howard-lynn-ye Python quant stack — router
Your training prior on this ecosystem is stale in ways that silently produce wrong numbers rather than errors. Read §1 before writing code. Then jump from §2 to the one skill that owns your task; do not try to answer from memory.
Every fact in this library carries a verification date. Where a fact is likely to have moved,
re-check the primary source listed in shared/live-sources.md rather than trusting the cache.
1. Version drift — where your prior is wrong
(verified 2026-09-03 against the PyPI JSON API and each project's own repo)
| Area | Stale prior you probably hold | Current reality |
|---|---|---|
| TA-Lib install | "needs the C library compiled by hand; no Windows wheels" | Solved. 0.7.1 ships 54 wheels incl. cp311-win_amd64. pip install TA-Lib just works. |
| QuantLib install | "a nightmare, build from source" | Solved. 1.43 ships cp39-abi3-win_amd64; one wheel covers 3.9+. No sdist at all. |
| IBKR client | ib_insync |
Dead (last release 2023-07-02). Successor is ib_async (2.1.0, 2025-12-08). |
| TD Ameritrade | tda-api |
Dead (2022-06). TDA absorbed into Schwab → schwab-py. |
| Alpaca | alpaca-trade-api |
Deprecated (2024-01). → alpaca-py. |
pandas_datareader |
pdr.get_data_yahoo(...) |
Removed in 0.11.0. Yahoo, Stooq, Tiingo, IEX, Quandl, AlphaVantage, Morningstar, Options — all deleted. It is now a macro-only library (FRED, Fama-French, OECD, Eurostat, EconDB, BoC). Every equity tutorial using it is broken. |
yfinance adjustment |
yf.download() returns raw OHLC + Adj Close |
auto_adjust=True since 1.0. OHLC are already adjusted and there is no Adj Close column — df['Adj Close'] raises KeyError. |
backtrader |
"a standard choice" | Unmaintained (last release 2023-04-19) and GPL-3.0. |
mlfinlab |
"the AFML reference implementation" | Gone. Not installable from PyPI; the GitHub source is stubbed (every function body is pass); proprietary all-rights-reserved. Use RiskLabAI + purgedcv. |
pandas-ta |
"the pandas-native TA-Lib" | See signal-construction — there is a supply-chain caution on the current PyPI package. Maintained successor: pandas-ta-classic. |
openbb |
MIT | Relicensed to AGPL-3.0 on 2024-05-14. Network copyleft. |
backtesting.py |
permissive | AGPL-3.0. |
PyBroker / vectorbt |
open source | Both are Apache-2.0 + Commons Clause — you may not sell a product or service deriving substantially from them. Not OSI-open-source. |
nautilus_trader |
"works everywhere" | Requires Python >=3.12,<3.15. Will not install on 3.11. |
fracdiff |
the fractional-differencing package | Archived 2023-12; requires_python <3.10. Unusable on modern Python. |
| Microsoft Qlib CN data | qlib_data --region cn downloads it |
Official dataset disabled ("more restrict data security policy"). Use the community mirror chenditc/investment_data. |
| Alpha Vantage free tier | 500 requests/day | 25 requests/day. Effectively a demo. |
polars |
"a fast Rust dataframe library" | The polars wheel is now an empty 865 KB shim (py3-none-any, no compiled code) that hard-depends on polars-runtime-32. A lockfile listing only polars does not pin the engine. |
polars.join_asof |
"pandas is the one that silently misjoins" | Inverted. pandas.merge_asof raises on unsorted keys; polars.join_asof does NOT check sortedness when by= is given — silently wrong rows. |
SEC XBRL frames API |
"a clean cross-section" | Not point-in-time and cannot be made so — no filed field. 53% of its CY2023 values come from filings made in 2026. Never backtest on it. |
2. Quick task reference
Written in the words you would actually use. -> is a literal path to read next.
"get stock prices" / "download OHLCV" / "which data source" / yfinance / polygon / databento / EODHD / tiingo / survivorship-free universe / delisted tickers / trading calendar / split & dividend adjustment
-> market-data-sourcing skill. Start with its plugins/fin-core/skills/market-data-sourcing/references/_decision-table.md.
"as-of join" / merge_asof / "join quotes to trades" / Parquet / polars / DuckDB / ArcticDB / "store tick data" / "my timestamps are wrong" / "too big for memory" / "different numbers when I parallelise"
-> market-data-engineering skill. Its as-of-join section is where look-ahead most often enters a pipeline.
"fundamentals" / 10-K / 10-Q / 8-K / EDGAR / XBRL / "point-in-time financials" / "as-of-date fundamentals" / earnings dates / CIK / restatements
-> fundamental-and-macro-data skill, §SEC. Read its restatement and acceptanceDateTime sections before writing any event study.
