Available from May 2027
Aditya Bhatia
Financial engineer finishing an MS at Stevens with a 4.0, graduating May 2027. I build hedging models that know what trading costs, and I write the code that proves them. Looking for quant research, strats and derivatives roles.
GPA4.0MS Financial Engineering, Stevens
U.S. rank, IMC Prosperity 4164thTop 1% of 22,000+ teams
NeurIPS 2026 WorkshopSole authorDeep hedging, and why the batch size is part of the risk measure
Stevens 25 Under 25HonoreeNamed in 2026
IndustryIndustry Capstone Fellow, Brown Brothers Harriman, derivatives on GPU compute
Research, ongoingGraduate RA in Statistics under Prof. Choudur Lakshminarayan, leading four undergraduate RAs
GitHub ↗
abhatia10@stevens.edulinkedin.com/in/adityajbhatiagithub.com/Adi7710medium.com/@abhatia10
I don't trust a model until it has met the evidence.
So here's the oldest evidence in the field. Drop balls through 12 rows of pegs:
Chaos, averaged, looks suspiciously like a bell.Galton board · 12 rows · Binomial(12, ½) · 0 landed
Research
The batch size is the risk measure
Deep hedging trains a neural network to hedge an option by minimizing risk, one minibatch at a time. My paper shows that the batch size quietly changes what the network is minimizing.The Batch Size Is Part of the Risk Measure: Entropic Deep Hedging with an Explicit Sample Size. Sole author. NeurIPS 2026 Workshop poster, Atlanta, December 2026.
The fix. Put the batch size into the risk measure and the bias disappears. TRACE, my multi-agent hedger, then learns when to sit still.
In plain EnglishTrain a hedging model in batches and the batch size changes the target. I showed why, fixed it, and built a system that finds the best hedge for every level of trading cost. When trading gets expensive, it stops over-trading.
Try it
At a trading cost of λ = 25 bps the band is ±0.14 wide and the hedge trades 23 times a year instead of 252.
auto
drag across the chart to set the cost
trades this year 0 · cost follows the slider above
drag to rotate
RA work: risk when n meets d Ongoing
As a Graduate RA in Statistics under Prof. Choudur Lakshminarayan, since January 2026 and ongoing, I study why portfolio risk breaks for volatile, fat tailed assets. Across 50 energy portfolios I run GMV optimization, VaR at several confidence levels, normality tests on the tails and demonstrations of how unstable the covariance becomes. I also lead four undergraduate RAs and wrote their teaching guide.
Why it is hard: the Bahadur term
VaR is a quantile, and the Bahadur representation says exactly how wrong a sample quantile is:
ξ̂p = ξp + (p − Fn(ξp)) / f(ξp) + Rn
The remainder Rn dies off like n−3/4 up to log factors, so the middle term is the error: its standard deviation is √(p(1 − p)/n) / f(ξp). Two things blow it up at once.
Fat tails. Deep in the tail the density f(ξp) is tiny, so 1/f is huge. At the same volatility, a Student t with 3 degrees of freedom leaves a 99% VaR estimate 2.6× noisier than the normal model, from the same data.
n close to d. With d assets and n days, the sample covariance spreads its eigenvalues over [(1 − √(d/n))², (1 + √(d/n))²]. As n → d the smallest one goes to zero, S−1 explodes, GMV weights go extreme and out of sample variance inflates by roughly 1/(1 − d/n).
Put together, the number a risk desk wants most is the hardest one to estimate, built from the noisiest weights you can make.
auto
d = 50 assets · d/n 0.20 · smallest eigenvalue 0.31 · GMV variance ×1.25 · VaR error ×2.6 vs normal
Also on my desk
Deep Hedging Benchmark. One fixed protocol to ask whether learned hedgers beat classical transaction cost bands, and how much survives a regime shift. Stage 0 complete and audited, 95 tests passing. github.com/Adi7710/deep-hedging-benchmark ↗
The Fed Put, replicated. Cieslak and Vissing-Jorgensen (2021) rebuilt with NLP on FOMC text, matched to three decimals, under Prof. Victoria Li.
