howard-lynn-ye
UserSource-verified knowledge base of the Python quantitative-finance ecosystem, packaged as Agent Skills. Every claim dated and marked verified/secondhand/unverified.
Categories
Indexed Skills (81)
asia-pacific-markets
Data and trading for Asia-Pacific outside mainland China. TRIGGER - Hong Kong, HKEX, 港股, Stock Connect, VCM, per-stock lot sizes; Taiwan, TWSE, 台股, shioaji, FinMind, twstock; Japan, TSE, JPX, J-Quants, jquants-api-client, the 2022 TSE restructuring; Korea, KRX, KOSPI, KOSDAQ, pykrx, FinanceDataReader, CSAT, Korean short-selling ban dates, Korean delisted-stock lists; India, NSE, BSE, kiteconnect, Zerodha, Upstox, STT, Muhurat; Singapore SGX, Australia ASX; futu-api, tigeropen. Also documents four verified exchange_calendars defects that affect US work too. SKIP for mainland A-shares (china-ashare-data, china-trading-stack).
backtest-validation
Decide whether a result survives the number of things you tried. TRIGGER - "I tried N strategies or parameters and the best one..."; overfitting, p-hacking, data snooping, multiple testing, "is this result real", "is this statistically significant"; deflated Sharpe, DSR, PSR, probability of backtest overfitting, PBO; purged or combinatorial cross-validation, embargo, walk-forward, "cross validation on time series without leaking"; a grid search, hyperopt or AutoML picked a winner; triple-barrier labeling, meta-labeling, fractional differentiation, mlfinlab. Load whenever a Sharpe ratio is about to be offered as evidence for trading. SKIP only for plain performance metrics - Sharpe, Sortino, CAGR, drawdown (portfolio-and-risk); the deflated and probabilistic Sharpe ratios are computed HERE.
backtesting-engines
Choose a backtesting engine and know what it silently models wrong. TRIGGER - "backtest this", backtest a crossover or a moving-average strategy, simulate a strategy, walk-forward, parameter sweep, "test this trading idea"; comparing or choosing backtest frameworks; vectorbt, backtesting.py, backtrader, zipline, PyBroker, bt, nautilus_trader, LEAN, freqtrade, jesse; how an engine models fills, slippage, commissions; what slippage to assume, partial fills, margin, shorting or delistings; taking a strategy from backtest to live; "my backtest looks too good"; "works in backtest but loses money live". Several popular engines fill at the signal's own bar close by default. SKIP for judging whether a finished result is real (backtest-validation) for A-share rules (china-trading-stack), for crypto funding and perpetuals (crypto-data-and-execution), for options, which need assignment and settlement no engine here models (options-backtesting), and for measuring fills you already have (execution-cost-analysis).
broker-execution-apis
Connect to a broker and place orders without accidentally trading live money. TRIGGER - connect to Interactive Brokers, TWS, IB Gateway, ib_async, ib_insync, ibapi, Alpaca, Schwab, schwab-py, Tastytrade, Tradier or Robinhood; place, modify or cancel an order; read positions or balances; set up paper trading; order types, time-in-force, bracket or OCO orders, client order ID; FIX, quickfix, simplefix; "make sure I don't send a live order"; a broker connection being refused. Load before any code that can transmit an order. SKIP for crypto exchanges and ccxt (crypto-data-and-execution), and for vnpy, CTP, QMT or any Chinese broker gateway (china-trading-stack). This skill answers TWS/Gateway port and connection failures on its own - the fin-libraries deep dive is optional and most installs will not have it.
china-ashare-data
Get China A-share and Greater China market data without the ecosystem's silent traps. TRIGGER - A股, 沪深, 北交所, 科创板, 创业板; akshare, tushare, baostock, efinance, adata, qstock, mootdx, easyquotation, jqdatasdk, 聚宽, rqdatac, 米筐, Wind, 万得, Choice, 东方财富; 复权, qfq, hfq, 前复权, 后复权; ST, 退市, delisted A-share tickers, 退市股票列表; 停牌 suspension; 公告日 versus 报告期; CSI300, HS300, 中证 index membership. Three popular libraries default to forward-adjusted prices, which are rewritten retroactively and are therefore look-ahead contaminated. SKIP for backtesting or trading A-shares (china-trading-stack) and for Hong Kong, Taiwan, Japan or Korea (asia-pacific-markets).
china-trading-stack
Backtest and execute Chinese-market strategies under the rules a Western engine gets wrong. TRIGGER - T+1 settlement, 涨跌停 price limits, 印花税 stamp duty, 夜盘 night session, 集合竞价, the lunch break, 240 minute bars; vnpy, CTP, QMT, miniQMT, easytrader, tqsdk, 天勤, RQAlpha, wondertrader, 掘金; Microsoft Qlib's China dataset; backtesting or live-trading A-shares, Chinese futures, options or convertible bonds; porting a US strategy to China. SKIP for sourcing the data itself (china-ashare-data).
