The difference between the lowest ask and highest bid: ask − bid. Tighter spreads mean cheaper round-trips for traders; wider spreads earn more per fill for the market maker but reduce fill frequency. The desk's quote width directly sets its passive spread.
see also:
Mid Price
The arithmetic midpoint between best bid and best ask: (bid + ask) / 2. Used to mark unrealized P&L each tick and as the baseline for computing edge. A mid move against a fill is the primary adverse selection signal.
see also:
Fair Value
The desk's model estimate of a contract's true probability in cents (forecastProb × 100). Edge = fair value − synthetic mid. The desk requires positive estimated edge above a minimum threshold before acting.
see also:
Bid
The highest price a buyer is currently willing to pay. Bids are sorted highest-first in the L2 book. The desk's passive bid rests here when it is offering to buy.
see also:
Ask (Offer)
The lowest price a seller is currently willing to accept. Asks are sorted lowest-first. The desk's passive ask rests here when it is offering to sell.
see also:
Maker / Taker
A maker adds liquidity by resting a limit order that others can fill. A taker removes liquidity by sending a market order that fills against resting orders. The desk is a maker when posting quotes and a taker when executing LIFT or HIT actions aggressively.
see also:
Order Flow
The stream of buy and sell orders arriving at the book each tick. The mix of participant types — noise, informed, liquidity-providing — determines whether flow is benign or adversarial for the desk.
see also:
Queue Priority
In a price-time priority book, orders at the same price are filled in the order they arrived. The synthetic book uses price priority only — the desk's quotes compete with synthetic participant orders at the same level.
see also:
Latent Probability
The model's internal belief about the true underlying probability — not directly observable. Derived from three layers: a long-run anchor, an adaptive base that mean-reverts toward it, and shock overlays from events and network contagion. The reported forecast probability is a smoothed version of this.
see also:
Adaptive Base
A slow-moving probability estimate that mean-reverts toward the long-run anchor. It drifts when latent risk builds (e.g. from planned maintenance), creating gradual price pressure before a visible event fires.
see also:
Shock Regime
A four-state market stability classification: calm → alert → shock → recovery. Driven by the magnitude of recent forecast changes. Shock regime increases overlay multipliers, suppresses aggressive market-making, and tightens the CVaR constraint in the optimizer.
see also:
Confidence
A 0–1 scalar capturing the model's certainty in its probability estimate — not the likelihood of the contract resolving YES. Low confidence (roughly below 46%) is caused by high volatility or regime uncertainty and suppresses aggressive desk actions. Confidence is a property of the model, not of the outcome.
see also:
Volatility Regime
A continuous EWMA estimate of how rapidly the forecast probability has been changing. High volatility leads to lower confidence, wider required edge, and more conservative optimizer plan selection. Shown as the 'market instability' bar in the forecast panel.
see also:
Inventory
The desk's signed net position in a market. Positive = long (net bought); negative = short (net sold). Bounded by hard limits. High inventory increases exposure to adverse price moves and triggers the optimizer's inventory skew constraint.
see also:
Inventory Limits
Hard position caps of ±8 contracts per market. As inventory approaches the limit, the desk widens quotes and may switch to flatten mode. The inventory_skew binding constraint fires in the optimizer when this threshold is approached.
see also:
Realized P&L
Profit or loss locked in by completed fills: (exit price − entry average cost) × quantity. Once realized, it does not change with subsequent market moves.
see also:
Unrealized P&L
The paper gain or loss on current open inventory, marked to the current mid price each tick: (mid − avg cost) × inventory. Changes with every tick until the position is closed.
see also:
P&L (Profit and Loss)
Total P&L = realized + unrealized. Realized is locked from prior fills. Unrealized is the current mark-to-market exposure. Total P&L is the primary performance metric for the desk.
see also:
Adverse Selection
The risk that a fill was executed against a counterparty with superior information. Flagged when mid price moves against the desk shortly after a fill — suggesting the other side knew something. Persistent adverse selection erodes P&L even when the quoted spread is theoretically positive.
see also:
Informed Flow
Orders from participants believed to hold an informational edge — primarily informed transit traders (tracking real MTA signals) and latency arbitrageurs. Filling against informed flow is the primary source of adverse selection.
see also:
Tail Risk
The risk of extreme losses in the low-probability tail of the return distribution. Measured via CVaR. In shock regime, tail risk is elevated because the probability model is less stable and large moves are more likely.
see also:
Quote Width
The distance between the desk's bid and ask (ourAsk − ourBid). A wider quote earns more per fill but is hit less often and is less likely to adversely select against informed flow. Shock regime and high inventory both push toward wider quotes.
see also:
Spread Capture
P&L earned by quoting a two-sided market passively and letting both sides fill over time. If the desk buys at the bid and later the same inventory is sold at the ask, the spread is captured. This is the core P&L mechanism for a passive market maker.
see also:
Aggressive Buy (LIFT)
A market order that crosses the spread and fills against the synthetic ask — the desk pays the full offer price to ensure a fill. Used when edge is large enough to justify the spread cost. Labelled LIFT OFFER in the Decision Engine.
see also:
Aggressive Sell (HIT)
A market order that fills against the synthetic bid — the desk sells at the bid price. Used when the model is bearish with sufficient edge and confidence. Labelled HIT BID in the Decision Engine.
see also:
Passive Maker
The desk's default mode: resting a two-sided limit order in the book and waiting for incoming orders to fill against it. HOLD maintains the current quote; WIDEN increases the spread for defensive protection. Earns the spread but carries adverse selection risk.
see also:
Flatten
Aggressively reducing inventory toward zero by crossing the spread — selling if long, buying if short. Triggered when inventory approaches the hard limit or when the optimizer determines tail risk from the position outweighs the spread crossing cost.
see also:
CVaR (Conditional Value at Risk)
The expected loss in the worst 10% of simulated future trajectories — a measure of tail risk. The JuMP optimizer uses CVaR as a hard constraint: plans whose tail risk exceeds the limit are rejected even with positive expected return. The limit tightens in shock regime.
see also:
Expected Return (μ return)
The mean P&L across all simulated rollout trajectories for a given plan. A plan with high expected return but elevated CVaR may be rejected. The optimizer selects the plan that maximizes expected return subject to risk constraints.
see also:
Fill Probability
The fraction of simulated trajectories in which the desk's passive quote gets hit by an incoming order. Tighter quotes increase fill probability but raise adverse selection exposure. The optimizer explicitly trades fill probability against per-fill P&L.
see also:
Scenario Rollouts
For each candidate plan (action + continuation policy), the optimizer simulates 50 stochastic trajectories over a 15-step horizon using the same probability dynamics model as the live simulation. Statistics across these trajectories feed the CVaR LP.
see also:
JuMP Optimizer
A Julia-based stochastic linear program solved using the JuMP modelling framework. Each tick, the system builds rollout trajectories for all candidate plans, computes CVaR statistics, and solves a constrained LP that maximizes expected return subject to risk limits. The solution selects the optimal plan.