Every model is a bet on what you can safely ignore. I spent three years at Nomura watching those bets meet production, and I'm finishing an MS in Financial Engineering at NYU Tandon to make better ones. I lag every input to the date it was published, test before I trust, and publish what didn't work.
One profile, four seats. Research, trading, development and risk draw on the same habits: careful data, honest statistics and code that is tested before it is trusted. Here is the evidence for each, with the résumé written for it.
Cross-sectional alpha on 495 names, scored with Newey–West inference and false-discovery control. Purged walk-forward learning. A look-ahead found in a published regime backtest, then measured. Every t-statistic here is a claim that had to survive its own robustness check.
Almgren–Chriss calibrated to a live Deribit book and stress-tested against a learning agent. SABR fitted to the BTC volatility smile. A Deribit options market maker in progress, with latency-aware fills, delta hedging and P&L after fees.
A zero-allocation C++20 matching engine clearing ~21M messages a second. An options library that prices three independent ways and agrees to within 0.6%. Python and SQL pipelines at Nomura that removed 50+ hours of manual work a month.
Three years owning VaR and sensitivity models across equities, rates, FX and commodities at Nomura, with Basel 2.5 and FRTB submissions reconciled 100% to front-office systems. Since then, a historical CVA that measures wrong-way risk by experiment rather than assuming it.
A C++20 engine that replays recorded Deribit order books, runs several quoting strategies side by side, simulates fills with latency, hedges delta on the BTC perpetual, and reports P&L after fees and adverse selection. The same engine then runs live on paper. If a strategy loses money after costs, that is the result. Three milestones are done: the engine replays a full recorded day, 15.9M events, in 120 seconds, and per-expiry SABR fits on recorded data (β = 0.5, chosen over 1.0 across 98,799 fits) price every quote.
Every figure below is reproducible from its repository, covered by tests in CI, and reported with the assumption that could break it. Where a published effect failed to replicate, I say so. That is the standard a desk should expect: a number that can be defended in front of a risk committee, not one that only survives a slide.
Eight cross-sectional signals on 495 S&P 500 names, 2010–2026, six built from prices and two from the tone of 15,575 SEC filings, are scored by Spearman rank information coefficient with Newey–West t-statistics, then Benjamini–Hochberg across all eight tests. One survives: Amihud illiquidity. Low volatility and the lottery signals come out with the wrong sign.
So the obvious move, z-score everything and average, loses 7.8% a year at a Sharpe of −0.57. Learning the combination with LightGBM under purged walk-forward CV with a one-month embargo, charging 10 bps a side, reaches 0.79 out of sample: within a quarter-point of the best single signal at two-thirds of its drawdown.
| Signal | Mean IC | NW t | Hit | BH q |
|---|---|---|---|---|
| amihud | 0.032 | 5.32 | 62% | 0.000 |
| tone_chg_naive | 0.007 | 1.37 | 57% | 0.46 |
| mom_12_1 | 0.008 | 0.69 | 54% | 0.60 |
| strev_1m | 0.006 | 0.64 | 48% | 0.60 |
| tone_chg | 0.003 | 0.46 | 55% | 0.65 |
| skew_120d | −0.004 | −0.70 | 46% | 0.60 |
| max5 | −0.015 | −1.16 | 48% | 0.49 |
| lowvol_60d | −0.023 | −1.50 | 47% | 0.46 |
| Portfolio, net | Ann. ret | Vol | Sharpe | Max DD |
|---|---|---|---|---|
| amihud | 8.5% | 8.2% | 1.03 | −13.5% |
| ml_combo | 7.3% | 9.3% | 0.79 | −9.1% |
| mom_12_1 | 2.0% | 16.0% | 0.12 | −36.7% |
| equal_weight | −7.8% | 13.7% | −0.57 | −69.1% |
Calibrated to a live Deribit BTC-PERPETUAL book, with volatility, volume and the touch spread taken from the feed and impact scaled as 1/ADV, Almgren–Chriss takes 38% off TWAP's cost risk and pays for it. On a 10%-of-ADV order, $49.0M over five days on the October 2, 2026 book, that is 7.5 bps of expected impact against 4.7.
A policy-gradient agent trained on 207,361 real five-minute bars, cut into 5,752 three-hour episodes and split chronologically, sees only inputs estimated from bars that closed before each episode opened. It was never shown the closed form and converged to it anyway: 0.060 bps apart out of sample, paired t = 0.10.
| Out of sample, bps | Shortfall | Risk | Objective | σ / episode |
|---|---|---|---|---|
| TWAP | 1.499 | 0.047 | 1.546 | 49.2 |
| Almgren–Chriss | 1.623 | 0.030 | 1.653 | 39.8 |
| REINFORCE agent | 1.686 | 0.026 | 1.713 | 33.8 |
Macro regimes from 120+ FRED-MD indicators, reduced by PCA and clustered by two-step K-Means, drive an MSCI sector allocation with every macro input lagged by its publication delay. The published backtest fitted the regime detector once on 1959–2023 data, so every month it traded before 2024 carried a label fitted with hindsight. Refitting the detector every January, on only the data available then, takes the headline Ridge book from a Sharpe of 0.65 to 0.35, below equal weight, while the simpler regime-mean forecaster holds up as the best long-short book in both universes: 0.77 in developed markets and 0.73 in emerging.
