Auto-deleveraging (ADL) activates on leveraged trading platforms whenever a user's position falls below required margins. The exchange must claw back funds from other open accounts to cover the shortfall. How those accounts are selected defines the ADL rule. Path-independent ADL rules, as characterized in this paper, do not depend on the sequence in which positions are unwound—they produce the same result regardless of traversal order. The authors show that such rules form an extended log-exponential family, a structure that connects them to well-known statistical distributions and enables clean mathematical analysis.
Immunity to account splitting, merging, and reopening further narrows the class. If an account is split into sub-accounts, merged with another, or reopened after being closed, a path-independent rule should allocate the deleveraging burden consistently. These invariance properties eliminate many natural candidates and push the formulation toward a smaller, more structured set.
Homogeneity—the property that scaling all account sizes scales the deleveraging proportionally—essentially restricts the class to a one-parameter family. Every rule in this family can be written as
$$ \delta_i = \frac{s_i^\theta}{\sum_j s_j^\theta} \cdot D $$
where \(s_i\) is the size of account \(i\), \(D\) is the total amount to deleverage, and \(\theta\) is the single parameter governing the rule. When \(\theta = 1\), the rule reduces to proportional division: each account pays a share proportional to its size. As \(\theta\) grows, the rule begins to favor larger accounts, behaving more like a water-filling allocation that protects smaller accounts up to a threshold and then charges larger ones disproportionately.The split-proof members of this family continuously interpolate between proportional division (\(\theta = 1\)) and water-filling (\(\theta \to \infty\)). This interpolation is not a discrete jump; every value of \(\theta\) generates a valid ADL rule, and the resulting allocations shift smoothly. The parameter \(\theta\) thus becomes a dial that trades off two objectives: fairness, which is maximized at \(\theta = 1\), and platform risk, which decreases as \(\theta\) grows because larger accounts absorb more of the burden.
Among these rules, decreasing \(\theta\) reduces platform risk at the cost of increasing inequality in how accounts are treated. Smaller accounts bear a disproportionately larger share when \(\theta < 1\), while larger accounts pay less than their proportional share. The authors argue that the parameter should be selected through ex-ante calibration: historical data on past margin calls and ADL events can be used to estimate a risk bound, and the largest \(\theta\) satisfying that bound yields the maximally fair rule under the constraint.
To test this framework, the authors perform a counterfactual replay of Hyperliquid's ADL episode on 10 October 2025. They re-run the deleveraging using various \(\theta\) values and compare the resulting allocations and subsequent risk statistics against the actual observed allocation. A range of intermediate \(\theta\) values—neither purely proportional nor fully water-filling—significantly improves on the observed allocation in both burden dispersion and downstream risk metrics. The improvement is consistent: these rules spread the deleveraging burden more evenly while keeping subsequent risk statistics closer to acceptable thresholds.
In practice, a platform operator could estimate the distribution of account sizes and historical margin shortfalls, then compute the minimal \(\theta\) that keeps the probability of exceeding a risk threshold below a target level. The resulting rule would be the fairest possible under that constraint. Because the family is one-dimensional, the calibration step is straightforward—line search over \(\theta\) with a risk estimator built on historical data.
The paper acknowledges limitations typical of a first characterization. The analysis assumes path-independence and homogeneity, properties that may not hold in all exchange designs. Real-world order-book dynamics, funding rates, and position correlation structures are not modeled in detail. The counterfactual replay uses a simplified snapshot of Hyperliquid's state at the time of the event; full-margin simulations with realistic fee structures and liquidation cascades were beyond the scope of the abstract.
For a working developer, the takeaway is a concrete, parameterized family of ADL rules with a clear fairness-risk trade-off. The one-parameter structure means no gradient-based optimization or multi-dimensional search is needed to tune the rule—simply calibrate \(\theta\) against historical data. The counterfactual results suggest that many existing exchanges operating near pure proportional division could reduce future risk without a large fairness hit by moving to an intermediate \(\theta\). The code and data needed to replicate the Hyperliquid replay are not publicly listed in the abstract, but the mathematical form is explicit and could be implemented directly.
The framework does not prescribe a single "correct" ADL rule but instead provides a calibrated dial for system designers. By grounding the rule in a log-exponential family and tying the parameter to an ex-ante risk bound, the work offers a principled alternative to ad hoc deleveraging policies that have, in practice, sometimes favored large accounts or produced unpredictable burden distribution. The Hyperliquid counterfactual replay demonstrates that the theory translates into measurable improvements on real trading data, which is a rare bridge between abstract mechanism design and on-chain practice.
Read the paper on arXiv