crypto-data-and-execution
Crypto market data and execution, and how a 24/7 market breaks equity tooling. TRIGGER - crypto, Bitcoin, BTC, Ethereum, ETH, digital assets, perpetuals, perps, funding rate, crypto order book; ccxt, cryptofeed, python-binance, freqtrade, jesse, hummingbot, OctoBot; a crypto exchange, testnet or sandbox; annualizing crypto returns; porting an equity strategy to crypto. Annualization is 365 not 252, funding is frequently larger than the alpha being measured, and exchange pair lists are chronically survivorship-biased. SKIP for equity and futures brokers (broker-execution-apis).
derivatives-pricing
Price options and fixed income, and get the Greeks and conventions right. TRIGGER - option pricing, implied volatility, IV, Greeks, delta, gamma, vega, theta, rho; Black-Scholes, binomial, Monte Carlo, American exercise, early exercise, exotics; volatility surface, smile, skew, SVI, SABR, ZABR; QuantLib, vollib, py_vollib, financepy, rateslib, optionlab; option chains; or yield curve, discount factor, curve bootstrapping, day count, swap and bond pricing. Greek scaling differs by 100x and 365x between the three common libraries, and one popular fixed-income package is not open source. SKIP for option lifecycle events - assignment, pin risk, expiry - live or backtested, and for historical option chains and vendors (options-backtesting).
etf-mechanics
Why an ETF's price series does not behave like the index it tracks - daily-reset leverage, NAV vs price, distributions, holdings files and fees. TRIGGER - "why is my 3x ETF down when the index is flat", TQQQ decay, SQQQ, leveraged ETF long term, inverse ETF, volatility drag, daily reset; premium to NAV, discount to NAV, iNAV, creation/redemption, "bond ETF trading below NAV"; ETF distribution, capital gains distribution, return of capital, "ETF dropped on the ex-date", phantom drop; ETF holdings file, constituents CSV, index reconstitution, Russell rebalance, "I used today's holdings for the backtest"; expense ratio drag, tracking difference vs tracking error, "ETF returned less than the index". SKIP for downloading price series and vendor adjustment defaults (market-data-sourcing) - holdings files stay here, for UNG, USO or VIXY roll yield and contango (futures-continuous-contracts), for auditing a finished backtest (research-integrity-guards), and for weights, Sharpe or drawdown (portfolio-and-risk).
execution-cost-analysis
Measure what your execution actually cost instead of assuming a number - implementation shortfall, benchmark choice, impact models, and the gap between the cost you assumed and the cost you paid. TRIGGER - transaction cost analysis, TCA, implementation shortfall, arrival price, decision price, slippage analysis, execution quality, fill quality, did I get a good fill; VWAP or TWAP benchmark, beat VWAP, participation rate, POV, percentage of volume, child orders, order slicing; market impact, temporary vs permanent impact, square-root law, Almgren-Chriss, price reversion after my order; "how much size can this strategy take", capacity, alpha decay with size. SKIP for a slippage assumption inside a backtest and for "works in backtest, loses live" with no measured fills (backtesting-engines), for whether that assumption is credible (research-integrity-guards), and for broker order types (broker-execution-apis).
external-skill-index
A verified index of every public finance Agent Skill repository — 139 repos, 4,851 SKILL.md files — so you can find what already exists instead of rebuilding it, and avoid the third that is legally unusable. TRIGGER - looking for an existing skill, plugin or marketplace for anything financial; "is there already a skill for X"; choosing between competing finance skill packs; before writing a new finance skill; checking whether a skill repo's licence permits use; or asked what the Claude/agent finance skill ecosystem contains. Also load before recommending any third-party finance skill repository, because 44 of the 139 declare no usable licence and 12 advertise skills while shipping none.
factor-and-timeseries-research
Judge whether a cross-sectional factor predicts returns, and forecast financial series. TRIGGER - information coefficient, IC, quantile returns, factor decay, turnover, alphalens; Fama-French, Fama-MacBeth, PanelOLS, linearmodels, cross-sectional asset pricing; event study, abnormal returns, CAR, BHAR; Alpha101, Alpha158, symbolic alpha mining, gplearn; or forecasting with ARIMA, GARCH, volatility models, arch, Nixtla, statsforecast, mlforecast, sktime, darts, Prophet or a time-series foundation model. SKIP for computing the indicator itself (signal-construction) and for portfolio weights or Sharpe (portfolio-and-risk).
finance-agent-architectures
How the mainstream finance agent systems are built, and how to stage a research-to-execution pipeline whose gates are code. TRIGGER - build a multi-agent trading system; TradingAgents architecture; the ai-hedge-fund repo; RD-Agent for quant; Vibe-Trading; FinRobot vs FinGPT; FinMem layered memory; a LangGraph, CrewAI, AutoGen or Claude Agent SDK pipeline for stock research; analyst, researcher, trader and risk-manager agents, bull-bear debate; an agent that reads 10-Ks and trades; "how should the pipeline be staged", where the LLM sits, human-in-the-loop gates, prompt injection through scraped filings, agent reproducibility; 交易 agent 架构, 多智能体 pipeline. SKIP for whether any of it makes money or whether you should build one at all (llm-finance-agents), choosing an MCP server (finance-mcp-servers), RL agents (rl-and-ml-trading), and order safety at the broker (broker-execution-apis).