Position sizing is a convex program with trading costs inside the objective: an L1 turnover penalty, a Cholesky second-order-cone risk term and a quadratic impact term, re-solved monthly over a receding horizon. The impact coefficient is assumed, not fitted, so the useful test is sweeping it. Under walk-forward regimes the multi-period book holds a net Sharpe between 0.50 and 0.70 from 0 to 100 bps of linear cost, while the frictionless book falls to −0.28.
A discount curve bootstrapped from FRED Treasury constant-maturity par yields marks a ten-swap, $255M book from 2 to 30 years. Every par instrument reprices to under 1e-9. Daily P&L splits into carry, roll-down and level, slope and curvature moves through key-rate durations, with convexity booked as an explicit residual so the parts reconcile to a full revaluation.
The same curves then drive counterparty exposure: 8,478 joint daily observations of the curve and Moody's credit spreads (1986–2026), block-bootstrapped into 20,000 thirty-year paths. Peak EPE $2.35M, peak PFE $13.5M. Wrong-way risk was priced by experiment: resampling rates and credit from the same days keeps their −0.38 correlation, resampling them from different days destroys it. The coupling is worth $89,553, 16.9% of CVA, and both runs share their curve paths bit-for-bit, so the difference cannot be Monte Carlo noise.
A dependency-free C++20 engine values the same contract three independent ways: closed-form Black–Scholes with analytic Greeks, Monte Carlo with exact terminal sampling, and a Crank–Nicolson finite-difference solver on a 400×400 grid — Thomas algorithm for Europeans, projected SOR on the early-exercise complementarity problem for Americans.
They agree to roughly 0.6% on an at-the-money call, pinned as regression tests. On barriers they don't, and that disagreement is the interesting part.
| Contract | Black–Scholes | Monte Carlo | Crank–Nicolson |
|---|---|---|---|
| European call | 10.4506 | 10.4794 | 10.4481 |
| European put | 5.5735 | 5.5616 | 5.5711 |
| American put | — | — | 6.0875 |
| Up-and-out call | — | 3.5162 | 3.3326 |
A deterministic limit order book and matching engine in C++20 enforces strict price–time priority. Resting orders sit in intrusive doubly-linked lists per price level, so cancelling is a pointer splice rather than a search. A pre-warmed memory pool means the matching loop never allocates — which is what keeps the tail flat, not just the mean.
Prices are fixed-point unsigned integers with no floating point on the hot path, because float comparison in a price-priority book is a correctness defect rather than a slowdown.
| Operation | Cost | Why |
|---|---|---|
| Add — existing level | O(1) avg | hash + tail append |
| Add — new level | O(log L) | ordered-map insert |
| Cancel | O(1) avg | intrusive splice |
| Execute at best | O(1) / fill | FIFO head pop |
| Best bid / ask | O(1) | map front |
The calibration recovers α, ρ and ν to within 0.01 of the truth under 30bp of injected quote noise. On Deribit's 18SEP26 expiry it fits 21 strikes at 0.43 vol points RMSE, with the worst fit in the 95,000-strike wing.
There are no labels, so it ranks rather than classifies. Isolation Forest and ECOD over 27 order-flow features agree on 76% of their top-5% flags; the ensemble exists for the other quarter.
Across WTI and Henry Hub settlements from EIA (1990–2024), the detector flags 112 events, 79 of them corroborated by an Isolation Forest. It had been blind to the −$37.63 WTI print because a log transform turned it into NaN.
Five crypto perpetuals give 10 Engle–Granger tests, run on the first 60% of the sample only and put through Benjamini–Hochberg. Two pairs clear a raw p < 0.05 and none survives the correction, and the raw survivors change with the training window. The honest result is that there is nothing to trade.
It scans Polymarket multi-outcome events whose prices don't sum to 1, net of the quoted spread. A completeness bug once made 32 live legs of a 71-leg event, priced at 0.613, look like a 32% riskless arbitrage; two tests now pin the fix. On the October 2, 2026 snapshot, 8 of 200 events stay positive after the spread, the largest 1.7%.
Trees, ensembles, SVM, MLP, LSTM and PCA, all written by hand in NumPy. Logistic regression, linear SVM and a decision tree match scikit-learn's test accuracy to four decimals; the random forest does not (0.937 against 0.958), and the README says so.