zLend: A Dual-Scope Cash-Flow Reconstruction Framework for On-Chain Credit Underwriting

Girish G N; Ashutosh Sahoo; Akshay SP; Gurukiran S; Dhanashekar Kandaswamy

Source-linked AI summary

zLend: A Dual-Scope Cash-Flow Reconstruction Framework for On-Chain Credit Underwriting

Girish G N, Ashutosh Sahoo, Akshay SP, Gurukiran S, Dhanashekar Kandaswamy

arXiv:2608.16856v1Aug 17, 2026q-fin.RMcs.LG

TL;DR

Because on-chain lending lacks income and liability records, zLend reconstructs wallets under stablecoin-only and total-wealth scopes to derive repayment-capacity signals. Its validated sensitivity analysis finds that loan size principally governs tier assignment, while coverage and drawdown criteria bind on different wallets, and the framework is deployed in lending applications.

Problem
On-chain underwriting lacks income and liability records, while total wallet value conflates liquid repayment capacity with volatile or illiquid holdings.
Method
zLend reconstructs each wallet’s daily balances under stablecoin-only and total-wealth scopes, then derives parallel liquidity, cash-flow, and risk signals.
Results
Tier assignment is governed principally by the reference loan size, while coverage and drawdown criteria bind on disjoint wallets.
Takeaways & Limitations
zLend turns public transfer history into an underwriting signal used by real lending applications at loan-decision time.
Takeaways & Limitations
The sensitivity analysis uses six constructed reference wallets and characterizes the tier decision surface rather than wallet-population frequencies.

Abstract

from arXiv · show

Decentralized lending lacks a credit bureau: a borrower's capacity to repay must be inferred entirely from public on-chain activity, without income verification or a liability record. This paper presents zLend, a deployed cash-flow underwriting framework that reconstructs a wallet's daily balance history from raw token transfers and derives short-duration repayment-capacity signals from it. The reconstruction is performed twice per wallet, once restricted to a fixed stablecoin basket and once over all fungible transfers, on the premise that a wallet's total token holdings and its liquid, spendable balance are distinct quantities whose conflation misprices risk. From each series we derive liquidity coverage against a fixed loan size, cash-flow volatility and regularity, a drawdown-and-recovery statistic adapted from quantitative finance, and a recurring-counterparty detector that identifies salary-like payment cadence from transfer timing alone. The two views are then compared: a wallet with large aggregate holdings whose stablecoin reserve rarely covers the loan size is flagged as a liquidity mismatch irrespective of total wealth. We specify the pipeline formally, document the golden-master methodology used to verify a cross-language production migration to numerical tolerance 1e-9, and characterize the tier function's parameter sensitivity with an independent reimplementation validated to exact agreement (78 of 78 field assertions) against the deployed system's reference fixtures. Tier assignment is governed predominantly by the reference loan size, with four of six reference wallets changing tier across loan sizes from USD 10 to USD 25,000; the drawdown and coverage criteria bind on disjoint wallets, so neither subsumes the other; and no criterion in the tier rule is inert. zLend is deployed in production, informing real lending decisions via third-party API integrations.

1 Introduction

zLend addresses the absence of credit information in DeFi by separating liquid stablecoin capacity from total token wealth through dual-scope balance reconstruction. It turns this distinction into underwriting signals and is deployed in production through third-party lending applications.

Core framework: zLend reconstructs each wallet twice: once from stablecoins and once from all fungible transfers, treating divergence between liquid capacity and aggregate wealth as underwriting signal.Both views use identical reconstruction procedures and feed the same liquidity, cash-flow, and risk statistics.
Core framework: The formal pipeline derives daily balances from unordered transfers using cumulative net flow and a non-negativity offset without requiring the opening balance.This is specified as Algorithm 1.
Underwriting signals: The signal family measures liquidity coverage, inflow regularity, recurring counterparties, maximum drawdown and recovery, outflow concentration, and normalized trends.These signals are derived from the reconstructed balance series.
Underwriting signals: Divergence between the two scopes becomes graded liquidity- and flow-mismatch severity and a four-tier underwriting signal.The cross-scope comparison layer is specified completely in Tables 2 and 3.
Deployment and verification: 10^-9 numerical tolerance verifies the cross-language production migration, while the deployed framework reaches real borrowers through third-party lending applications.The implementation deliberately replicates summation, variance-convention, and rounding-mode semantics.

2 Related Work

Prior work shows behavioral regularity can substitute for missing credit histories, while DeFi lending research focuses on collateral, liquidation, and off-chain trust. zLend positions reconstructed daily balances and drawdown analysis as tools for short-duration repayment-capacity underwriting.