finance-mcp-servers
Pick a finance MCP server, and know its licence and blast radius before connecting it. TRIGGER - an MCP server for market data, filings, macro data, brokerage or trading; adding, choosing, comparing or debugging a finance MCP; Alpaca MCP, Alpha Vantage MCP, Polygon or Massive MCP, OpenBB MCP, SEC EDGAR MCP, FRED MCP, yfinance MCP, QuantConnect MCP; "which finance MCP should I install"; granting an MCP server the ability to place orders. Several are AGPL-3.0, one places real trades, and the most-starred one is over a year stale. SKIP for the Python libraries behind them (market-data-sourcing).
fundamental-and-macro-data
Company fundamentals and macro series with correct point-in-time semantics. TRIGGER - 10-K, 10-Q, 8-K, 13F, Forms 3/4/5, filings, EDGAR, XBRL, accession number, CIK, "which CIK is this ticker", ticker-to-CIK mapping, edgartools; parsing an income statement or balance sheet out of a filing; revenue, EPS or balance-sheet history as it was known on a past date; restatements; earnings dates; or CPI, GDP, payrolls, unemployment, interest rates, FRED, ALFRED, data vintages and revisions. Load before joining ANY fundamental or macro series to prices: the obvious join is a look-ahead bug, and the SEC frames API cannot be made point-in-time. SKIP for price and OHLCV vendors (market-data-sourcing) and Chinese filings (china-ashare-data).
futures-continuous-contracts
Build and use a futures price series correctly — a continuous contract does not exist in the market, it is stitched, and the stitching method changes your answer. TRIGGER - futures, continuous contract, back-adjusted, Panama adjustment, ratio adjustment, roll, roll yield, contango, backwardation, front month, expiry, first notice day, open interest roll, CME, Globex, ES, CL, NG, VX, GC, ZN; joining futures bars to an equity calendar; "my futures backtest returns look wrong"; negative prices in a price series; norgatedata, databento continuous symbols, yfinance CL=F or ES=F. SKIP for Chinese futures and 夜盘 (china-trading-stack) and for crypto perpetuals, which have funding rather than rolls (crypto-data-and-execution).
fx-markets
Trade and backtest FX correctly — quote conventions, pip sizing, and the carry that a spot-only backtest silently omits. TRIGGER - forex, FX, currency pair, EURUSD, USDJPY, GBPUSD, AUDUSD, USDCHF, USDCAD, NZDUSD; pip, pipette, lot sizing on a currency pair; carry trade, swap points, rollover, interest rate parity, covered or uncovered parity; NDF, forward points, T+2 value date; "there is no official FX close"; DukasCopy or free tick FX; forex-python or a similar package. SKIP for crypto pairs, which have funding rather than swap (crypto-data-and-execution), and for FX options and vol surfaces (derivatives-pricing).
intraday-microstructure
Measure the market at the tick level and know when the measure is lying. TRIGGER - build dollar bars, volume bars, tick bars, imbalance bars; classify trades as buyer or seller initiated, Lee-Ready, tick rule, bulk volume classification, BVC; quoted vs effective vs realized spread, price impact, Kyle's lambda, Amihud illiquidity, order flow imbalance, OFI, VPIN; reconstruct the order book from MBO, market by order, L2 vs L3, queue position, NBBO, odd lots, dark prints; "my tick strategy works on bars but not on ticks", exchange clock vs vendor clock, the 5-second rule. SKIP for downloading tick data or picking a vendor (market-data-sourcing), for as-of joins and tick storage (market-data-engineering), for the cost of your own fills (execution-cost-analysis), for indicator look-ahead on bars (signal-construction), for how an engine fills orders or what spread to assume in a bar backtest (backtesting-engines) and for short-sale or margin rules (us-market-rules).
lib-akshare
akshare is the widest free Chinese-market scraper (1,103 public interfaces) and it purges its own PyPI history, so you cannot pin it. TRIGGER - akshare, "import akshare as ak", pip install akshare, stock_zh_a_hist, stock_zh_a_daily, index_stock_cons_csindex, stock_zt_pool_em, stock_zh_a_stop_em, adjust="qfq"/"hfq", 复权, 前复权, 后复权, 涨跌停, 东方财富, 新浪财经, A股数据, 沪深300成分股, "No matching distribution found for akshare==", akshare 报错, akshare 封 IP. akshare ships roughly 2.3 releases a week and deletes the old ones, so any signature, column name or version pin you remember is probably already gone. SKIP for lib-tushare, which is the skill for point-in-time fundamentals and dated index membership. SKIP when the question is WHICH library to choose, or names no library at all - both belong to the domain skill.
