Hiding a Market Maker's Inventory: Quantized and Jittered Quote Skew against an Explicit Attacker

A market maker that shades its quotes against its inventory reveals that inventory to an observer who knows its quoting rule and the inputs to that rule. I construct an explicit attacker against an Avellaneda–Stoikov maker in an agent-based limit order book calibrated to the Binance.US bitcoin–dollar (BTC/USD) market. An attacker that is told which quotes are the maker’s, observes the price the maker quotes against and the maker’s horizon, and is trained on true inventory recovers that inventory exactly; with the full-book mid in place of the maker’s reference price, its out-of-sample R² is 0.03 ± 0.35 (mean ± standard error). Averaging its observations lets the attacker partially recover a skew jittered by independent noise, but produces no detectable gain against a quantized skew. On average, quantization preserves more inventory control than removing the skew, yet each of the four widths tested between 0.75 and 1.50 standard deviations of inventory leaks information relative to a zero-skew control, and a negative out-of-sample R² cannot show that a defense blinds the attacker. When an attacker trades, the maker’s seven-day profit falls by $108 ± 33, an effect the design cannot attribute to leakage, and at a larger quote size a symmetric one-hour strategy using true inventory does not cover its trading costs.