B.A. Mathematics & Computer Science · NYU 2026

Brayden K. Mi

Researcher at Trexquant. I build research systems that are honest about costs, fills, and what the data can't tell you.

Projects Research Résumé Email

Experience

May 2025 – Jan 2026 · New York
Commodity option pricing, VaR under stress, and the data pipelines underneath.
- Statistical models for commodity option pricing from vol surface dynamics and forward curves.
- Scenario analysis and delta-gamma / full-revaluation VaR under stressed markets.
- Python/SQL pipelines for macro and options data; 42% faster than the previous query methods on large calibration sweeps.
Sep 2024 – Sep 2025 · NYU
Teaching assistant for Data Structures.
- Office hours and problem sets on recursion, complexity and object-oriented design.
May – Aug 2024 · Houston
Deep learning for medical image segmentation; co-author on the MIST preprint and first author on the RSNA abstract.
- PyTorch segmentation pipelines for biliary ducts: mean Dice 0.83 vs 0.62 for nnU-Net (+33.6%) on held-out cases.
- New loss functions for the MIST toolkit, including clDice; statistical modeling on noisy clinical data.
Sep 2023 – May 2024 · NYU
Volatility modeling and backtesting tools.
- Asset price and volatility modeling with ARIMA, LSTM and stochastic simulation.

Education

Projects

Kalshi Backtester · featuredPython · 2026
Monte Carlo Portfolio SimulatorFastAPI · React
Solar Generation ForecastingTime series
LSTM Equity SignalsPyTorch

Research

MIST: Medical Imaging SegmentationCo-author · arXiv 2024
Biliary Segmentation for PTBDFirst author · RSNA 2024
Reverse-Time Migration for USCTSole author · high school

Contact

braykzmi@gmail.com

GitHub LinkedIn Résumé

Project

Kalshi Backtester

Python 3.11 · asyncio · Parquet · tested for lookahead, synthetic recovery and live/backtest parity

A research and execution framework for systematic strategies on Kalshi event contracts: rules-based signals run across a universe of recurring markets, evaluated statistically, combined into a portfolio, and executed with realistic costs. It is built correctness first, because the failure that matters in a backtester isn't a crash. It's a number that looks right and isn't.

Architecture

Data flows one way; alpha code cannot reach settlement or future bars.

Order book & fill modeling

For resting orders, the engine replays each taker order's prints through the recorded book plus our own orders, in price order. Volume that printed at a level was resting before us, so it fills first. Our order fills only if the taker clearly went through it. Step through it:

A recorded L2 snapshot of the YES asks.

There are two execution paths. A bar engine models latency in bars, caps size at a fraction of each bar's volume, and fills limit orders only on trade-through. The L2 tape replay above is queue-aware and pessimistic. Both share one fee model, charged per fill and rounded up to the cent, with maker fees looked up per series. A parity test holds the live engine's replay to the research backtest: the same order intents, and P&L within 1%.

Not fooling yourself

Lookahead tests: settlement and future bars are unreachable from signal code, and future external data can't change past outputs.
Synthetic recovery: on data with a known generating process, the engine books exactly the P&L that process implies.
Walk-forward evaluation, candidates registered before they are run, and a holdout spent once with its pass rule fixed in advance.

Known limitations

No adverse-selection model: a passive order fills whenever price trades through it, which flatters passive strategies. Maker fees are charged in full partly for that reason.
In the bar engine, queue position is a blunt proxy (a strict trade-through), not a model.
Public prints include block and crossed volume that a small order could not have taken.

Status

A small set of alphas is in paper trading on the live feed. Strategy details aren't public.

Project

Solar Generation Forecasting

Time series · Seasonal ANN · Regression

GitHub repo ↗

Forecasting monthly U.S. solar generation, utility-scale and small-scale, from U.S. Energy Information Administration data for January 2022 to October 2024. The series is dominated by a strong 12-month cycle riding on fast capacity growth, so the question was which models capture both without overfitting 34 months of data.

Approach

Six models, from simplest to most flexible: two sinusoidal regressions (a seasonal sine around an exponential or logistic midline), SARIMAX, a plain feed-forward neural network, a time-lagged neural network, and a seasonal neural network (SANN) that sees the same month in prior years. The series was supplemented with a growth rate from a regression on installed solar capacity and with monthly total solar irradiance.