lib-alpaca-py
Alpaca's current Python SDK, which defaults to the paper host but lets url_override silently send live orders from a client that believes it is in the sandbox. TRIGGER - import alpaca, pip install alpaca-py, TradingClient, StockHistoricalDataClient, CryptoHistoricalDataClient, submit_order, LimitOrderRequest, MarketOrderRequest, OrderSide, TimeInForce, client_order_id, url_override, paper=True, BaseURL.TRADING_PAPER, paper-api.alpaca.markets, bracket OCO OTO orders, trail_percent, IEX vs SIP feed, Algo Trader Plus, alpaca-trade-api, APCA_API_BASE_URL; an order rejected asynchronously for time-in-force or price precision. Memory is stale here: alpaca-trade-api was deprecated in 2024 and defaulted to LIVE, whereas alpaca-py 0.44.0 (2026-08-11) declares paper=True - the widely repeated warning is inverted. SKIP for Interactive Brokers and for the general order-safety patterns (broker-execution-apis). SKIP for choosing between libraries, or when no library is named - the domain skill's job.
lib-alphalens
alphalens-reloaded scores cross-sectional factors, and its forward return starts at date t's OWN price - it never lags your factor. TRIGGER - alphalens, alphalens-reloaded, "import alphalens as al", get_clean_factor_and_forward_returns, compute_forward_returns, factor_information_coefficient, mean_return_by_quantile, factor_returns, quantile_turnover, factor_rank_autocorrelation, create_full_tear_sheet, MaxLossExceededError, max_loss=0.35, cumulative_returns, information coefficient, IC decay, quantile spread, "pip install alphalens". The original quantopian package is dead at 0.4.0 (2020-04-27) and most snippets you recall target it or its removed pandas internals. SKIP for lib-qlib, which is the skill for the feature pipeline and model. SKIP when the question is WHICH library to choose, or names no library at all - both belong to the domain skill.
lib-arch
The reference GARCH implementation in Python, and the home of SPA/StepM/MCS - which all take LOSSES, so passing returns silently inverts the test and names your worst strategy as the best. TRIGGER - arch, arch_model, arch.bootstrap, arch.univariate, SPA, RealityCheck, StepM, MCS, optimal_block_length, StationaryBootstrap, superior_models, spa.pvalues, mcs.included, arch.unitroot, GARCH, EGARCH, GJR-GARCH, TARCH, APARCH, FIGARCH, HARCH, HAR-RV, skewt, conditional_volatility, or a GARCH fit emitting convergence warnings. Memory is stale on licence and version - it is 8.0.0 (2025-10-21) under NCSA, not one of the three usual permissive licences. SKIP for PSR and DSR (backtest-validation), for PBO and CPCV (lib-purgedcv) and for reporting Sharpe (lib-quantstats). SKIP for choosing between libraries, or when no library is named - the domain skill's job.
lib-backtesting-py
Single-asset bar-loop backtester with honest next-open fills, an AGPL-3.0 licence, and an indicator API that computes over the entire series before slicing. TRIGGER - from backtesting import Backtest, Strategy; pip install backtesting, bt = Backtest(df, MyStrategy), bt.run(), bt.optimize(), self.I(), self.buy(), self.sell(), self.data.Close, trade_on_close, exclusive_orders, finalize_trades, commission, spread, backtesting.lib crossover, _OutOfMoneyError, "kernc"; wanting a portfolio, a universe or a second instrument inside it. Memory is stale here: it is alive at 0.6.6 (2026-07-22), it is AGPL-3.0-or-later rather than MIT, and trade_on_close fills at data.Close[-2] rather than the current bar's close. SKIP for multi-asset or cross-sectional work and for engine choice generally (backtesting-engines). SKIP when the question is WHICH library to choose, or names no library at all - both belong to the domain skill.
lib-ccxt
The unified MIT client for 100+ crypto venues - and not a backtester, with an OHLCV endpoint that silently truncates and returns an unclosed final bar. TRIGGER - import ccxt, import ccxt.pro, import ccxt.async_support, pip install ccxt, fetch_ohlcv, fetchOHLCV, load_markets, fetch_markets, create_order, watchOrderBook, watchTicker, watchMyTrades, set_sandbox_mode, enableRateLimit, amount_to_precision, price_to_precision, exchange.has, options defaultType, parse8601, implicit methods like fapiPrivateGetPositionRisk, CCXT Pro subscription expiry, funding rate history; an order rejected on precision or min-notional, fewer candles returned than requested. Memory is stale here: CCXT Pro was merged into the free MIT package at v1.95, prediction markets landed at 4.5.66, and 4.5.77 shipped 2026-09-01. SKIP for equity and futures brokers (broker-execution-apis). SKIP when the question is WHICH library to choose, or names no library at all - both belong to the domain skill.
