ai4finance-foundation-finrl

name	ai4finance-foundation-finrl
description	Build FinRL-inspired reinforcement-learning research datasets and offline strategy evaluations using QVeris-routed market data. Use for train-validate-trade experiments, state and reward design, risk controls, walk-forward evaluation, and paper-trading research without live order execution.

FinRL-Inspired Research with QVeris

Use the upstream FinRL project as a methodology reference. Keep model training, environment simulation, reward calculation, and backtesting local.

Data boundary

Acquire every external market input through QVeris and require QVERIS_API_KEY for real calls. Use Discover → Inspect → Call and the capability plan in qveris.manifest.json. Do not call exchange, broker, or market-data provider APIs directly.

This evaluation skill is research-only. Do not place, modify, or cancel orders and do not access brokerage accounts.

Workflow

1. Define the universe, market calendar, adjusted-price policy, rebalance frequency, transaction-cost model, and decision horizon.
2. Discover and call adjusted bars, technical indicators, realized-volatility, and market-breadth capabilities.
3. Record QVeris search_id, tool_id, provider, retrieval time, timezone, and requested data window.
4. Sort and deduplicate observations, align them to the exchange calendar, and reject non-monotonic or future-dated rows.
5. Build point-in-time state features. Fit scalers and imputers on the training window only.
6. Use chronological train, validation, and test windows. Prefer walk-forward evaluation over random splits.
7. Define actions, portfolio constraints, turnover penalties, transaction costs, and the reward before evaluating a policy.
8. Use realized volatility and breadth as the declared risk-state substitute while QVeris has no native VIX coverage.
9. Compare against simple benchmarks and report return, volatility, Sharpe, maximum drawdown, turnover, exposure, and cost sensitivity.

Failure rules

• Stop if adjusted-price semantics or trading calendar cannot be verified.
• Do not silently substitute a different market, frequency, or symbol.
• Do not claim live-trading readiness from an offline backtest.
• Mark optional VIX data missing rather than fetching it outside QVeris.

Output contract

Return the experiment specification, QVeris data trace, feature dictionary, chronological split, environment assumptions, benchmark comparison, risk metrics, leakage checks, and reproducibility limitations.

Validation

node scripts/qveris_candidate_runner.mjs validate .
node --test tests/smoke.test.mjs
node scripts/qveris_candidate_runner.mjs preflight .
QVERIS_API_KEY=... node scripts/qveris_candidate_runner.mjs live .