做市商队列模型评估：闭式动作回放方法
[quant trading]
面向量化做市从业者，提供了一套可复现的队列模型评估方法（闭式动作回放），强调动作保真度而非盈亏，值得研究参考。
[quant trading]
[量化交易]
suppose you run a live market maker and now want to assess whether your queue modelling is any good.
假设你运营一个实时做市商，现在想评估你的队列建模是否有效。
how do you do this? a mm system is typically state-dependent against inventory. A backtest simulation of your mm is therefore path dependent and cumulative errors make the comparisons moot.
你如何做到这一点？做市系统通常依赖于库存状态。因此，你的做市回测模拟是路径依赖的，累积误差使得比较变得毫无意义。
you do this by performing a closed-action replay. you need two objects:
你通过执行封闭动作重放来实现。你需要两个对象：
- live order trace: submission and cancellation intents, exchange timestamps, order IDs, acknowledgements, fills and terminal states.
- public ticks: L2 updates, trades, exchange and receive timestamps, and sequence ordering.
- 实时订单轨迹：提交和取消意图、交易所时间戳、订单ID、确认、成交和最终状态。
- 公共行情：L2更新、交易、交易所和接收时间戳，以及序列排序。
you then run the simulation engine against only the order trace instead of the mm logic. order/cancellation latency distributions are backed out against the live trace. your simulation generates a comparable trace object under the parametric queue model assumptions.
然后，你仅针对订单轨迹而不是做市逻辑运行模拟引擎。订单/取消延迟分布根据实时轨迹进行反推。你的模拟在参数化队列模型假设下生成可比较的轨迹对象。
you can now create a similarity vector, like fill precision and recall scores, first fill time deltas with smooth decay function, etc. based on your use case, a fitness function scores this similarity vector. this queue model is then be "tuned" against the exchange, where the objective is to produce high action fidelity rather than generate better pnl
你现在可以创建一个相似度向量，例如成交精确率和召回率分数、带有平滑衰减函数的首次成交时间差等。根据你的用例，适应度函数对该相似度向量进行评分。然后，该队列模型将针对交易所进行“调整”，其目标是产生高动作保真度，而不是产生更好的盈亏。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力