lib-edgartools
edgartools is the default free SEC EDGAR client - typed objects for 20+ form types, XBRL statements, no API key - and it 403s on every request until you call set_identity(). TRIGGER - edgartools, "from edgar import Company, set_identity", set_identity, Company("AAPL"), get_filings, filing.xbrl(), get_facts, get_financials, accession number, CIK, ticker-to-CIK, 10-K, 10-Q, 8-K, 13F, Forms 3/4/5, EDGAR full-text search, "HTTP 403" from sec.gov, SEC User-Agent required, SEC 10 requests per second, edgartools MCP server, edgartools[ai]. 442 releases have moved the API repeatedly and the classifier is still Beta, so any snippet recalled from memory is probably wrong for the installed version. SKIP for lib-fredapi, which is the skill for macro series and revisions. SKIP when the question is WHICH library to choose, or names no library at all - both belong to the domain skill.
lib-fredapi
fredapi wraps FRED/ALFRED and is the primary anti-look-ahead tool in macro - and three of its four vintage methods are buggy in source. TRIGGER - fredapi, "from fredapi import Fred", Fred(api_key=), FRED_API_KEY, get_series, get_series_all_releases, get_series_as_of_date, get_series_first_release, get_series_vintage_dates, realtime_start, realtime_end, ALFRED, vintage, data revision, revised GDP, CPI or payrolls, Philadelphia Fed Real-Time Data Set, full-fred, DBnomics, "Bad Request. The value for variable api_key is not registered". Frozen at 0.5.2 since 2024-05, so these bugs are current behaviour, not history you remember from an old version. SKIP for lib-edgartools, which is the skill for company filings and fundamentals. SKIP when the question is WHICH library to choose, or names no library at all - both belong to the domain skill.
lib-freqtrade
freqtrade is a live-first crypto bot with the best bias detectors in the field and a backtester that assumes zero slippage always. TRIGGER - freqtrade, "freqtrade backtesting", freqtrade trade/hyperopt/download-data, lookahead-analysis, recursive-analysis, startup_candle_count, IStrategy, populate_indicators, populate_entry_trend, populate_exit_trend, custom_stoploss, stoploss_on_exchange, minimal_roi, trailing_stop, VolumePairList, StaticPairList, dry_run, dry_run_wallet, config.json, user_data/strategies, FreqAI, freqtrade GPL. Monthly YYYY.M releases have renamed the strategy callbacks repeatedly, so remembered method names are usually the old ones. SKIP for backtesting-engines, the skill for equity and futures bar engines. SKIP when the question is WHICH library to choose, or names no library at all - both belong to the domain skill.
lib-ib-async
The maintained Interactive Brokers Python client - successor to the archived ib_insync - where one digit of the port number is all that separates paper from live. TRIGGER - import ib_async, from ib_async import IB, pip install ib_async, ib.connect, clientId, ports 7496 7497 4001 4002, TWS, IB Gateway, reqHistoricalData, reqMktData, reqTickersAsync, placeOrder, managedAccounts, reqPositions, reqOpenOrders, reqExecutions, Master Client ID, Read-Only API, orderRef, ibflex, ibapi, ib_insync; "Enable ActiveX and Socket Clients", pacing violations, error 1102, a DU or U account prefix. Memory is stale here: ib_insync was archived 2024-03-14 after its author died, ib_async 2.1.0 (2025-12-08) is the successor and does not wrap ibapi, its main branch has been static about nine months, and ib_fut does not exist. SKIP for non-IB brokers and for the general order-safety patterns (broker-execution-apis). SKIP for choosing between libraries, or when no library is named - the domain skill's job.
lib-nautilus-trader
Event-driven Rust-core engine with the strongest execution modelling in open source, gated to Python 3.12-3.14, where a wrong ts_init silently makes every bar visible one interval early. TRIGGER - import nautilus_trader, pip install nautilus_trader, BacktestEngine, BacktestNode, TradingNode, Strategy.on_bar, ts_init, ts_event, FillModel, prob_fill_on_limit, prob_slippage, LatencyModel, base_latency_nanos, RiskEngine, OmsType NETTING HEDGING, BookType.L1_MBP, bar_execution, bar_adaptive_high_low_ordering, high-precision build, ClientOrderId; "could not find a version that satisfies nautilus_trader", a Rust source build on an Intel Mac or Alpine. Memory is stale here: 1.231.0 shipped 2026-08-02, a 2.0 line is in release candidates that moves the fill model to nautilus_trader.execution, and the licence is LGPL-3.0-or-later. SKIP for choosing among engines generally (backtesting-engines). SKIP when the question is WHICH library to choose, or names no library at all - both belong to the domain skill.
lib-polars
The polars wheel is now an empty 865 KB py3-none-any shim hard-pinned to polars-runtime-32, so a lockfile listing only polars does not pin the engine. TRIGGER - polars, "import polars as pl", LazyFrame, scan_parquet, collect(), pl.col, join_asof, group_by, with_columns, polars-runtime-32, polars-runtime-64, polars-lts-cpu, polars 2.0.0rc1, "pip download polars", vendored or air-gapped polars install, polars wheel has no compiled code, porting pandas merge_asof to polars.join_asof, polars sortedness. The runtime split landed at 1.34.0b2 on 2025-09-26, so install matrices, wheel audits and lockfiles written from memory are wrong. SKIP for market-data-engineering, the skill for storage formats and time-series stores. SKIP when the question is WHICH library to choose, or names no library at all - both belong to the domain skill.