"macro data" / FRED / CPI / GDP / NFP / "revised data" / vintage / ALFRED / World Bank / IMF / Eurostat
-> fundamental-and-macro-data skill, §Macro. GDP is revised for years — plugins/fin-core/skills/fundamental-and-macro-data/references/fredapi.md covers the vintage API and its three bugs.
"backtest this" / "which backtesting library" / vectorbt / zipline / nautilus / LEAN / freqtrade / "my backtest looks too good"
-> backtesting-engines skill. If the backtest looks too good, go to research-integrity-guards first.
"connect to IBKR" / TWS / ib_async / place an order / paper trading / Alpaca / Schwab / Tastytrade / "make sure I don't send a live order"
-> broker-execution-apis skill. Its order-safety patterns are mandatory reading before any code that can transmit an order.
"RSI" / "MACD" / "moving average" / TA-Lib / "which indicator library" / "does this indicator repaint" / streaming indicators
-> signal-construction skill.
"is my factor any good" / IC / alphalens / quantile returns / Fama-French / Fama-MacBeth / event study / abnormal returns / CAR
-> factor-and-timeseries-research skill.
"forecast returns" / ARIMA / GARCH / volatility model / Nixtla / sktime / darts / Prophet / time-series foundation model
-> factor-and-timeseries-research skill, §Forecasting. For volatility specifically, plugins/fin-core/skills/factor-and-timeseries-research/references/arch.md.
"optimize portfolio weights" / mean-variance / Black-Litterman / risk parity / HRP / CVaR / efficient frontier / covariance shrinkage
-> portfolio-and-risk skill, §Optimization.
"Sharpe ratio" / drawdown / tearsheet / quantstats / pyfolio / "annualize returns" / attribution / VaR
-> portfolio-and-risk skill, §Analytics. Read its metric-correctness audit — several popular libraries compute these wrong.
"is this result real" / overfitting / "I tried 200 strategies" / deflated Sharpe / PBO / walk-forward / purged CV / p-hacking / "how many trials"
-> backtest-validation skill. This is the single highest-value skill in the library.
"price an option" / Greeks / implied vol / vol surface / SABR / QuantLib / yield curve / bond pricing / swap
-> derivatives-pricing skill.
"look-ahead bias" / "survivorship bias" / "data leakage" / "point-in-time" / "is my backtest honest" / "review my research design"
-> research-integrity-guards skill. Also load this whenever you are about to report a number.
A-share / 沪深 / akshare / tushare / baostock / 复权 / 涨跌停 / T+1 / vnpy / qlib / 北交所
-> fin-china plugin: china-ashare-data, china-trading-stack.
Hong Kong / Taiwan / Japan / Korea / India / SGX / ASX / HKEX / TWSE / KRX / NSE / Stock Connect / J-Quants / 港股 / 台股
-> fin-asia plugin: asia-pacific-markets. Its exchange_calendars defect list applies to US work too.
crypto / ccxt / Binance / perpetuals / funding rate / freqtrade / hummingbot
-> fin-crypto plugin: crypto-data-and-execution.
reinforcement learning / RL agent / FinRL / gym / gymnasium / stable-baselines3 / TensorTrade / LSTM / Transformer for returns / "does deep learning beat linear"
-> fin-llm plugin: rl-and-ml-trading. FinRL does not install; the evidence section is the point.
alternative data / satellite / credit card / web traffic / ESG / Numerai / WallStreetBets / news sentiment sources
-> plugins/fin-core/skills/market-data-sourcing/references/alternative-data.md
"LLM trading agent" / TradingAgents / FinGPT / FinRobot / RD-Agent / "does AI trading work" / finance MCP server
-> fin-llm plugin: llm-finance-agents (read its evidence section before building anything), finance-mcp-servers.
3. Searching this library
Most library detail lives inline in each skill's own tables, not in separate files. Some skills
additionally carry references/*.md for material too long to inline — a decision table, a
methodology, a single deep library note. catalog/index.json lists exactly which, generated from
frontmatter, so check there rather than assuming a file exists.
When the table above does not name your library, grep the skill bodies first:
# Which skill covers a given package?
grep -ril "riskfolio" plugins/*/skills/*/SKILL.md
# What does it get wrong? (traps are marked with a siren in every skill)
grep -i -B2 -A6 "quantstats" plugins/fin-core/skills/portfolio-and-risk/SKILL.md
# Only some skills have reference files; list them before reading
ls plugins/*/skills/*/references/
4. Three rules that override any library's defaults
- A library's default is not a safe default. Adjustment mode, CV splitter, risk-free rate, and annualization factor all default to something wrong for research in at least one popular package. Set them explicitly, always.
- Never report a backtest number without its trial count. How many variants were tried is part
of the result. See
backtest-validation. - Code license ≠ data license. yfinance is Apache-2.0; the Yahoo data it fetches is personal-use-only. The permissive license on a scraper grants you nothing about the scraped data.