Results

Model	RMSE	MAE
Sinusoidal exponential-midline regression	889.8	748.7
Sinusoidal logistic regression	3354.0	2628.7
SARIMAX	1021.9	880.2
Artificial neural network	3294.4	2828.0
Time-lagged neural network	1360.3	1141.3
Seasonal artificial neural network (SANN)	817.2	637.8

The SANN had the lowest RMSE and MAE and held up across training runs, so it was the recommended model. The exponential-midline sinusoid came close with a far simpler model, which makes it the better choice where simplicity matters, though it can't use exogenous inputs.

Limitations

Only 34 monthly observations, too few for the flexible models to learn long-run trend and seasonality reliably.
Weather, economic and policy variables were left out; the forecasts assume history repeats without a structural break.
Trained on a laptop, which limited features, epochs and cross-validation folds.

Full reportOpen ↗

Research

Automated Biliary Segmentation for PTBD

RSNA 2024 abstract · first author · MD Anderson

B. Mi, R. Kempen, K. Pal, A. Celaya, I. Paolucci, R. A. Sheth, E. Y. Lin, A. L. Tam, D. Fuentes. Automated biliary segmentation for percutaneous transhepatic biliary drainage: a comparative study of neural network models. RSNA 2024.

Patients with obstructive jaundice often can't continue chemotherapy until their bilirubin comes down, and percutaneous transhepatic biliary drainage (PTBD) is the usual way to get there. But there are no clear guidelines on which patients, by biliary anatomy, benefit most. Quantifying how dilated the bile ducts are could help select them. This study built an automated segmentation model for that. I was first author on the RSNA abstract.

Data and models

A single-institution retrospective set of PTBD patients from January 2016 to May 2024 with pre- and post-procedure CT, giving 89 cases with ground-truth segmentations from an interventional radiologist (56 training, 33 validation). Five models were trained with identical settings: nnU-Net and four MIST variants differing in loss (Dice cross-entropy, and clDice with constant, linear and cosine schedules). Mean Dice scores were compared with ANOVA.

Results

Model	Dice
MIST, Dice cross-entropy	0.8379
MIST, clDice constant	0.8313
MIST, clDice cosine	0.8307
MIST, clDice linear	0.8206
nnU-Net	0.6215

The MIST models scored well above nnU-Net, a significant difference by ANOVA (p = 0.01).

My part

I led the study: built the models, ran the comparison and the statistics, and wrote the abstract. Ground-truth segmentations were drawn by an interventional radiologist.

Limitations

One institution and 89 cases, with a single annotator, so the results still need external validation before they can guide patient selection.

AbstractOpen ↗

Dice scores by model, with an automated segmentation exampleOpen ↗

Research

Reverse-Time Migration for Ultrasound Tomography

High school research project · St. John's School · sole author

B. K. Mi. Adapting an accurate reverse-time migration method to medical ultrasound imaging. High school research project, St. John's School.

Reverse-time migration (RTM) is how the oil industry images the subsurface: propagate the recorded wavefield backward in time, the source wavelet forward, and apply an imaging condition at each step. It handles two-way propagation, multiple arrivals and complex velocity structure far better than the ray-tracing methods common in medical ultrasound. Seismic imaging and ultrasound computed tomography (USCT) share the same physics, so this paper adapts RTM to breast USCT, where the goal is to image tumors by their velocity contrast with healthy tissue.

Method

The wavefield is extrapolated with a phase-shift-plus-interpolation (PSPI) propagator: exact phase shifts for several reference velocities, interpolated in the frequency-space domain to the local velocity. The main change from the seismic setting is geometry. Seismic data is recorded on a flat surface, while USCT places sources and receivers on a ring around the organ, so the propagators start from a circular datum.

Synthetic data came from finite-difference modeling with a 9-point Laplacian on a 0.25 m × 0.25 m model at 0.5 mm sampling, with a 500 kHz wavelet, using a synthetic breast (1600 m/s tissue, 1700 m/s tumors, in 1500 m/s fluid) and then a velocity model of a real cancerous breast.

Results

Circular acquisition gave better illumination than a flat datum, especially along the diagonals.
Shot density matters: the 4-shot image shows significant migration noise, while the 72-shot image is far cleaner, because stacking suppresses unfocused energy.
On the real breast model, the tumor is imaged even with 4 shots and clearly resolved with 72.

Limitations

All data is synthetic and 2D. RTM images impedance boundaries, so sharp contrasts like tumors image well but gradual tissue transitions do not, and it needs an accurate velocity model from tomography first.

My part

A high school research project at St. John's School. I was the sole author.