lib-purgedcv
The only genuinely sklearn-protocol-compliant purged and embargoed splitter, and the one that refuses to run until you state when each label resolved - understate evaluation_times and it silently reintroduces the leak while reporting clean folds. TRIGGER - purgedcv, PurgedKFold, PBO, CPCV, probability of backtest overfitting, CombinatorialPurgedCV, PurgedGroupKFold, WalkForwardSplit, prediction_times, evaluation_times, purge_horizon, embargo, audit_splitter, reconstruct_paths, path_metrics, overlapping labels, triple-barrier touch time, mlfinlab.cross_validation. Memory is stale or absent - this package first shipped 2026-05-16 and is at 0.1.6. SKIP for CPCV over portfolio models (lib-skfolio) and for SPA/StepM/MCS (lib-arch). SKIP when the question is WHICH library to choose, or names no library at all - both belong to the domain skill.
lib-pyportfolioopt
Textbook mean-variance and Black-Litterman optimizer whose HRPOpt silently accepts a price matrix where it requires returns and returns plausible garbage. TRIGGER - pypfopt, PyPortfolioOpt, EfficientFrontier, HRPOpt, CovarianceShrinkage, DiscreteAllocation, BlackLittermanModel, EfficientCVaR, EfficientSemivariance, CLA, mean_historical_return, capm_return, clean_weights, max_sharpe, min_volatility, portfolio_performance, risk_models.risk_matrix, "efficient frontier", "whole-share allocation". Memory is stale - the repo moved to the PyPortfolio org and 1.6.0 shipped 2026-02-26 after three dormant years under a new maintainer. SKIP for Marcenko-Pastur denoising, HERC or NCO (lib-riskfolio) and for GridSearchCV over portfolio models (lib-skfolio). SKIP for choosing between libraries, or when no library is named - the domain skill's job.
lib-qlib
Microsoft Qlib (pip name pyqlib, imported as qlib) ships Alpha158/Alpha360 and a default normalizer that leaks your test set into training, silently. TRIGGER - qlib, pyqlib, "pip install pyqlib", qlib.init, provider_uri, REG_CN, Alpha158, Alpha360, DataHandlerLP, ZScoreNorm, CSZScoreNorm, fit_start_time, fit_end_time, qrun, workflow_config yaml, "qlib_data --region cn", LABEL0, "Ref($close, -2)", qlib expression engine, qlib .bin dataset, investment_data qlib_bin. Wheels are cp38-cp312 with no sdist so it fails outright on Python 3.13+, packaging lags the repo by about a year, and the official China dataset was switched off. SKIP for lib-alphalens, which is the skill for scoring a factor you already have. SKIP when the question is WHICH library to choose, or names no library at all - both belong to the domain skill.
lib-quantlib
The only broadly-permissive, mature, full-coverage derivatives library in Python, whose global evaluationDate returns an NPV of exactly 0.0 with no warning once it is past expiry. TRIGGER - QuantLib, "import QuantLib as ql", QuantLib-SWIG, QuantLib-Python, Settings.instance(), evaluationDate, ql.Date, YieldTermStructureHandle, VanillaOption, AmericanExercise, QdFpAmericanEngine, FdBlackScholesVanillaEngine, BinomialVanillaEngine, BaroneAdesiWhaleyApproximationEngine, SABRInterpolation, SviSmileSection, thetaPerDay, "NPV is zero". Memory is stale on packaging and engine names - it is at 1.43 and ships 26 wheels but no sdist. SKIP for vanilla European IV and Greeks in pure Python (lib-vollib). SKIP for choosing between libraries, or when no library is named - the domain skill's job.
lib-quantstats
The tearsheet library whose cagr(rf=...) accepts your risk-free rate and silently discards it - "cagr" sits on an exclusion list inside _prepare_returns, which dispatches on the caller's function name. TRIGGER - quantstats, "import quantstats as qs", qs.reports.html, qs.stats.sharpe, qs.stats.cagr, qs.stats.value_at_risk, expected_shortfall, gain_to_pain_ratio, rolling_volatility, qs.extend_pandas, tearsheet, quantstats-lumi; or a wildly negative Sharpe. Memory is stale on status and correctness - 0.0.81 shipped in a single-day hotfix burst on 2026-01-13 with no default-branch commits since, and the cagr bug survived it. SKIP for optimizing against these measures (lib-riskfolio, lib-skfolio) and for PSR/DSR, which it does not have (backtest-validation). SKIP when the question is WHICH library to choose, or names no library at all - both belong to the domain skill.
