Qlib joins the research steps that usually live in separate notebooks
Qlib covers market-data access, feature construction, model training, backtesting, portfolio analysis, and order execution. Its qrun command reads a workflow configuration and records the experiment through MLflow. That is useful when your problem is repeatability: the same dataset and evaluation path can compare LightGBM, neural models, or a custom estimator without rebuilding the surrounding machinery for each paper.
The scope is much wider than a charting backtester. Qlib includes supervised-learning examples, changing-market research, reinforcement-learning work, online model management, and nested execution. commit be72549 contained 626 files and about 75,401 lines of source in an 8.2 MB checkout. A quant team can grow inside that structure. Someone testing one moving-average rule may spend more time learning the framework than testing the rule.
The 12-second npm install sets up commit checks, not Qlib
Our sandbox detected the root package.json and took the Node path. It installed 75 packages in 12 seconds and used 62 MB on disk. That package calls itself qlib-ci; its only dependencies are Commitlint packages for checking pull-request titles. A clean npm result therefore says nothing about whether Python can import Qlib, LightGBM can compile, or a sample workflow can read market data.
This distinction is easy to miss in automated evaluations because the repository really does contain a valid Node manifest. The actual project metadata is in pyproject.toml, where pyqlib requires Python 3.8 or newer and lists data, experiment, optimization, plotting, and storage libraries. The README documents Python 3.8 through 3.12. It recommends Conda because missing headers can break installation outside a managed environment.
What happened when we ran it
Our run at commit be72549 installed the npm side in 12 seconds: 75 packages, 62 MB, and 0 known vulnerabilities from npm audit. No build script existed, so the lab skipped that step. No test script existed either, so it skipped tests. The repository did have 6 CI workflow files, a Dockerfile, and a tests directory, but those signals belong to the wider Python project.
The container had 3 CPUs, 8 GB of RAM, Node 22, no secrets, and no elevated privileges. We did not run pip install, qrun, the Python tests, a backtest, or a data download. There is no honest way to turn the npm result into a claim about model speed or Python reliability. It exposed a detection trap and nothing more.
Data ownership starts before the first backtest
Qlib's quick start currently cannot use Microsoft's official dataset because the README says that download is temporarily disabled. It points readers to a community-maintained release, while the included collectors can build data from Yahoo Finance. The same document warns that the Yahoo data may be imperfect and recommends supplying higher-quality data. That warning matters more than the convenience of a one-command workflow.
The offline dataset also cannot simply be extended with incremental updates because some fields were removed to make it smaller. The README tells users who need ongoing data to collect Yahoo history from scratch and then append to it. A health-check script can inspect missing values and large price or volume jumps. You still have to decide whether those thresholds fit your market, instruments, and trading calendar.
Model breadth comes with conflicting environments
The benchmark catalog covers many model families, but they do not share one painless environment. Qlib explicitly says each baseline has different dependencies. Its TFT example is tied to Python 3.6 or 3.7 because of TensorFlow 1.15, even though the current package supports Python 3.8 through 3.12. Reproducing an older paper may therefore mean isolating it from the main Qlib environment.
Apple Silicon adds another documented wrinkle: LightGBM may need OpenMP installed before Qlib can build. Analysis plots are optional dependencies, and reinforcement-learning support brings Torch plus a NumPy constraint. None of these is disqualifying for a research group used to per-model environments. They do make Qlib a poor choice for a team expecting every catalog entry to run from one lockfile.
An open drawdown bug deserves a pre-production check
Issue 2358 reports that risk_analysis can ignore a loss in the first period when calculating maximum drawdown. Its reproduction on commit be72549 shows a series opening with a 10% loss returning a drawdown of 0. The report also traces that calculation into the standard portfolio-analysis output. For a research platform, that is a decision-relevant defect, not a cosmetic edge case.
GitHub showed 484 open issues and pull requests, split into 305 issues and 179 pull requests when checked. The last push was September 22, 2026, and several late-September changes addressed pandas compatibility, drawdown logic, Yahoo updates, and documentation. That activity argues against calling the project abandoned. It also means adopters should pin a commit, verify the metrics they use, and read the issue queue before trusting a recorded result.
Qlib pays off when the workflow is the product
Choose Qlib when your team repeatedly turns market data into features, trains several models, records experiments, and evaluates strategies under one research convention. Its MIT license and broad model catalog make that a credible foundation. The price is ownership: data quality, environment isolation, and finance-specific regression checks stay with you.
For a single strategy, Backtrader or Zipline is easier to reason about. For large vectorized parameter sweeps, vectorbt is more direct. Qlib earns its complexity when several researchers need the same path from data to analysis. Before adopting it, run the Python suite and one representative workflow on your own dataset. Our 12-second npm result cannot answer the question that matters.

