Quant jobs / Millennium

ML Research Scientist – Deep Learning & Transformer Architectures

A long-horizon research role building a purpose-made Transformer for tokenised intraday market data, trained from scratch. It sits inside a new systematic equities pod with large compute.

What you’d work on

Our reading of the posting: what a new hire in this role would most likely spend their first months on.

Design a custom decoder-only Transformer for financial time series
Devise tokenisation for price, volume and order flow
Build mixed-precision multi-GPU training pipelines
Run ablation studies and evaluate signal quality
Optimise inference for low-latency deployment

What they ask for

Required

PhD in ML or a related field
Transformers implemented from scratch
Expert PyTorch
Linear algebra, probability, optimisation, information theory
Training models of 100M+ parameters
Python and C++
Self-directed research planning

Preferred

Deep learning on financial time series
Tokenisation of continuous data
Publications at top ML venues
Market microstructure
Model compression and inference optimisation

Education: PhD in machine learning, computer science, statistics, applied maths or similar with deep learning focus
Experience: Large-scale model training experience; research record in deep learning
Languages and tools: Python
C++
PyTorch
Markets: equities

How to prepare

Implement a small decoder-only Transformer from scratch in PyTorch
Try tokenising limit order book events and predict next moves
Learn mixed-precision and multi-GPU training
Benchmark against a simple baseline to see how little signal there is

What to study

The NLP and ML Quant Researcher roadmap sets out the stages for this kind of role, with free resources for each.

Open the roadmap

Apply

Read the full posting on Millennium’s site before applying: requirements, location and pay can change after we check them.

View the posting on Millennium’s site ↗

SelfMadeQuant wrote this summary from the employer’s public posting, last checked 6 November 2026. SelfMadeQuant isn’t affiliated with Millennium and doesn’t handle applications.