Market Impact Simulation

Simulates how aggressive orders consume resting liquidity across multiple price levels in the Limit Order Book (LOB). The model calculates the instantaneous price impact by walking the order book, accounting for order book depth and resilience. This allows agents to learn the cost of urgency versus patience.

• Models volume-weighted average price (VWAP) slippage
• Accounts for hidden and iceberg orders
• Simulates quote stuffing and fleeting liquidity events

Captures the temporary price distortion that reverts after a trade is executed, distinct from permanent information impact. The simulator models the exponential decay of this mechanical pressure, teaching agents to exploit mean-reversion windows.

• Uses Hawkes Process kernels for impact propagation
• Distinguishes between transient and permanent impact components
• Calibrated to empirical decay half-lives observed in equity markets

Employs Generative Adversarial Networks (GANs) or Diffusion Models to create realistic, non-stationary order flow that adapts to the agent's strategy. The simulator acts as an adversary, generating spoofing patterns and predatory latency arbitrage to stress-test execution algorithms.

• Generates stylized facts like volatility clustering and fat tails
• Trains agents via self-play against evolving market conditions
• Exposes strategies to market manipulation simulation scenarios

Replicates the complexity of modern fragmented markets by simulating multiple competing exchanges, dark pools, and Smart Order Routers (SOR). Agents learn to optimize execution across venues with varying fee structures, latencies, and fill probabilities.

• Models maker-taker and inverted fee schedules
• Simulates regulatory trade-through protection rules
• Incorporates latency arbitrage between geographically dispersed venues

Integrates tail-risk measures directly into the agent's reward function during training. Beyond minimizing expected slippage, the simulator penalizes strategies that exhibit high Conditional Value at Risk (CVaR) or fail during regime-switching events.

• Optimizes for CVaR and maximum adverse excursion
• Incorporates volatility regime conditioning
• Penalizes execution shortfall variance, not just mean

Uses Neural SDEs and Signature Wasserstein GANs (SigCWGAN) to calibrate simulators directly to tick-level market data. The system learns the joint distribution of order arrivals, cancellations, and price movements, ensuring synthetic environments replicate the rough volatility and Hawkes process dynamics of real markets.

• Matches empirical order book shape and resilience
• Reproduces Epps effect in correlated assets
• Validated via adversarial validation against out-of-sample data

A set of consistent statistical properties observed across financial time series that any realistic market simulator must replicate. These are the acceptance criteria for synthetic market data.

• Volatility Clustering: Large price changes tend to be followed by large changes, and small by small.
• Fat Tails: The distribution of returns exhibits excess kurtosis compared to a normal distribution.
• Absence of Autocorrelation: Raw returns show almost no linear autocorrelation, while absolute returns show long memory.
• Leverage Effect: Volatility tends to increase when prices drop.