Autodeleveraging as Online Learning

Battle Count: 51

While not directly about trading strategies, this paper is highly relevant to quantitative traders in perpetual futures markets. ADL directly affects position management, leverage decisions, and tail-risk exposure. Understanding ADL mechanics helps traders assess the true cost of holding leveraged positions during stress events. The paper's analysis of queue vs. pro-rata policies informs which venues offer better protection for profitable positions. For market makers and HLP participants, the execution-price estimation and liquidity feedback analysis is directly applicable. The October 10 event analysis provides concrete dollar-impact estimates relevant to risk modeling.

The theoretical framework (online convex optimization, dynamic regret bounds) is moderately complex but well-established. The empirical replay methodology requires careful handling of public data, markout estimation, and policy comparison under fixed state trajectories. The vector mirror descent algorithm is straightforward to implement. The ILP for integer pro-rata and the queue instability diagnostics add engineering complexity. The main challenge is the execution-price estimation and the separation of ex ante/ex post benchmarks, which requires careful accounting.

The paper uses public Hyperliquid replay data from October 10, 2025, with open-source reconstruction code (GitHub: pluriholonomic/autodeleveraging-analysis and ConejoCapital/HyperMultiAssetedADL). Replay assumptions and observation model are explicitly stated. However, the full internal-ledger reconstruction is not available, and some quantities (execution prices, markout horizons) require estimation. Theoretical proofs are provided in appendices.