Modified Gerber Statistic. An independent implementation, in progress on WRDS data.
Projects
Built under σ > 0
Seven projects, each set by a market, a client or a clock. First, the market itself:
regime calm
drag to rotate
Paper traded on Kalshi
Prediction market prices overshoot, then drift back, so the first model I reached for was Ornstein Uhlenbeck, dX = θ(μ − X)dt + σ dW, with a fade when the price strays too far. Reversion speed θ = 3.0 a day, so a shock halves in 5.5 h. Enter at 1.5σ from the mean, exit at the mean.
Then I tested the obvious strategy on 749 settled Kalshi markets, and it lost money: contracts far from 50¢ are more accurate, not less. What worked was calibration. KalshiQuant, the public repo, has isotonic calibration fitted on the settlement tape (Brier 0.046), fee aware sizing, a Bernoulli copula CVaR in place of Gaussian VaR, vol scaled triple barrier exits, a Heston pricer for short dated binaries, a look ahead bias scanner and 160 passing tests. It runs against the live Kalshi feed in paper mode only; live trading is off by design.
github.com/Adi7710/kalshiquant ↗
fills 0 · paper P&L 0¢ · flat
grab the price and pull it
auto
Thesis-Delta Engine
Demoed at Freedom Capital Markets
An LLM pipeline that spots the moment a thesis changes.
IMC Prosperity 4
Finalist · 164th in the U.S. · top 1%
Algorithmic trading against 22,000+ teams. Made the finals.
Compute as an Asset Class
Brown Brothers Harriman · Capstone
Designing how GPU compute could trade like a commodity.
Vanguard ETF Challenge
Sharpe ranked · Fall 2026 · Team
Scored on Sharpe, so every rebalance is a bet on volatility.
Bodh
Hackathon · 3rd place
Money basics for students who never got them in school.
Orbit
SteelHacks · Engineering lead · 200+ commits
The honest planner for students. Your calendar says 13h 35m free; you have 9h 07m. Built in 24 hours with a team I met that morning.
github.com/Adi7710/orbit ↗
Also shipped: the CME Group University Trading Challenge, IMC Prosperity 3, Stevens Nexus, and a voice agent for blind users built solo in 24 hours.
Career
One path, so far
Most recent first. Scroll and the path draws itself.
Sep 2026 to PresentBrown Brothers Harriman
Industry Capstone Fellow, Innovation and Strategy
Jan 2026 to PresentStevens Institute of Technology
Graduate Research Assistant, Quantitative Finance
May 2026 to Aug 2026Stevens Institute of Technology
Summer Research, FOMC and the Fed Put
Feb 2025 to Aug 2025NeuraMinds AI
Machine Learning Analyst
Jan 2024 to Jun 2024Bhatia Graphica
Data Analyst
Sep 2023 to Dec 2023Saksham Valuers
Investment and Valuation Intern
Aug 2023 to Dec 2023Heramb Finserv
Options and Derivatives Modeling Intern
Story
It started in Mumbai
Many paths lead out of Mumbai. This one is mine.Brownian bridges, pinned at both ends
Grew up around printing presses.Bhatia Graphica, my family's flexo prepress business in Mumbai, nearly three decades old.
Studied finance. Priced options. Checked them against the market every single day.
No CS degree. Five years of YouTube and stubbornness. One paper, written alone.
Education & skills
The toolkit
- 2025 to 2027
- MS Financial Engineering, Stevens. GPA 4.0. Provost Master's Scholarship. Stevens 25 Under 25 honoree, 2026.
Stochastic Calculus · Shreve IIFE 680 Advanced DerivativesFA 541 Applied StatisticsDeep LearningFA 690 NLP
- Undergraduate
- BCom Finance with Distinction, Narsee Monjee, Mumbai.
- Code
- Python (NumPy, pandas, SciPy) and PyTorch, plus WRDS and the Anthropic API. CFA Level I candidate.
