Ethan Jin
All projects

Market microstructure / Simulation

Adverse Selection Lab

What happens to a market maker when the trader on the other side knows more?

View on GitHub Read the research memo

The experiment

I built a synthetic market with one dealer, informed traders, and noise traders. Three quoting strategies face the same seeded scenarios, so their differences can be compared directly. The experiments vary informed order flow, volatility, spreads, and signal quality to study how each changes the dealer’s profit and risk.

Selected results

What the simulations show

01

Inventory control reduces risk.

Shifting quotes in response to inventory reduced exposure and drawdown in these experiments. The paired comparisons did not establish a statistically conclusive improvement in profit.

Inventory-skewed quotes show lower average absolute inventory and maximum drawdown than fixed quotes in the simulated inventory-control sweep.
Saved simulation results: 250 seeded scenarios per setting; error bars show 95% confidence intervals. Select the figure to enlarge.
02

A wider spread comes with a tradeoff.

Quoting farther from fair value changes the profit per trade, but fewer traders accept the quote. The spread sweep tracks fill rates alongside total profit and adverse-selection losses.

Simulation figure showing how dealer profit, fill rate, and adverse-selection loss vary as the quoted spread widens.
Saved spread-sweep results with 95% confidence intervals. These are outcomes in the synthetic market.
03

Adapting to order flow needs informative history.

Toxicity-adaptive quotes improved simulated profit when clustered trading made earlier fills informative. With one order opportunity per interval, there was no earlier same-interval fill to learn from, and the adaptive strategy matched the fixed strategy.

Comparison of fixed and toxicity-adaptive quoting across different numbers of order opportunities per interval.
Saved order-flow experiments. Strategies use the same scenario tapes; error bars show 95% confidence intervals.

Scope & limitations

A controlled model of a market.

These experiments isolate mechanisms in a fully synthetic setting. The model uses one dealer, one-unit trades, Gaussian shocks, and an observed current efficient value. It omits competing makers, queue position, cancellations, partial fills, fees, latency, and market impact.

The results do not establish real-world trading profitability. The repository includes the simulation code, experiment settings, accounting checks, and instructions for reproducing the saved results.

Explore the code and methodology