I study where and when information enters prices, and what it costs the market maker who is quoting.

Order-level liquidity around the EIA releases

A pre-registered study of 60 EIA inventory releases (March to October 2026, 440 million CME order-book records), each against four matched control days, with the analysis frozen before the data were opened. After the release, depth within 5 ticks of the mid stays depleted for the equivalent of an extra 34 seconds in crude and 48 seconds in natural gas of fully missing depth over the first five minutes. Before the release, the book thins because fewer new orders arrive in crude (partly a general 10:30 effect) and more orders execute in gas, not because traders cancel more, contrary to my initial hypothesis. Failed placebos are reported alongside.

Status. Confirmatory analysis done and independently replicated. Out-of-sample replication on releases from mid-October 2026 in progress.

How wrong are backtest queue models?

Six open-source queue-position models that backtesters run on price-level data, scored against 57.7 million real resting orders reconstructed from CME order-by-order data (141 crude and 125 natural-gas trading days) under a frozen, pre-registered protocol. Mean fill-time error ranges from 0.31 to 1.07 seconds in crude and from 0.68 to 1.98 seconds in gas. After the gas storage report two models look more accurate, but mostly because more orders fill, and faster, not because they track the queue better. Built on mbo-queue, my C++20 and Python library that rebuilds CME FIFO queues from market-by-order data.

Status. Confirmatory analysis done, independently replicated and adversarially reviewed.

Crude vs Micro WTI: which quotes move first

On 66 pre-registered days of CME order-by-order data, crude futures quotes lead Micro WTI below one millisecond on every day: within 1 ms, a crude quote move is followed by a same-direction micro move on net 24% of the time, against 12% the other way. Micro WTI resting orders filled during an adverse crude move mark out about 0.6 ticks worse, the cost of being filled during a cross-product move.

Status. Done; independently replicated and adversarially reviewed.

WH Trading, University of Chicago Project Lab

Quantitative research: forecasting realized volatility across horizons in futures markets against implied volatility, classifying momentum and mean-reversion regimes, and testing whether regime-aware delta hedging beats fixed-threshold rules on P&L, risk and turnover.

Quantitative Modeling Intern, Endless Moments

Revenue and expense forecasts with ARIMA and LSTM models, about 15% lower out-of-sample RMSE than the baseline, validated walk-forward over two to three years.

Quantitative Research Intern, Morrison School of Agribusiness (USDA-funded)

Jacobian-corrected Box-Cox likelihood profiles that reproduce OLS exactly at λ = 1, in Python and Stata/Mata, and a parallelized estimation pipeline on more than 100 million forest-visitation records that cut runtime from about 10 hours to 3.

Optiver × UChicago Explaining Markets

Ranked second globally. A live earnings-reaction model explained an extra 0.150 of R² beyond the earnings surprise on 517 events (95% CI 0.10–0.20), 27% above the strongest LLM baseline. Every model change had to pass out-of-sample replication, and the live LLM pipeline processed 584 of 584 events with no errors.

IMC Prosperity 4

256th of 18,803 worldwide (top 1.4%). Market making around Ornstein–Uhlenbeck fair values, with quotes skewed on book imbalance to limit adverse selection; on a 10-strike call chain the edge was only near the money.

C++ limit order book and market-making simulator

A price-time-priority matching engine with partial fills, cancel/replace and self-trade prevention, at 125 ns median latency. On a synthetic deep-queue stream of 1M events it produces exactly the same fills as Liquibook at 2.4 to 2.8 times its throughput (on a near-empty book Liquibook is faster). Avellaneda–Stoikov backtests show quote fills collapsing from about 60–74% to under 2% under ITCH-calibrated flow, because executions are sparse.

Honors thesis: Bayesian sequential decision-making under regime switches

Re-run as a pre-registered, independently replicated simulation study: neither the volatility feature nor the CUSUM trigger speeds regime detection at matched false-alarm rates, but the volatility feature cuts false alarms from 26.4 to 2.1 per 1,000 steps at the thesis threshold, at the cost of a longer delay.

Cross-venue price discovery in BTC-USD

Coinbase leads Kraken at short horizons: Granger-significant after Benjamini–Hochberg correction in 6 of 9 usable sessions, against 1 of 9 in reverse, on quotes aligned to a 100 ms grid under a pre-specified protocol.

SPY volatility surface research

On a filtered set of 273 liquid out-of-the-money quotes across four expiries, median SABR calibration error is 0.0060 (0.0143 on the full surface), and SABR beats Heston on every expiry. One usable snapshot so far, which the repository states plainly.