ENV-FIN-103 · Quantitative Research
Systematic Alpha Research. Produce statistically sound, reproducible research, not strategies that merely backtest well.
The quantitative research lifecycle: validate market data, reconstruct factors, implement and backtest strategies, test robustness and document conclusions, with data vintages, corporate actions, survivorship bias, transaction costs and regime changes built in.
Overview. What the environment trains, and the work it's built around.
Core competency
Produce statistically sound, reproducible research, not strategies that merely backtest well.
- Building factors from point-in-time data without look-ahead
- Translating a strategy specification into correct, reproducible code
- Validating market data and handling corporate actions
- Testing robustness out of sample and adjusting for multiple testing
- Reporting conclusions the evidence supports
Built around the work of
Quantitative Researcher · Quantitative Developer · Data Scientist
Simulated systems
- Market-data service
- Research workspace
- Backtesting engine
- Experiment registry
Sample task
A factor research notebook reports unusually strong returns over its historical test period. Review the supporting datasets, identify any information that would not have been available at the relevant decision dates, and reproduce the strategy using valid point-in-time inputs. Deliver corrected factor values, a revised backtest and a report explaining material differences from the original results.
Specification. The environment, field by field.
- product_type:
- Executable RL training and evaluation environment
- category:
- Quantitative Research · Agentic Coding · Financial ML
- commercial_model:
- Environment development and validated task packages
- environment_id:
- ENV-FIN-103
- target_roles:
- Quantitative Researcher · Quantitative Developer · Data Scientist
- core_competencies:
- Building factors from point-in-time data without look-ahead
- Translating a strategy specification into correct, reproducible code
- Validating market data and handling corporate actions
- Testing robustness out of sample and adjusting for multiple testing
- Reporting conclusions the evidence supports
- data:
- Synthetic equity universe with daily prices, corporate actions, delistings and fundamentals stored by vintage. Licensed historical market data can be added where an engagement allows it.
- simulated_systems:
- Market-data service · Research workspace · Backtesting engine · Experiment registry
- interfaces:
- Structured market-data API · Sandboxed Python with common scientific libraries · Backtesting engine · Notebook inspection
- interaction:
- Stateful, multistep tool use with code execution. The agent submits code, data and a report, and the verifier reruns them.
- state:
- Versioned market data and fundamentals by vintage, a research workspace and an experiment registry. Each episode starts from an isolated, seeded snapshot.
- disturbances:
- Look-ahead in fundamentals, missing delisted securities, unadjusted splits, omitted dividends and costs, overlapping train and test periods, and over-searched parameters.
- task_families:
- TASK-FIN-103-01 · Point-in-time factor reconstruction & leakage audit
- TASK-FIN-103-02 · Strategy implementation & backtest repair
- TASK-FIN-103-03 · Market data integrity & corporate action handling
- TASK-FIN-103-04 · Strategy robustness & overfitting audit
- evaluation:
- Submitted code is rerun on held-out data. Point-in-time compliance, corporate-action handling, split integrity and the injected errors are checked deterministically.
- reward:
- Deterministic checks on methodology and reproducibility, plus rubric scoring of the report. Returns never raise the reward on their own.
- training_use:
- RL training, held-out evaluation and SFT on reference research notebooks.
- deployment:
- Containerized, isolated execution with sandboxed code, and a customer-controlled deployment option.
- licensing:
- Commercial use rights. Exclusivity, ownership and any third-party data licenses are agreed for each engagement.
Workflow coverage. What agents do, and what the verifier checks.
Workflow
Factor construction
Example tasks
Rebuild a fundamental factor using filing dates and vintages.
What is verified
No input is used before it was available.
Workflow
Strategy implementation
Example tasks
Implement a specified strategy and fix a backtest's bugs.
What is verified
Code matches the specification, costs are modelled and results reproduce.
Workflow
Data integrity
Example tasks
Validate a price history and handle splits, dividends and delistings.
What is verified
Adjusted series and the universe match the corporate-action record.
Workflow
Robustness review
Example tasks
Audit a strategy for overfitting and leakage.
What is verified
Out-of-sample design, purging and multiple-testing adjustments are correct.
Task families. Each family has its own datasheet: prompt, verification criteria and failure modes.
- TASK-FIN-103-01
Point-in-time factor reconstruction & leakage audit
Temporal reasoning, data lineage and feature engineering.
Quantitative Researcher
- TASK-FIN-103-02
Strategy implementation & backtest repair
Quantitative coding, financial calculations and translating a strategy into code.
Quantitative Developer
- TASK-FIN-103-03
Market data integrity & corporate action handling
Data validation, adjusted prices and historical consistency.
Data Scientist
- TASK-FIN-103-04
Strategy robustness & overfitting audit
Statistical testing, out-of-sample analysis and methodological validity.
Quantitative Researcher
Agent tools. What the agent can call.
The agent queries point-in-time market and fundamental data, runs Python in a sandbox, uses the backtesting engine and the experiment registry, and submits code, data and a report for the verifier to rerun.
query_market_dataget_fundamental_vintageget_corporate_actionsinspect_research_notebookexecute_pythonrun_backtestget_experiment_resultssubmit_research_artifacts
Verification methodology. Checked against state the verifier computes itself.
Point-in-time data availability, corporate-action handling, separation of training and evaluation periods, reproducible calculations, transaction-cost and portfolio assumptions, and whether the injected methodological errors were found.
The verifier checks
- 01Every input was available at its decision date, by filing date and vintage.
- 02Corporate actions, dividends and delistings are handled as recorded.
- 03Training and evaluation periods don't overlap, including through label horizons.
- 04Submitted code reruns and reproduces the submitted results.
- 05Transaction costs and portfolio construction follow the specification.
- 06Injected methodological errors are found and corrected.
- 07Conclusions are consistent with the corrected evidence.
Reward
High returns or an attractive Sharpe ratio never raise the reward on their own. Correct methodology is the objective, and statistically defensible alternative approaches are accepted.
Scores zero on
- Using information unavailable at the decision date
- Overlapping training and evaluation periods
- Results that can't be reproduced from the submitted code
- Conclusions the evidence doesn't support
Included deliverables. What you receive, and what we agree before we build.
Delivered with the environment
- Containerized environment with a sandboxed research workspace and versioned data
- Validated task instances with reference implementations
- Verifier code that reruns submissions on held-out data
- Agent traces and failure-mode analysis across repeated runs
- Technical datasheet and integration documentation
Agreed before development
- Data universe and history
- Execution and cost assumptions
- Compute and execution limits per episode
- Tolerance thresholds for reproduced results
- Any third-party data licenses