microsoft-qlib

name	microsoft-qlib
description	Build Qlib-inspired quantitative research datasets and offline alpha experiments with QVeris-routed data. Use for point-in-time universe construction, factor engineering, model training, walk-forward backtests, attribution, and portfolio research without direct market-data APIs or live execution.

Qlib-Inspired Quant Research with QVeris

Use the upstream Qlib project as a methodology reference. Keep feature computation, model training, backtesting, risk modeling, and portfolio optimization local.

Data boundary

Acquire every external market or fundamental input through QVeris. Require QVERIS_API_KEY for real calls. Use Discover → Inspect → Call and the capability plan in qveris.manifest.json; do not call market-data providers directly.

Workflow

1. Define the market, universe, benchmark, calendar, frequency, label horizon, adjustment policy, and as-of cutoff.
2. Discover and call adjusted price history, index constituents, industry classification, financial ratios, and realized volatility.
3. Record search_id, tool_id, provider, retrieval time, field definitions, currency, timezone, and coverage window.
4. Construct a point-in-time data panel. Lag fundamental values by their public availability date and prevent survivorship leakage.
5. Calculate features locally and version their definitions. Fit preprocessing on training windows only.
6. Use chronological train, validation, and test windows with walk-forward retraining when appropriate.
7. Evaluate information coefficient, turnover, capacity proxies, transaction costs, exposure, drawdown, and benchmark-relative performance.
8. Attribute results by factor, industry, regime, and time period. Run perturbation and missing-data sensitivity checks.
9. Treat index rebalance events as an explicit current gap; disclose any approximation in historical-universe reconstruction.

Failure rules

• Stop if adjustment factors, symbol identity, or point-in-time dates are ambiguous.
• Do not mix provider schemas without a documented normalization step.
• Do not infer production readiness from one backtest.
• Do not place live orders.

Output contract

Return dataset and feature manifests, QVeris trace, split definitions, model and benchmark settings, performance and risk metrics, attribution, leakage tests, coverage gaps, and reproduction commands.

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 .