Numerai Research

Overview

This skill is a “meta-workflow” that sequences existing Numerai skills so research requests reliably produce: (1) runnable configs, (2) executed experiments, (3) a full written report + plots, and (4) a deployable pickle when requested.

Workflow (always follow this order)

1) Design the experiment (use numerai-experiment-design)

• Follow the numerai-experiment-design skill to:
  • clarify the idea (or run quick scout interpretations if ambiguous)
  • choose baseline + feature set alignment (default ender20 baseline)
  • create an experiment folder under numerai/agents/experiments/<experiment_name>/
  • write configs in configs/
  • run training via PYTHONPATH=numerai python3 -m agents.code.modeling --config <config> --output-dir <experiment_dir>
  • track metrics with BMC as primary (bmc_mean, bmc_last_200_eras)
  • iterate in rounds (typically 4–5 configs per round), and keep going until you hit a plateau (per the experiment-design skill)
  • scale winners (bigger feature set and/or full data) before finalizing the best model

2) Implement new model types if needed (use numerai-model-implementation)

Only if the idea requires new code (new model wrapper, new fit/predict behavior, etc.):

• Follow the numerai-model-implementation skill to add the model type and register it.
• Add at least one smoke-test config and verify the pipeline runs.

3) Report the research (use report-research)

After you have iterated through multiple rounds and stopped finding improvements (plateau), and after any confirmatory scale runs:

• Follow the report-research skill to:
  • write a full experiment.md (abstract + methods + results + decisions + next steps)
  • generate the standard show_experiment plot(s)
  • link plots and artifacts in the report

4) Package and upload (use numerai-model-upload)

If (and only if) the user wants deployment:

• Follow the numerai-model-upload skill to create a Numerai-compatible pickle and upload it via the Numerai MCP.
• Remember: only Classic (tournament 8) supports pickle uploads.

Defaults (unless user specifies otherwise)

• Scout first on downsampled data; scale only winners.
• Run experiments in rounds (4–5 configs per round) and stop only after a plateau + confirmatory scale step.
• Benchmark reference: v52_lgbm_ender20.
• Always record corr + BMC metrics and include the standard plot in the report.