Agent-Based Models & Market Simulation

The last course handed you one honest answer to “I only have a single market history.” You learned a generator — a GAN, a VAE, a diffusion model — that absorbs the distribution of your data and emits new paths that look like it. Powerful, but top-down and opaque: the model never explains why returns are fat-tailed, only that it can copy the fact, and a memorizing generator can leak your history straight back into your tests.

This course is the other answer, and it could not be more different in spirit. Instead of learning the market from the top down, you build it from the bottom up. You specify the agents — zero-intelligence traders who quote at random, market makers who post two-sided liquidity, momentum players who chase trends, value players who fade them, and the learned RL bots from Deep RL for Execution & Market Making — drop them into a matching engine, and let them trade. You never write down “the price.” The price emerges from their collective order flow. This is an agent-based model (ABM): a mechanistic, interpretable simulator where macro-scale market behavior is an output, not an assumption.

The headline result — the one that makes ABMs worth a whole course — is emergence: when you wire up heterogeneous agents and let them interact, the price path spontaneously grows the stylized facts you spent Time-Series Finance cataloguing. Fat tails, volatility clustering, the slow decay of absolute-return autocorrelation — nobody programs them in. They fall out of the feedback between trend-chasers and mean-reverters, exactly as they do in the real market. A generative model can reproduce fat tails because it saw them; an ABM can produce fat tails because it contains the mechanism that makes them.

This course assumes the two prerequisites cash out three habits. From Generative Models for Synthetic Market Data you carry the stylized-facts rubric any synthetic market must satisfy and — more importantly — the “did you fool yourself?” discipline of evaluation: a model that looks right is not the same as a model that is right. From Deep RL for Execution & Market Making you carry the agent’s-eye view of trading and the sim-to-real gap, because the most demanding thing you can do with an ABM is use it as the environment an RL agent trains against. We lean on both relentlessly, always asking the skeptic’s question: did this market structure really cause that, or did I just tune the model until it agreed with me? The arc:

• Why build a market from agents — the bottom-up vs top-down split restated: simulate a mechanism (this course) vs learn a generator (the last one). What emergence means, why interaction is the engine, and the honest trade-off — interpretable and counterfactual-ready, but slow, fiddly, and just as capable of fooling you as any black box.
• Zero-intelligence agents: the shocking baseline — Gode & Sunder’s discovery that traders quoting at random, constrained only by their budgets and the market’s rules, still push prices to near-perfect allocative efficiency. The lesson that reframes everything: a lot of what looks like “smart” market behavior is the institution, not the intelligence — and ZI traders are the baseline any fancier agent must beat.
• The agent zoo — the heterogeneous cast that makes a market interesting: market makers, momentum/trend followers, fundamental/value traders, noise traders, and learned RL agents. The lineage that built this field — the Santa Fe Artificial Stock Market and the modern ABIDES-style multi-agent limit-order-book simulators — and how you assemble a population.
• Emergent stylized facts — the payoff. How the interaction of trend-followers and value-traders manufactures volatility clustering and fat tails from nothing, why a market of identical or non-interacting agents stays boringly Gaussian, and how to read the emergence as a genuine explanation rather than a coincidence.
• Calibrating an agent-based model — the hard part. An ABM has knobs (how many of each agent, how aggressive, how much noise) and no likelihood you can just maximize. The simulation-based toolkit: method of simulated moments / simulated minimum distance, matching the stylized-facts vector, and the modern simulation-based-inference and surrogate approaches — plus why ABM calibration is brutally ill-posed.
• Validating without fooling yourself — the same gauntlet, pointed at a new target. In-sample fit is not validation; you need out-of-sample stylized facts, parameter identifiability, sensitivity analysis, and the honesty to admit that a model flexible enough to match everything has explained nothing. The ABM version of deflated Sharpe.
• What an ABM gives you that a backtest cannot — the killer apps. Counterfactual market structure (change the tick size, add a circuit breaker, slow everyone down) that no historical tape can answer; flash-crash and feedback-loop stress testing; and a non-stationary, reactive training ground for RL agents — a market that pushes back instead of replaying a fixed past. The closing contrast: ABM (bottom-up, mechanistic, interpretable) vs deep generative models (top-down, learned, opaque), and when to reach for each.

By the end you’ll be able to assemble a population of agents, watch real stylized facts emerge from their interaction, calibrate the model to data and validate it without lying to yourself, and wield it for the counterfactuals and stress tests the one historical path can never deliver. A graded final exam runs the whole discipline back at you, one locked question at a time.