Overview In this role you will design and train a purpose-built decoder-only Transformer for tokenized intraday market data, advancing a proprietary model aimed at next-token price movement predictions. You will work within a newly formed systematic equities pod to push state-of-the-art transformer architectures from first principles. The position involves heavy research with substantial compute, spanning tokenization, attention design, and low-latency inference for live trading. You'll collaborate across teams to integrate predictions into the trading pipeline and document results. This is a long-term, impact-driven research opportunity at scale.
Compensation / Benefits- base salary
- discretionary performance bonus
- comprehensive benefits
Responsibilities- Design and implement a custom decoder-only Transformer for financial time-series.
- Develop a novel tokenization scheme for intraday market data (price, volume, order flow, cross-sectional features).
- Build efficient, multi-GPU training pipelines in PyTorch with mixed-precision and gradient checkpointing.
- Design attention mechanisms suited to financial data, including temporal and cross-asset patterns.
- Create evaluation frameworks for next-token accuracy, signal quality, and trading performance.
- Optimize inference for low-latency deployment (quantization, KV-cache, speculative decoding).
- Conduct ablations to validate architectural choices and training methodology.
- Collaborate to integrate model predictions into the live trading pipeline.
- Document research methodology, experiments, and architectural decisions.
Key requirements- PhD in Machine Learning, Computer Science, Statistics, Applied Mathematics, or a related field with a focus on deep learning.
- Proven ability to implement Transformer architectures from scratch (not just fine-tuning).
- Deep understanding of attention, positional encodings, tokenization strategies, and training dynamics.
- Expert PyTorch skills including custom modules, training loops, mixed-precision, and multi-GPU training.
- Strong mathematical foundations in linear algebra, probability, optimization, and information theory.
- Experience training models at scale (100M+ parameters).
- Strong Python and C++ programming for performance-critical components.
- Self-directed researcher capable of executing a multi-month agenda.
- Familiarity with AI-assisted development tools (Cursor, Claude Code).
- self-directed
- collaborative
- strong communication of complex results
- custom Transformer implementation
- tokenization strategies for time-series
- PyTorch with mixed-precision training