Alternative and cash-flow credit scoring: Behavioral regularity, including mobile usage and digital-footprint variables, can predict repayment and default comparably to or better than bureau scores.zLend’s recurring-counterparty and income-cadence detectors are presented as an on-chain analogue of this behavioral signal.
DeFi lending risk: DeFi lending risk research examines liquidation concentration, sophisticated searchers, and reflexive links between collateral value and solvency.These risks motivate the broader context for undercollateralized lending.
Undercollateralized lending protocols: Production undercollateralized protocols substitute different forms of off-chain trust for full on-chain collateral.Goldfinch uses off-chain backers and first-loss junior capital after traditional due diligence, while senior passive lenders bear residual risk.
Drawdown as a risk measure: zLend applies maximum drawdown to a reconstructed on-chain balance series as an underwriting criterion for unsecured credit, an application not identified in prior work.Maximum drawdown is otherwise standard in portfolio and strategy risk assessment.
Positioning: zLend narrows the zScore reputation lineage from general-purpose wallet reputation to short-duration repayment capacity using reconstructed daily balances instead of aggregate transaction statistics.The distinction is both credit-scoped and representational.

3 Problem Formulation and Balance Reconstruction

zLend reconstructs a wallet’s unobserved balances from token-transfer flows alone, using identical stablecoin and total-wealth scopes. The method aligns both reconstructed series on a shared calendar while treating the inferred opening offset as a data-implied lower bound.

3.1 Setting: A wallet is represented by timestamped transfers carrying USD value, normalized token symbol, type, direction, and counterparty, with no directly observed balance.
3.2 Scopes: The stablecoin scope uses a fixed 15-symbol basket, while symbol normalization folds bridged and chain-specific variants together.
3.2 Scopes: The stablecoin scope approximates capital deployable toward repayment without price risk, whereas total wealth captures the wallet’s full economic footprint.
3.3 Daily Aggregation and Derived Balance: Daily balances are reconstructed from cumulative net flow and shifted by the minimum offset that keeps the series non-negative.
3.4 Daily Spine: Both sparse scope series are projected onto one contiguous calendar spanning the earliest and latest transfer days, preserving direct index-wise comparability.
3.3 Daily Aggregation and Derived Balance: The offset is the smallest opening balance consistent with the observed history, not an estimate of true pre-history holdings; level statistics inherit it, but ratios and variation statistics do not.
3.4 Daily Spine: Reconstruction costs O(T log T + D) time and O(D) space, while the statistic suite is effectively linear in transfer count.

4 Signal Derivation

Signal derivation computes balance-based repayment, inflow, counterparty, drawdown, and flow-trend measures separately over rolling tail windows, while tracking data-quality limitations. Coverage is assessed both against the loan size and against each wallet’s operating balance scale.

Windowing: Rolling windows use the final min(W, |D|) spine days for W ∈{30, 60, 90}, avoiding zero-padding when wallet history is shorter.All statistics are computed per scope, and Table 1 lists the free parameters.
Liquidity coverage: Coverage measures the fraction of window days when derived balance covered loan size L outright, with a companion statistic using θdyn, the wallet’s 90-day median balance.The companion measures coverage relative to operating scale rather than an absolute figure.
Income cadence: Inflow frequency, inflow-amount coefficient of variation, and counterparty recurrence characterize payment regularity and income-like cadence.A counterparty must appear as an inflow source on at least krec = 3 distinct days, with day-gap criteria admitting weekly-to-monthly cadences.
Drawdown and recovery: Maximum drawdown tracks balance declines from running peaks, while duration records consecutive below-peak days and recovery time measures days until the prior peak returns.Recovery time is undefined when the pre-drawdown peak is not regained within the window.
Data quality: A wallet is flagged insufficient-history when |D| < 90, or insufficient-counterparty-data when either scope’s 90-day resolved-counterparty coverage falls below 0.8.The 0.8 threshold is diagnostic and distinct from the 30-day insufficiency criterion; incomplete counterparty data can understate recurrence.

5 Cross-Scope Comparison and Tier Assignment

zLend compares identical signals across stablecoin and total-wealth scopes to distinguish primary repayment capacity from contextual wealth. Tier assignment uses conjunctive, non-substitutable criteria, while mismatch severity and trend alignment preserve graded cross-scope differences.

Cross-scope comparison: The framework’s central design decision is to compare identical Section 4 signals computed under both scopes.The comparison is where the cross-scope design operates.
Mismatch analysis: Both mismatch flags use graded severity, and a four-way trend-alignment label records whether the scopes’ 30-day trend labels agree.The alignment labels are aligned-positive, aligned-flat, aligned-negative, and divergent.
Signal roles: The stablecoin scope determines the primary credit signal, while the total-wealth scope provides parallel context that qualifies but never overrides it.Table 3 specifies both signals.
Tier assignment: Primary-tier criteria are conjunctive and non-substitutable, so failing any one criterion prevents reaching that tier regardless of other strengths.Coverage measures usual solvency against the loan, whereas drawdown measures recent ability to withstand near-total loss.

6 Implementation and Verification

zLend is deployed as an in-process production wallet-scoring module whose outputs are persisted and consumed by third-party lending applications. A golden-master verification suite established cross-language agreement by reproducing numeric fields within 10^-9 tolerance and categorical fields exactly, requiring faithful replication of numerical semantics.

