Published work
from the frontier.

We introduce a diffusion-model-based approach to unsupervised regime detection in financial time series. Our method identifies latent structural breaks at hourly resolution with 91% out-of-sample accuracy, outperforming HMM and clustering baselines across six asset classes.

We revisit the VPIN metric at nanosecond granularity using FPGA-captured market data. At sub-millisecond resolution, classical toxicity measures break down; we propose an adaptive flow imbalance estimator that remains predictive of adverse selection even at extreme frequency.

We benchmark a family of transformer architectures against LSTM baselines on mid-price prediction at 10ms intervals. Sparse attention mechanisms with causal masking yield a 1.8× Sharpe improvement in out-of-sample testing across US equity markets.

We train a PPO agent to minimize implementation shortfall across liquid equity markets. The agent learns adaptive TWAP/VWAP-hybrid schedules that respond to intraday liquidity conditions, reducing market impact cost by 23% vs. static benchmarks.

We describe a hardware-software co-design for incremental model parameter updates within a live inference loop. Using custom FPGA logic and a reduced-precision gradient accumulator, we achieve model refresh cycles of under 400 nanoseconds with no interruption to inference throughput.

Using synchronized order book data from 11 venues, we characterize the speed and persistence of price dislocations across fragmented equity markets. We show that cross-venue price convergence follows a power-law decay with venue-specific exponents driven by co-location proximity.

We develop a multi-agent simulation environment where learned strategies can be stress-tested against adversarial counterparts. The framework surfaces strategy fragility invisible to historical backtests, and has informed the retirement of three live strategies since 2024.