Senior Data Scientist

Job Description

We need someone who can build high-quality forecasting models for UK energy balancing markets — not a generalist who's touched a bit of everything, but a specialist who genuinely understands time series, knows how to extract signal from massive feature sets, and can produce reliable probabilistic forecasts.

You'll spend significant time on tasks like: engineering features from raw market data, selecting the most predictive subset from hundreds of thousands of candidates, building gradient boosting models that output well-calibrated prediction intervals, and rigorously validating everything to avoid the subtle leakage problems that plague time series work.

You won't be responsible for deployment — we have experienced DevOps for that. But you'll need to hand off models that are well-documented, reproducible, and actually work in production. If you find satisfaction in the craft of building models that hold up under scrutiny — rather than just hitting a metric on a test set — this role is for you.

Feature Engineering and Selection

• Engineer predictive features from energy market data (prices, volumes, grid conditions, weather, calendar effects)

• Work with feature sets in the hundreds of thousands — you'll need systematic approaches, not manual inspection

• Apply and evaluate feature selection methods (mRMR, importance-based selection, recursive elimination) to build parsimonious models

• Analyse feature importance and stability across time periods and market conditions

• Understand the domain well enough to create features that reflect how the balancing market actually works

Model Development

• Build gradient boosting models (XGBoost, LightGBM, CatBoost) for multi-horizon forecasting

• Produce probabilistic forecasts — prediction intervals, quantile regression, or distribution outputs — not just point estimates

• Handle class imbalances appropriately when the problem requires classification

• Design proper time series cross-validation schemes that respect temporal ordering

• Diagnose and fix target leakage — you should be able to explain why a 'too good' result is suspicious

Validation and Testing

• Test pipeline components using synthetic/artificial data where ground truth is known

• Validate that preprocessing steps (missing value imputation, outlier handling) don't introduce leakage

• Build confidence that models will generalise, not just interpolate

Experiment Tracking and Reproducibility

• Track experiments systematically (MLflow or similar)

• Maintain reproducible training pipelines with proper configuration management

• Document model decisions, hyperparameter choices, and validation results clearly

Domain Understanding

• Invest time learning UK energy balancing markets — BM units, settlement periods, system prices, imbalance dynamics

• Translate domain knowledge into model improvements (better features, appropriate loss functions, sensible constraints)

• Collaborate with colleagues who understand the data infrastructure and market context

Requirements

Must Have

• Deep time series experience — you understand why random CV splits fail for forecasting, how to handle multiple horizons, and the pitfalls of lookahead bias

• Strong feature engineering and selection skills — you've worked with high-dimensional feature sets and know multiple approaches to reduce them systematically

• Gradient boosting expertise — XGBoost, LightGBM, or CatBoost are your core tools; you understand their hyperparameters and when each matters

• Probabilistic forecasting ability — you can produce calibrated prediction intervals or quantile forecasts, not just point predictions

• Rigorous validation mindset — you're paranoid about leakage, you test your assumptions, and you don't trust results that seem too good

• Python fluency — clean, testable code; comfortable with pandas/Polars, scikit-learn, and the GBM libraries

• SQL competence — you can pull and reshape data from PostgreSQL without friction

• Clear communication — you document your work and can explain model behaviour to non-ML colleagues

Nice to Have

• Experience with MLflow, Hydra, Metaflow, or similar tooling for experiment tracking and pipeline management

• Polars experience (we're migrating some workloads from pandas)

• Background in energy, utilities, trading, or other domains with similar forecasting challenges

• Familiarity with UK energy markets, Elexon data, or grid balancing

• Experience with conformal prediction or other modern uncertainty quantification methods

Highly Desirable — Agentic AI Coding Experience

We value candidates who can build software using agentic AI coding systems. This is fundamentally different from using code completion tools or chat-based assistants.

What we're NOT looking for: - GitHub Copilot (code completion/autocomplete) - ChatGPT or similar chat interfaces for generating isolated code snippets - Any tool that only provides single-turn question/answer interactions

What we ARE looking for: Hands-on experience with agentic coding systems such as Claude Code, Codex (OpenAI's agentic coding tool), Open Code, or Cursor.

Ideal candidates will demonstrate:

- Breadth of experience — proficiency with at least 2 agentic systems (experience with only one is insufficient)

- End-to-end development — ability to design and build software from the ground up using these tools, not just generating isolated snippets

- Multi-agent orchestration — demonstrated experience orchestrating multiple agents using skills, tools, and agent coordination, not just one-shot problem solving

- Deep system knowledge — familiarity with hooks, permission systems, MCP (Model Context Protocol) servers, custom skills and tool definitions, and context management

Benefits

• Plenty of opportunities for learning and professional growth
• B2b contract with a paid vacation