lib-riskfolio
The 26-risk-measure portfolio optimizer whose stateful API optimizes against stale or missing mu and Sigma - with no error - if you forget assets_stats(). TRIGGER - riskfolio, Riskfolio-Lib, "import riskfolio as rp", rp.Portfolio, rp.HCPortfolio, assets_stats, port.optimization, hcp.optimization, model="HRP"/"HERC"/"NCO", rm="CVaR"/"CDaR"/"EVaR"/"RLVaR"/"EDaR"/"RLDaR", denoiseCov, riskfolio.src.AuxFunctions, ParamsEstimation, entropy_pooling, OWA, MVSK, "solver did not converge". Memory is stale - it is at 7.3.0 (2026-05-31) with an unusual 18 open issues against 4,480 stars. SKIP for GridSearchCV over portfolio models (lib-skfolio) and for whole-share allocation (lib-pyportfolioopt). SKIP when the question is WHICH library to choose, or names no library at all - both belong to the domain skill.
lib-skfolio
The sklearn-compatible portfolio estimator library whose CombinatorialPurgedCV breaks sklearn's own split() contract - it yields (train, [test_0, ...]), and normal two-variable unpacking mis-partitions your data without raising. TRIGGER - skfolio, skfolio.optimization, MeanRisk, RiskBudgeting, HierarchicalRiskParity, HierarchicalEqualRiskContribution, NestedClustersOptimization, skfolio.moments, DenoiseCovariance, GerberCovariance, ImpliedCovariance, EmpiricalPrior, EntropyPooling, VineCopula, CombinatorialPurgedCV, WalkForward, purged_size, embargo_size, RiskMeasure. Memory is stale and will break code - 1.0.0 landed 2026-08-23, so every recalled snippet predates the API stability commitment. SKIP for the widest risk-measure menu (lib-riskfolio) and for a strictly sklearn-compliant purged splitter (lib-purgedcv). SKIP for choosing between libraries, or when no library is named - the domain skill's job.
lib-talib
The C reference implementation of technical indicators, where every pure-Python port disagrees during warm-up and none of them say so. TRIGGER - import talib, pip install TA-Lib, ta-lib-python, talib.RSI, talib.MACD, talib.ATR, talib.ADX, talib.BBANDS, talib.OBV, from talib import abstract, talib.get_functions, unstable period, indicator warm-up, an indicator differing between two libraries or between backtest and live; "Exception: input array type is not double", a failed ta-lib C build or missing ta_libc.h, and. Memory is stale here: the install pain is solved - 0.7.1 (2026-07-16) ships 54 prebuilt wheels bundling the C library, including cp311-win_amd64 - while pandas-ta's repo, homepage and release history are all gone. SKIP for whether a signal actually predicts returns (factor-and-timeseries-research) and for leak-free signal construction generally (signal-construction). SKIP for choosing between libraries, or when no library is named - the domain skill's job.
lib-tushare
tushare is the cheapest source of genuinely point-in-time A-share fundamentals, and it sends your token over plaintext HTTP. TRIGGER - tushare, tushare pro, "import tushare as ts", ts.pro_api, pro_bar, adj="qfq", stock_basic, list_status, daily_basic, adj_factor, income, balancesheet, f_ann_date, ann_date, update_flag, 报告期, 公告日, tushare token, 积分, waditu, api.waditu.com, "抱歉，您没有接口访问权限", tushare 权限不够. The public GitHub repo has been idle since 2024-03 while PyPI kept shipping through 2026, so recalled behaviour does not match the installed wheel. SKIP for lib-akshare, which is the skill for breadth of free Chinese coverage rather than PIT. SKIP when the question is WHICH library to choose, or names no library at all - both belong to the domain skill.
lib-vectorbt
Vectorized Numba/Rust backtester built for parameter sweeps, whose from_signals fills at the signal's own bar close by default. TRIGGER - import vectorbt as vbt, pip install vectorbt, vbt.Portfolio.from_signals, from_orders, from_holding, ma_crossed_above, vbt.MA.run, vbt.IndicatorFactory, .vbt.signals.fshift, price=np.inf, reject_prob, allow_partial, stop_conflict_mode, cash_sharing, FlexArray, vectorbt[rust], VectorBT PRO, "this PRO example does not work"; a 10,000-combination grid, "my backtest looks too good", an equity curve that dies live. Memory is stale here: v1.0 (2026-04-22) was a breaking rewrite with an optional Rust engine, 1.1.0 shipped 2026-07-05, and the licence is Apache-2.0 plus Commons Clause - not OSI open source. SKIP for choosing among engines (backtesting-engines) and for judging a finished result (backtest-validation). SKIP when the question is WHICH library to choose, or names no library at all - both belong to the domain skill.
lib-vollib
Machine-precision implied volatility with no bracketing, behind a package name restructured in 2026 - py_vollib is now a DEAD SHIM with four files and zero library code, and every pre-2026 tutorial installs it. TRIGGER - vollib, py_vollib, py_vollib_vectorized, lets_be_rational, "Let's Be Rational", Jaeckel, black_scholes, black_scholes_merton, implied_volatility, greeks.analytical, BelowIntrinsicException, AboveMaximumException, "py_vollib is deprecated", or implied volatility returning 0.0. Memory is stale on the package name, on the C++ dependency (it is now pure Python) and on the Greek scaling. SKIP for American exercise, exotics, curves or conventions (lib-quantlib). SKIP when the question is WHICH library to choose, or names no library at all - both belong to the domain skill.
