Work

Projects & experiments.

Systematic strategies, market microstructure, derivatives pricing, and applied machine learning. The deep dives are interactive - every model on those pages runs live in your browser.

01Deep dives - quant methods, interactive
Market Microstructure01

Order Flow: From Poisson to Hawkes

A live limit order book with a trading ladder, driven by three models of order flow: homogeneous Poisson, a time-varying Poisson fitted to the intraday U-shape by thinning, and a self-exciting Hawkes process with MLE calibration.

L3 Tick DataPoisson / ThinningHawkesExplore →
Portfolio Allocation02

Markowitz in Practice

The efficient frontier built live on ten S&P 500 stocks - min-variance vs max-Sharpe out of sample - and Random Matrix Theory eigenvalue clipping to fix the covariance matrix that betrays the optimiser.

MarkowitzRMTLive S&P 500 dataExplore →
Computational Methods03

Monte Carlo & the Volatility Smile

Euler vs Milstein strong convergence measured live, Heston Monte Carlo with full-truncation, and implied-volatility smiles backed out by bisection - drag ρ and watch the equity skew appear.

Euler / MilsteinHestonImplied volExplore →
Numerical Optimization04

Optimization Under Constraints

Gradient descent with a live step-size stability limit, plus constrained solvers (projection, penalisation, Uzawa) on a 2-D problem - the KKT machinery behind Markowitz and support vector machines.

Gradient descentUzawaKKTExplore →
Derivatives05

Free Boundaries & Exotic Pricing

Pricing American and Bermudan options by solving the Black-Scholes PDE with finite differences - early-exercise boundaries, and the same machinery applied to Bermudan swaptions.

PDE / FDStochastic CalculusFixed IncomeExplore →
Machine Learning06

Deep Learning in Finance

An autoencoder trained live in your browser to compress the cross-section of returns and flag anomalous days, plus a tour of CNN roughness estimation, no-arbitrage pricing networks, LSTMs and GANs for synthetic markets.

AutoencodersCNN / LSTMGANsExplore →
Reinforcement Learning07

Q-Learning in a Maze

A tabular Q-learning agent trained live on a freshly generated maze: watch the value function flood backwards from the goal and the greedy policy snap onto the shortest path.

Q-learningε-greedyDynamic ProgrammingExplore →
02Featured
03Other projects
Geospatial / ML

Deforestation Detection

Automatic detection of illegal fires in primary forests using SAR imagery from the Copernicus programme.
PythonSARCopernicus
NLP

Insult Detector

NLP classification model for detecting insults in tweets, with feature engineering and model comparison.
PythonNLPscikit-learn