latent LP incentives can outlast the initial shock (Amouzgar, 18 Feb 2026, Xu et al., 23 Apr 2026, RajabiNekoo et al., 25 Jul 2025).

Taken together, the literature presents LWI as a unifying label for withdrawal-sensitive state variables rather than a single statistic. Its substantive meaning depends on whether liquidity is being studied as visible top-of-book depth, latent order revelation, multi-scale price-path improbability, optimal LP stopping, counterfactual systemic impact, or behaviorally gated outflow pressure.