Production implementation: 166 columns are written per evaluation into a monthly-partitioned append-only table with 90-day retention.The record includes 69 fields per scope, six cross-scope ratios, mismatch and tier fields, data-quality flags, and four latency columns.
Cross-language verification: 10^-9 absolute tolerance and exact categorical agreement were required when verifying the TypeScript port against live Python reference outputs.The golden-master suite exercised scenarios covering each branch of the pipeline.
Numerical fidelity: Kahan summation, population dispersion with denominator n, and round-half-to-even at ten decimal places preserve reference numerical semantics.These choices replace naive accumulation, sample-statistic defaults, and the target language’s native rounding mode.
Numerical fidelity: Ratios with absolute denominators below 10^-12 produce NaN and explicit null at the API boundary rather than zero.This preserves the distinction between undefined and zero.
Production integration: Third-party lending applications integrate with the production API and depend on exact specification-defined values on every call.The passage presents numerical fidelity as enabling zLend’s production deployment.

7 Worked Examples

The worked examples use verified synthetic fixtures to demonstrate how zLend distinguishes strong repayment capacity, severe drawdown risk, and liquidity mismatches between total holdings and stablecoin reserves. They also show that partial history and recurring-income cadence can support underwriting decisions.

Reference fixtures: Six verified synthetic wallets exercise distinct underwriting regimes from the deployed system’s reference fixtures.Each wallet was independently reproduced and checked against the reference fixtures.
Recurring income: $2,500 stablecoin payments on a seven-day cadence are correctly flagged as income-like by the recurring-income detector.The recurring-income wallet’s balance declines by no more than 9.2%.
Partial history: 36 days of observed activity is sufficient for strong-tier admission despite falling below the 90-day data-quality threshold.The history exceeds the 30-day insufficiency floor, so the framework scores available partial history rather than withholding judgment.
Drawdown risk: 0.95 maximum drawdown makes the drawdown wallet weak despite 100% trailing-30-day coverage and coverage of a $100 loan on every day.The wallet loses 95% of its $10,000 balance in one day, remains near $500 for three weeks, and later recovers past its original level.
Liquidity mismatch: $270,030 total-wealth balance alongside stablecoin coverage of 0.00 exposes a severe liquidity mismatch that total-value screening would miss.The balance ratio exceeds 6,000, stablecoin inflow is 0% of trailing total inflow, and the two screens diverge by roughly four orders of magnitude.

8 Sensitivity and Binding-Constraint Analysis

Sensitivity analysis shows that loan size is the dominant tier parameter, while drawdown is comparatively stable and coverage and drawdown provide nonredundant constraints. The validated reimplementation also finds that mismatch flags respond differently across parameters, with the balance-ratio threshold producing no tested transitions.

Validation and scope: 78 of 78 assertions matched at tolerance 10^-9 before parameters were perturbed in an independent reimplementation of the deployed pipeline.The analysis characterizes the tier function’s decision surface, not wallet-population frequencies; six reference wallets span designed operating regimes.
Loan-size sensitivity: Four of six wallets change tier across L ∈ [$10, $25,000], because coverage is defined relative to L and scale-free criteria leave two wallets invariant.A thin-stablecoin wallet is strong at L = $10 but weak at L = $50, while L represents product scope rather than merely a tuning constant.
Drawdown-bound sensitivity: Four of six wallets are invariant across the full strong-tier drawdown-bound range, and the deployed default of 0.35 is adjacent to no wallet’s realized value.The severe-drawdown wallet is admitted only at a bound of 0.95, where the criterion excludes nothing.
Binding constraints: Every criterion is decisive for at least one wallet, while coverage and drawdown bind on disjoint wallets and neither statistic subsumes the other.The severe-drawdown wallet fails on drawdown alone, whereas the thin-stablecoin wallet fails on coverage alone; Figure 5 depicts these orthogonal failures geometrically.
Mismatch-flag sensitivity: Zero flag transitions occur when the balance-ratio threshold moves from 1.5× to 10×, while the remaining mismatch thresholds are sensitive around observed wallet values.The two balance-ratio flags occur at approximately 6,668× and at the 10^-9 numerical floor; the liquidity threshold crosses at 0.5 for a wallet with 30-day coverage 0.40.

9 Conclusion

zLend reconstructs wallet cash flows under stablecoin-only and total-wealth scopes to derive short-duration underwriting signals, including liquidity coverage, regularity, drawdown recovery, and cross-scope divergence. Its specification is production-verified, and sensitivity analysis shows loan size principally governs tiers while coverage and drawdown capture distinct wallets.

Framework contribution: zLend reconstructs daily wallet balances under stablecoin-only and total-wealth scopes, deriving liquidity coverage, cash-flow regularity, and drawdown-and-recovery underwriting signals.The framework also compares aggregate holdings with liquid reserves as an underwriting signal.
Production verification: The production implementation is verified against its reference to 10−9 numerical tolerance.The specification is given in full before production verification.
Sensitivity and deployment: Sensitivity analysis establishes that tier assignment is governed principally by the reference loan size, while coverage and drawdown criteria bind on disjoint wallets.Their disjoint binding confirms that neither criterion subsumes the other; zLend runs inside real lending applications through the Section 6 API.