Public Trader Identity: Adverse Selection and Return Predictability

Battle Count: 93

This paper is directly and highly relevant to quantitative trading. It demonstrates that public wallet identity on decentralized exchanges provides statistically significant and economically meaningful incremental predictive power for short-horizon returns (13.2% R² gain at one second, t=9.2). The findings are immediately actionable for market makers on Hyperliquid who can condition quotes on toxic wallet activity. The paper quantifies the information lost when anonymous order-book data are used in place of identity-augmented features, directly relevant to adverse selection management, optimal quoting, and toxic flow detection. The methodology (wallet scoring, identity-augmented prediction, matched-wallet placebo) provides a template for similar analyses on other venues.

Implementation requires: (1) running a non-validating Hyperliquid node and collecting 17.1 billion Level-4 messages, (2) reconstructing the full-depth limit order book in consensus order with oracle-gate validation, (3) computing notional-weighted markouts for 147,113 wallets, (4) constructing 11 identity features on a 100ms grid, (5) fitting ridge and gradient-boosted tree models with expanding-day validation, (6) running 200 matched-wallet placebo cohorts (8.8 billion cohort-observations from 126 GB of replayed per-wallet quote tapes), and (7) handling infrastructure events (node failures, gate closures, media damage). The data volume and reconstruction pipeline are substantial, though the GitHub repository provides code.

The paper provides a GitHub repository (https://github.com/daojingzhai/public-trader-identity) with reconstruction code, full audit tables, and replication artifacts. Data is collected from a public blockchain (Hyperliquid) via a non-validating node. The December 2025 replication uses an independently collected public dataset. All model specifications, freezing protocols, variable definitions, and evaluation procedures are fully documented. The collection infrastructure events (node failures, gate closures) are transparently disclosed.