Risk-Based Auto-Deleveraging

ArXiv.org · 2026-03-16

Auto-deleveraging (ADL) mechanisms are a critical yet understudied component of risk management on cryptocurrency futures exchanges. When available margin and other loss-absorbing resources are insufficient to cover losses following large price moves, exchanges reduce positions and socialize losses among solvent participants via rule-based ADL protocols. We formulate ADL as an optimization problem that minimizes the exchange's risk of loss arising from future equity shortfalls. In a single-asset, isolated-margin setting, we show that under a risk-neutral expected loss objective the unique optimal policy minimizes the maximum leverage among participants. The resulting design has a transparent structure: positions are reduced first for the most highly levered accounts, and leverage is progressively equalized via a water-filling (or ``leverage-draining'') rule. This policy is distribution-free, wash-trade resistant, Sybil resistant, and path-independent. It provides a canonical and implementable benchmark for ADL design and clarifies the economic logic underlying queue-based mechanisms used in practice. We further study the multi-asset, cross-margin setting, where the ADL problem becomes genuinely multi-dimensional: the exchange must allocate a vector of required reductions across accounts with portfolios exposed to correlated price moves. We show that under an expected-loss objective the problem remains separable across accounts after introducing asset-level shadow prices, yielding a scalable numerical method. We observe that naive gross leverage can be misleading in this context as it ignores hedging within portfolios. When asset prices are driven by a single dominant risk factor, the optimal policy again takes a water-filling form, but now in a factor-adjusted notion of leverage, so that more effectively hedged portfolios are deleveraged less aggressively.

Risk-Based Auto-Deleveraging

arXiv (Cornell University) · 2026-03-16

Auto-deleveraging (ADL) mechanisms are a critical yet understudied component of risk management on cryptocurrency futures exchanges. When available margin and other loss-absorbing resources are insufficient to cover losses following large price moves, exchanges reduce positions and socialize losses among solvent participants via rule-based ADL protocols. We formulate ADL as an optimization problem that minimizes the exchange's risk of loss arising from future equity shortfalls. In a single-asset, isolated-margin setting, we show that under a risk-neutral expected loss objective the unique optimal policy minimizes the maximum leverage among participants. The resulting design has a transparent structure: positions are reduced first for the most highly levered accounts, and leverage is progressively equalized via a water-filling (or ``leverage-draining'') rule. This policy is distribution-free, wash-trade resistant, Sybil resistant, and path-independent. It provides a canonical and implementable benchmark for ADL design and clarifies the economic logic underlying queue-based mechanisms used in practice. We further study the multi-asset, cross-margin setting, where the ADL problem becomes genuinely multi-dimensional: the exchange must allocate a vector of required reductions across accounts with portfolios exposed to correlated price moves. We show that under an expected-loss objective the problem remains separable across accounts after introducing asset-level shadow prices, yielding a scalable numerical method. We observe that naive gross leverage can be misleading in this context as it ignores hedging within portfolios. When asset prices are driven by a single dominant risk factor, the optimal policy again takes a water-filling form, but now in a factor-adjusted notion of leverage, so that more effectively hedged portfolios are deleveraged less aggressively.

Quantifying the Value of Revert Protection

Lecture notes in computer science · 2026-01-01

A Framework for Combined Transaction Posting and Pricing for Layer 2 Blockchains

Lecture notes in computer science · 2026-01-01

Automated Market Making and Arbitrage Profits in the Presence of Fees

Lecture notes in computer science · 2025-01-01 · 7 citations

A Framework for Combined Transaction Posting and Pricing for Layer 2 Blockchains

ArXiv.org · 2025-05-26

This paper presents a comprehensive framework for transaction posting and pricing in Layer 2 (L2) blockchain systems, focusing on challenges stemming from fluctuating Layer 1 (L1) gas fees and the congestion issues within L2 networks. Existing methods have focused on the problem of optimal posting strategies to L1 in isolation, without simultaneously considering the L2 fee mechanism. In contrast, our work offers a unified approach that addresses the complex interplay between transaction queue dynamics, L1 cost variability, and user responses to L2 fees. We contribute by (1) formulating a dynamic model that integrates both posting and pricing strategies, capturing the interplay between L1 gas price fluctuations and L2 queue management, (2) deriving an optimal threshold-based posting policy that guides L2 sequencers in managing transactions based on queue length and current L1 conditions, and (3) establishing theoretical foundations for a dynamic L2 fee mechanism that balances cost recovery with congestion control. We validate our framework through simulations.

