PhD Researcher · The University of Tokyo
Ísak Pétursson
I build learning systems that forecast and decide under uncertainty, when the outcome is random and the rules keep changing.
My research covers forecasting full return distributions, from one minute to one year ahead, and reinforcement learning where the set of legal actions changes every turn. Before my PhD I spent four years in industry as a data scientist, building ML/NLP systems, data pipelines and scraping infrastructure.
Open to quant and ML research roles: internships, or full-time for the right fit. Remote or on-site, anywhere.
Selected work
The code for these projects is private while the work is in progress. Each page shows the results with interactive figures, and I'm happy to walk through the code on request.
qCast
Distribution prediction: forecasts the full return distribution for 98 assets (equities, FX, futures, crypto) at horizons from 1 minute to 1 year. A neural network supplies the shape of each shock and GJR-GARCH how its scale evolves.
Lowest mean pinball regret across all 10 frequency × period cells (0.79% vs 1.63% for the next best), and the lowest VaR error, against six benchmarks from the literature. preliminary
qStorm
Turns a return-distribution forecast into a local volatility and solves the pricing equation, so it prices options without calibrating to any quotes. Skew and fat tails come from the forecast alone.
Median IV error of 2.20 vol points against 3.93 for the best quote-free rival, on 768k held-out quotes. Three quotes are enough to rebuild a chain to 0.18 vol points. preliminary
EidolonZero
Extends MuZero-style planning to large action spaces that depend on the state. Instead of masking a fixed action list, it learns representations of whichever actions are legal.
Search with the learned model beats a published deep-RL baseline on 3 of 4 job-shop scheduling sizes, and lifts Morpha win rates on the dynamic-card decks from 42% to 53%. preliminary
Morpha
A card-game benchmark built to isolate one problem: action spaces that change with the state. It compares static-masked and dynamic-action PPO, DQN and AlphaZero.
Playable in your browser: the same Python engine the agents train on, running client-side.
Chaos into Order (LEC)
A physics-informed network that learns the expected solution of a linear stochastic PDE directly by sampling its noise, with no discretization grid. The extension handles noise that multiplies the equation's coefficients, not just additive forcing.
With multiplicative noise: 0.3% error where the unmodified method is off by 60–70%, and under 1% up to six dimensions for Brownian and OU noise. Failures are flagged, not hidden. preliminary
Publications
- Chaos into Order: Neural Framework for Expected Value Estimation of Linear SPDEs arXiv:2502.03670
- Generalized Distribution Prediction for Asset Returns arXiv:2410.23296
Experience
- 2023 – 2026Data Scientist, Origo Reykjavík
Designed and built ML/NLP systems end to end, including a retrieval-augmented generation system for working with documents on limited GPU hardware, which was well received by internal users. Also built internal tools and data pipelines connecting the company's systems.
- 2020 – 2022Data Scientist, Data Science Services Vienna
Built web-scraping infrastructure for automated data acquisition, and ML models for document classification.
Education
- 2025 – 2028PhD, Advanced Interdisciplinary Sciences The University of Tokyo
- 2023 – 2025M.Sc. Data Science Reykjavik University
GPA 9.23/10 · Alan Turing Award
- 2021 – 2023B.Sc. Business Administration (Finance) Mercy University, NY
GPA 4.0/4.0
Off the clock
Icelandic Yu-Gi-Oh! National Champion (2024). It's why my RL benchmark is a card game.