跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
搜索与研究

queue-position-modeling

Estimate passive-order queue position and fill probability from MBO or aggregated L2 data; compare conservative and probabilistic queue models and calibrate them against observed executions. Use for maker-fill simulation, quote placement, and queue-value research.

voyager67 · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add voyager67/JaneStreet-Agentic-Skills --skill queue-position-modeling

技能说明


name: queue-position-modeling description: Estimate passive-order queue position and fill probability from MBO or aggregated L2 data; compare conservative and probabilistic queue models and calibrate them against observed executions. Use for maker-fill simulation, quote placement, and queue-value research.

Queue Position and Passive Fill Modeling

Treat queue position as a latent state unless the feed and venue protocol expose sufficient order-level information. The objective is a calibrated fill distribution, not a single overconfident point estimate.

Workflow

  1. Establish matching priority. Confirm FIFO/price-time, pro-rata, size-time, or hybrid rules, plus special priority classes, hidden/reserve liquidity, self-trade prevention, and order modification priority.
  2. Classify available data. MBO with stable order IDs may support direct queue reconstruction. MBP/L2 reports aggregate size and requires an inference model. State exactly which facts are observable.
  3. Initialize queue ahead. Align the order's venue acknowledgement time with book events. Estimate displayed quantity ahead at the price when the order becomes live, accounting for exchange-to-feed and order-entry timing. Do not initialize from the local decision timestamp alone.
  4. Process events. Trades at the price consume queue ahead according to aggressor and matching rules. Additions behind do not move the order backward under FIFO. Cancellations/reductions in L2 are ambiguous: model the fraction ahead versus behind rather than assuming every decrement advances the order.
  5. Compare model families. At minimum, test:
    • Risk-averse bound: cancellations do not improve the simulated queue; only executions advance it.
    • Probabilistic model: allocate ambiguous depth reductions ahead/behind using a monotone function of estimated ahead and behind volume. Test several shapes rather than picking one because it produces more fills.
    • MBO reconstruction: exact order-level transitions only when the data and venue priority semantics justify it.
  6. Add timing uncertainty. Simulate feed latency, order-entry latency, acknowledgement latency, and cancel latency. Include the possibility that the market moves before the order is acknowledged or cancelled.
  7. Calibrate on held-out fills. Compare predicted fill probability, time-to-fill, partial-fill size, and queue-ahead evolution with actual order outcomes. Segment by depth, order size, spread, volatility, and time of day.
  8. Report bounds and uncertainty. Run conservative/base/aggressive assumptions and show the strategy's P&L and fill rate range. A backtest that only works under optimistic queue assumptions is not robust.

Diagnostics

  • Fill probability calibration curve and Brier/log loss where appropriate.
  • Time-to-first-fill and time-to-complete-fill distributions.
  • Predicted versus realized partial-fill ratios.
  • Fill rates by estimated queue-ahead bucket and order size/depth ratio.
  • Sensitivity to unknown cancels, feed batching, and latency distributions.

Guardrails

  • Never treat L2 queue position as ground truth.
  • Do not use fills to calibrate and evaluate on the same sample without a temporal holdout.
  • A cancel is not automatically a cancellation ahead of your order.
  • A full fill is not evidence the model is correct on unfilled orders; account for censoring.
  • Do not claim realistic maker backtesting if queue position and latency are absent.

Deliverable

State the matching rule, observability level, initialization timing, queue model equations/parameters, calibration split, fit diagnostics, scenario bounds, and the degree to which the result is identified by the data.

Further reading