Latency Advantages in Common-Value Auctions

ArXiv.org · 2025-04-02

In financial applications, latency advantages -- the ability to make decisions later than others, even without the ability to see what others have done -- can provide individual participants with an edge by allowing them to gather additional relevant information. For example, a trader who is able to act even milliseconds after another trader may receive information about changing prices on other exchanges that lets them make a profit at the expense of the latter. To better understand the economics of latency advantages, we consider a common-value auction with a reserve price in which some bidders may have more information about the value of the item than others, e.g., by bidding later. We provide a characterization of the equilibrium strategies, and study the welfare and auctioneer revenue implications of the last-mover advantage. We show that the auction does not degenerate completely and that the seller is still able to capture some value. We study comparative statics of the equilibrium under different assumptions about the nature of the latency advantage. Under the assumptions of the Black-Scholes model, we derive formulas for the last mover's expected profit, as well as for the sensitivity of that profit to their timing advantage. We apply our results to the design of blockchain protocols that aim to run auctions for financial assets on-chain, where incentives to increase timing advantages can put pressure on the decentralization of the system.

Tail-Optimized Caching for LLM Inference

ArXiv.org · 2025-10-16

Prompt caching is critical for reducing latency and cost in LLM inference: OpenAI and Anthropic report up to 50-90% cost savings through prompt reuse. Despite its widespread success, little is known about what constitutes an optimal prompt caching policy, particularly when optimizing tail latency, a metric of central importance to practitioners. The widely used Least Recently Used (LRU) policy can perform arbitrarily poor on this metric, as it is oblivious to the heterogeneity of conversation lengths. To address this gap, we propose Tail-Optimized LRU, a simple two-line modification that reallocates KV cache capacity to prioritize high-latency conversations by evicting cache entries that are unlikely to affect future turns. Though the implementation is simple, we prove its optimality under a natural stochastic model of conversation dynamics, providing the first theoretical justification for LRU in this setting, a result that may be of independent interest to the caching community. Experimentally, on real conversation data WildChat, Tail-Optimized LRU achieves up to 27.5% reduction in P90 tail Time to First Token latency and 23.9% in P95 tail latency compared to LRU, along with up to 38.9% decrease in SLO violations of 200ms. We believe this provides a practical and theoretically grounded option for practitioners seeking to optimize tail latency in real-world LLM deployments.

Outbidding and Outbluffing Elite Humans: Mastering Liar's Poker via Self-Play and Reinforcement Learning

arXiv (Cornell University) · 2025-11-05

AI researchers have long focused on poker-like games as a testbed for environments characterized by multi-player dynamics, imperfect information, and reasoning under uncertainty. While recent breakthroughs have matched elite human play at no-limit Texas hold'em, the multi-player dynamics are subdued: most hands converge quickly with only two players engaged through multiple rounds of bidding. In this paper, we present Solly, the first AI agent to achieve elite human play in reduced-format Liar's Poker, a game characterized by extensive multi-player engagement. We trained Solly using self-play with a model-free, actor-critic, deep reinforcement learning algorithm. Solly played at an elite human level as measured by win rate (won over 50% of hands) and equity (money won) in heads-up and multi-player Liar's Poker. Solly also outperformed large language models (LLMs), including those with reasoning abilities, on the same metrics. Solly developed novel bidding strategies, randomized play effectively, and was not easily exploitable by world-class human players.

Optimal Dynamic Fees for Blockchain Resources

Lecture notes in computer science · 2025-01-01 · 3 citations