Measuring the engine of a liquidation cascade: subcritical branching inside a first-order transition

Battle Count: 71

The paper is primarily a market-structure and financial-physics study rather than a direct trading strategy paper. However, findings are highly relevant for: (1) understanding that liquidation cascades are first-order (abrupt) rather than critical (gradual), meaning no reliable single-variable early-warning signal exists for timing entries/exits around crashes; (2) the liquidity-sector signature (impact spiking 3-9x, OI clearing 25-70%) provides actionable in-cascade regime detection; (3) the subcritical branching finding implies that within-venue feedback amplification is limited, but cross-venue contagion remains the risk channel; (4) the backstop mechanism design insight is relevant for venue selection in execution strategies. The work informs risk management and market-making decisions more than alpha generation.

High complexity due to: (1) multi-source data pipeline requiring careful timestamp alignment (Binance metrics stamp interval end vs. klines stamp start, requiring -5min realignment); (2) fill-log deduplication by liquidated-user leg to avoid double-counting; (3) correlation matrix construction, eigenvalue decomposition, and random matrix theory (Marčenko-Pastur edge) analysis; (4) Kyle regression with multiple specifications (price-conditioned, volume-native); (5) INAR/Hawkes process estimation on one-minute forced-sell counts; (6) finite-size subsampling analysis with 40 draws per size; (7) circular moving-block bootstrap (B=5000) for trend significance; (8) simulated power calculations for falsification tests; (9) regime-window decomposition of the branching ratio. Requires expertise in financial econometrics, random matrix theory, point processes, and statistical physics.

All three data sources are public and reconstructible without proprietary feeds: Binance panel and BTC series from public data.binance.vision dumps; Hyperliquid per-minute asset-context archive from s3://hyperliquid-archive (requester-pays); Hyperliquid node fill log from s3://hl-mainnet-node-data. Every figure and number is produced by a script from a frozen experiment record (EXP-012 through EXP-019b), which are append-only and fix question, data, method, numbers, and verdict at run time. Two load-bearing conventions are explicitly stated (timestamp alignment, fill-log deduplication). Analysis code available from author on request.