Skip to content

PRD.md

Latest commit

 

History

History
97 lines (79 loc) · 3.9 KB

PRD.md

File metadata and controls

97 lines (79 loc) · 3.9 KB
Raw

Product Requirements Document (PRD)

1. Summary

This repository provides a Gymnasium-compatible trading environment, a backtesting workflow, and an optional LLM-driven assistant layer. The project emphasizes reproducibility, safe separation of secrets, and a clear configuration model.

2. Goals

  • Provide a Gymnasium trading environment suitable for training and evaluating agents.
  • Provide a backtesting path that is reproducible and configurable.
  • Provide a system-wide configuration layer for non-secret defaults (LLM model/prompt, IO directories, concurrency) via system.ini.
  • Keep secrets out of source control and out of runtime logs.
  • Offer a minimal server surface for health/config inspection and future APIs.

3. Non-Goals (for now)

  • No guarantee of broker/exchange execution correctness for live trading.
  • No portfolio optimization suite beyond basic risk controls.
  • No UI dashboard requirements.

4. Personas

  • Solo Researcher (Primary): Runs gym/backtests locally, iterates quickly, values reproducibility.
  • Quant Engineer (Secondary): Integrates with external datasets/providers and CI.
  • Operator (Future): Runs paper/live mode with strict safety controls.

5. Configuration Model (Source of Truth)

5.1 Secrets / Environment (Pydantic)

  • Stored in .secrets.env and/or .env (not committed).
  • Examples: OPENROUTER_API_KEY.
  • Rule: secrets must never be written to logs or returned via API.

5.2 System-wide non-secret configuration: system.ini

Purpose: stable, system-wide defaults that are not part of a specific backtest/strategy run.

  • LLM provider + model + base_url + base prompt
  • IO folder conventions
  • Concurrency / behavior defaults

5.3 Gym/backtest runtime configuration: strategy.yaml

Purpose: run-specific configuration for training/backtesting.

  • strategy name, symbols, timeframe
  • execution assumptions (paper trading, fees, slippage)
  • risk constraints
  • data provider toggles

6. User Stories

  • As a researcher, I can run a backtest with a single strategy.yaml file and get deterministic outputs.
  • As a researcher, I can change LLM model/prompt without touching YAML or secrets.
  • As a developer, I can inspect loaded config at runtime without exposing secrets.

7. Functional Requirements

7.1 Config loading

  • Load system.ini into settings.system.
  • Load strategy.yaml into settings.runtime.
  • Load env secrets into Settings fields.
  • Provide a single settings object importable across modules.

7.2 IO paths

  • Create or expect a root IO directory (default ./io).
  • Subdirectories: ./io/logs, ./io/cache.

7.3 Logging

  • Central log level controlled by env (log_level) and/or system behavior defaults.
  • Redaction rule: never log secrets.

7.4 Server API (initial)

  • / returns a simple status string.
  • /health returns basic health and mode.
  • /config/system returns non-secret INI config.
  • /config/runtime returns YAML runtime config.
  • /config/secrets returns only boolean presence flags.

8. Security & Privacy Requirements

  • .secrets.env must be gitignored.
  • Do not return secret values via HTTP.
  • Avoid writing secret values to disk in IO outputs.

9. Milestones

v0 (Now)

  • Establish config model (env + system.ini + strategy.yaml).
  • Minimal server endpoints for inspection.
  • Document usage.

v1

  • Implement core gym environment + backtest runner pipeline.
  • Standardize output artifacts (metrics, trades, plots) under io_dir.
  • Add tests for config loading and redaction.

v2

  • Add provider integrations and caching strategy.
  • Add a pluggable LLM “assistant” layer for analysis/report generation.
  • Add CI checks and linting.

Appendix: Repository Files

  • system.ini: system-wide defaults
  • strategy.yaml: run-specific backtest/gym config
  • .secrets.env: secrets (not committed)
  • config.py: config loader / settings singleton
  • server.py: minimal HTTP surface