Domain Randomization

The engine randomly samples environment variables from predefined distributions at the start of each training episode. This prevents the agent from overfitting to a single market regime.

• Volatility (e.g., σ ∈ [0.1, 0.8])
• Spread width (e.g., 1-10 basis points)
• Arrival rate intensity of the Hawkes process
• Latency jitter to simulate network delays
• Correlation matrices between assets

By exposing the agent to extreme and varied conditions, the neural network is forced to discard spurious correlations tied to specific simulation parameters. The resulting policy relies on causal economic relationships rather than memorized noise patterns.

This directly addresses the sim-to-real gap by ensuring the learned strategy generalizes to the true data-generating process of live markets, not just the simulator's specific configuration.

Unlike static data augmentation, dynamics randomization alters the underlying transition function of the environment itself. The agent must adapt to changing physics of the market.

• Randomizing the order book matching engine rules
• Varying the tick size regime
• Switching between continuous and call auction mechanisms
• Altering the latency profile of market data feeds

In market simulation, 'vision' translates to the observation tensor fed to the agent. Randomization injects noise to simulate real-world sensor and data imperfections.

• Gaussian noise added to price and volume feeds
• Random packet drops simulating UDP market data loss
• Stale data injection where the agent acts on delayed quotes
• Outlier injection to simulate flash crashes or erroneous trades

Domain randomization is often paired with a curriculum that gradually expands the sampling bounds. The agent starts in a narrow, stable regime and progressively faces wilder parameter ranges.

• Phase 1: Tight spreads, low volatility
• Phase 2: Widen spread and volatility bounds
• Phase 3: Introduce adversarial market manipulation
• Phase 4: Full randomization with fat-tailed event shocks

An advanced technique where the sampling distribution itself adapts based on the agent's performance. The system automatically expands the randomization range when the agent masters the current difficulty.

• Uses Wasserstein distance to measure sim-to-real distribution gaps
• Expands bounds only when success rate exceeds a threshold (e.g., 80%)
• Contracts bounds if the agent fails catastrophically
• Results in a self-tuning curriculum without manual parameter engineering

The performance discrepancy that occurs when a trading model trained in a synthetic environment is deployed in live markets. Domain randomization directly addresses this by forcing the agent to learn invariant features rather than exploiting simulator-specific artifacts.

• Cause: Distributional mismatch between training and deployment
• Mitigation: Randomizing market parameters (volatility, spread, latency)
• Goal: Zero-shot transfer to unseen market conditions

A technique that trains a classifier to distinguish between training and test data distributions. When the classifier can reliably separate synthetic from real data, it signals that the simulator fails to capture critical market properties.

• Detects covariate shift in financial datasets
• Provides a quantitative metric for simulator fidelity
• Guides which parameters to randomize next
• Often implemented with gradient-boosted trees or simple neural networks

A set of consistent statistical properties observed across financial time series that synthetic data must replicate. Domain randomization ensures agents don't overfit to a single parameterization of these properties.

• Volatility clustering: Large changes follow large changes
• Fat-tail distributions: Extreme events more likely than normal
• Absence of autocorrelation in returns
• Leverage effect: Volatility rises when prices fall
• Gain/loss asymmetry: Drawdowns faster than run-ups

A reinforcement learning paradigm where multiple autonomous agents interact within a shared environment. Combined with domain randomization, MARL simulates the co-evolution of competing strategies under varying market regimes.

• Agents learn to adapt to adversarial opponents
• Randomizing agent populations prevents strategy monoculture
• Used to discover Nash Equilibrium strategies
• Critical for modeling market impact and liquidity dynamics

A self-exciting point process model where the arrival of an event increases the probability of future events. Used to simulate clustered order flow and trade arrivals with randomized intensity parameters.

• Captures endogenous market activity (trades triggering more trades)
• Randomizing decay rates and branching ratios creates diverse market microstructures
• Generates realistic bid-ask bounce and order book dynamics
• Complements GAN-based approaches for high-frequency simulation

A systematic bias where a trading strategy is excessively tailored to historical noise rather than true signal. Domain randomization mitigates this by training across thousands of plausible market scenarios rather than a single historical path.

• Symptom: Excellent in-sample performance, poor out-of-sample
• DR Solution: Agent must generalize across randomized environments
• Validation: Performance consistency across randomized test sets
• Often measured via deflated Sharpe ratio and probability of false discovery