Skills, by area
Machine learning
- Regularized regression: ridge, lasso, elastic net
- Tree ensembles: random forests, gradient boosting
- Walk forward validation for time series, no look ahead
- Model selection and evaluation for client problems (NeuraMinds AI)
Deep learning
- PyTorch: custom training loops, autograd, GPU batching
- Deep hedging policy networks trained on simulated paths
- Risk measure losses: entropic risk, CVaR
- How minibatch size biases the objective (my NeurIPS 2026 Workshop paper)
- Transformers and embeddings for financial text (FA 690)
- Autograd, backprop and attention typed from a blank file: micrograd-to-frontier ↗
Agentic AI in finance
- TRACE: multi agent hedger, every decision checked against evidence
- LLM pipelines with RAG over filings and research
- Structured outputs and tool calling (Anthropic API)
- Thesis-Delta Engine: flags when an investment thesis changes
Statistics in finance
- VaR at multiple confidence levels across 50 energy portfolios
- Fat tail diagnostics: QQ plots, Jarque Bera, Student t fits
- Quantile estimation error via the Bahadur representation
- Co-movement with the Modified Gerber Statistic (WRDS)
- Event regressions on FOMC text (Fed Put replication)
Linear algebra and matrices
- Eigendecomposition and PCA of covariance matrices
- Random matrix theory: Marchenko Pastur spectra when n ≈ d
- Conditioning, inversion and why S⁻¹ blows up
- Cholesky factorization for correlated simulation
Stochastic calculus
- Itô’s lemma and Itô integrals (Shreve II)
- Girsanov and change of measure to risk neutral pricing
- Martingale pricing and Feynman Kac
- SDE solutions for GBM, OU and Vasicek
Core probability
- Conditional expectation and martingales (FE 543)
- Laws of large numbers and the central limit theorem
- Order statistics and sample quantiles
- Brownian motion: scaling, reflection, bridges
Applied statistics
- OLS with diagnostics: heteroskedasticity, autocorrelation
- Maximum likelihood and hypothesis testing (FA 541)
- Bootstrap and resampling for estimation error
- Normality and goodness of fit tests
Portfolio theory
- Mean variance and the efficient frontier
- Global minimum variance portfolios
- Sharpe ratio optimization (Vanguard ETF Challenge)
- Estimation error in weights and out of sample variance
Market microstructure and trading
- Limit order books, spreads and queue dynamics
- Market making with inventory control under position limits (IMC Prosperity 3 and 4)
- Mean reversion on prediction markets (Kalshi, paper traded)
- Transaction cost aware hedging: the no trade band
- Futures trading (CME Group University Trading Challenge)
Advanced derivatives
- Black Scholes, Greeks and delta gamma hedging (FE 680)
- Implied volatility surfaces: smile, skew, term structure
- Hedging under transaction costs
- Market structure and derivatives on GPU compute (BBH)
Computational methods
- Monte Carlo with antithetic and control variates
- Euler Maruyama discretization of SDEs
- Convex optimization for portfolio construction
- Vectorized simulation in NumPy and PyTorch; WRDS data pipelines
Stochastic models I work with
Each one, the equation and where I have used it.
Brownian motion
coredX = μ dt + σ dW
The building block. Brownian bridges draw the story chart on this page.
Geometric Brownian motion
builtdS = μS dt + σS dW
Black Scholes pricing and Greeks, and the simulated paths TRACE learns to hedge on.
Ornstein Uhlenbeck
builtdX = θ(μ − X) dt + σ dW
Mean reversion with half life ln 2 / θ. The engine behind my paper traded Kalshi strategy.
Vasicek short rate
builtdr = κ(θ − r) dt + σ dW
Gaussian, mean reverting rates with closed form bond prices. Built a Monte Carlo simulator for it.
Cox Ingersoll Ross
courseworkdr = κ(θ − r) dt + σ√r dW
Mean reverting rates that stay positive when 2κθ ≥ σ².
Heston stochastic volatility
courseworkdv = κ(θ − v) dt + ξ√v dW, with ρ between dW and the price
Why the smile exists: variance itself mean reverts and correlates with returns.