Neural Stochastic Agent-Based Limit Order Book Simulation: A Hybrid Methodology
Neural Stochastic Agent-Based Limit Order Book Simulation: A Hybrid Methodology
This paper introduces a hybrid neural stochastic agent-based model (NS-ABM) for limit order book (LOB) simulation, combining the strengths of agent-based models (ABMs) and stochastic models (SMs). The core innovation is a neural stochastic background trader (BT) that learns aggregate market behavior from historical LOB data using a state-dependent parallel neural Hawkes process (sd-PNHP). The BT is embedded into the ABIDES simulation platform, where it can interact with other trading agents, such as trend followers and value investors. The model is validated by demonstrating that the simulated LOB reproduces a comprehensive set of ten stylized facts observed in real markets, both when the BT runs in isolation and when it interacts with other agents. Additionally, the simulation exhibits order flow impact and financial herding behaviors consistent with empirical studies. The hybrid approach overcomes limitations of pure ABMs (lack of data grounding) and pure SMs (lack of interactivity), offering a realistic and interactive environment for testing trading strategies and studying market dynamics.
Highlights
- 1Proposes a novel hybrid NS-ABM that combines neural stochastic background trader with agent-based simulation for limit order book modeling.
- 2The background trader uses a state-dependent parallel neural Hawkes process (sd-PNHP) trained on historical LOB data to generate realistic order streams.
- 3Demonstrates that the simulated LOB reproduces ten stylized facts of real markets, both with the background trader alone and when interacting with trend and value agents.
- 4Shows that the hybrid model exhibits order flow impact and financial herding behaviors consistent with empirical observations.
Methods
- MState-dependent parallel neural Hawkes process (sd-PNHP) for learning event arrival patterns from historical LOB data.
- MAgent-based simulation using the ABIDES platform, incorporating background trader, trend agents (momentum and mean reversion), and value agents (zero intelligence and heuristic belief learning).
- MSobol's global sensitivity analysis to assess parameter influence on simulation outputs.
Results
- RThe background trader alone generates LOB dynamics that replicate ten stylized facts of real markets, including volatility clustering, fat-tailed returns, and autocorrelation of order flow.
- RWhen interacting with trend and value agents, the stylized facts persist, and the simulation exhibits realistic order flow impact and herding behavior.
- RSensitivity analysis shows that parameters related to order price and volume distributions have the most influence on simulation outcomes.
Analyze Paper
Generate insights from "Neural Stochastic Agent-Based Limit Order Book Simulation: A...".