lib-yfinance
The default free Yahoo Finance downloader, whose yf.download() now returns pre-adjusted OHLC with no Adj Close column at all. TRIGGER - import yfinance as yf, pip install yfinance, yf.download, yf.Ticker, Ticker.history, auto_adjust, multi_level_index, ignore_tz, repair=True, get_shares_full, yf.Search, yf.Lookup, yf.WebSocket, yfinance-cache; errors "KeyError: 'Adj Close'", YFRateLimitError, "Too Many Requests. Rate limited", YFTickerMissingError, "possibly delisted", curl_cffi pin conflicts, one ticker returning MultiIndex columns. Memory is stale here: auto_adjust flipped at 0.2.51 and hardened at 1.0, intraday timezones changed at 1.4.0, the proxy= constructor kwarg is gone, and 1.7.0 shipped 2026-08-26. SKIP for choosing between data vendors (market-data-sourcing) and for A-share data (china-ashare-data). SKIP when the question is WHICH library to choose, or names no library at all - both belong to the domain skill.
llm-finance-agents
What the published evidence says about LLM trading agents, and the real status of the frameworks. TRIGGER - TradingAgents, FinGPT, FinRobot, FinMem, FinCON, FinAgent, AlphaAgent, RD- Agent, AI4Finance; building or evaluating an LLM-driven trading system, a multi-agent trader, or a news-sentiment-to-signal pipeline; "does AI trading work"; FinBERT and financial sentiment models; reproducing a Sharpe from an LLM-trading paper; whether a backtest window overlaps a model's training cutoff. No credible evidence exists that any of it produces alpha net of costs. SKIP for reinforcement learning and deep learning specifically (rl-and-ml-trading) and for MCP servers (finance-mcp-servers).
market-data-engineering
Store, join and parallelize market data you already hold, without corrupting it. TRIGGER - as-of join, merge_asof, join_asof, ASOF JOIN, "join quotes to trades", aligning signals to prices; reading or writing Parquet, Feather, HDF5 or CSV of market data; choosing between pandas, polars, DuckDB, pyarrow, dask or ray; a time-series store such as ArcticDB, QuestDB, ClickHouse, TimescaleDB or kdb; storing years of minute bars for thousands of tickers; a dataset too big for memory; partitioning; timestamps or timezones coming back wrong; float precision on prices or volume; "different numbers when I parallelise". SKIP for choosing a data VENDOR (market-data-sourcing) - this skill starts once the bytes are yours.
market-data-sourcing
Choose a market price or reference data vendor and use it without silently corrupting the numbers. TRIGGER - download, fetch, pull or load OHLCV, prices, quotes, bars or a ticker universe; compare vendors on cost, coverage or free-tier limits; need delisted US or global tickers, or a survivorship-free universe; two sources disagree; hitting 429 or rate limits; "KeyError: Adj Close"; split and dividend adjustment; trading calendars and holidays. Covers yfinance, yahooquery, defeatbeta, EODHD, Tiingo, Twelve Data, Finnhub, Alpha Vantage, Polygon/Massive, Databento, openbb, findatapy, financetoolkit, exchange_calendars, and alternative data. Also covers 美股 and global 行情数据 requests. SKIP for historical option chains (options-backtesting), Asian calendars and lot sizes (asia-pacific-markets), storing, partitioning or as-of joining data you already hold (market-data-engineering); for EDGAR filings, XBRL, CIK and macro vintages (fundamental-and-macro-data); and for A-share, 沪深 or 退市 queries (china-ashare-data).
options-backtesting
Options positions end in ways you do not control - live or in a backtest: assignment, expiry settlement, pin risk, multi-leg lifecycle, historical chain assembly, and the margin that decides whether the position fits. TRIGGER - backtest a covered call, cash-secured put, wheel, credit spread, iron condor, butterfly, calendar, diagonal, straddle, strangle, PMCC; short option assigned, early exercise, exercise by exception, expires in the money, pin risk, pinned at the strike; historical option chain, options history, chain panel, OSI symbol, adjusted option, non-standard deliverable; 0DTE, weeklies, third Friday, AM vs PM settlement, cash settled index options; option margin, naked margin requirement, portfolio margin, SPAN, buying power reduction; "my options backtest returns look too good"; optopsy, optionlab, an options backtesting library. SKIP for pricing a single option or fitting a vol surface (derivatives-pricing) and for futures rolls (futures-continuous-contracts).
portfolio-and-risk
Turn signals into weights, and compute performance metrics that are actually correct. TRIGGER - portfolio weights, allocation, rebalancing, mean-variance, Black-Litterman, risk parity, HRP, HERC, NCO, efficient frontier, covariance shrinkage or denoising, PyPortfolioOpt, riskfolio, skfolio, cvxportfolio; or computing Sharpe, Sortino, Calmar, CAGR, annualized volatility, max drawdown, VaR, CVaR, beta, alpha, a tearsheet, quantstats, pyfolio, empyrical, ffn, or performance attribution. Load before quoting any performance number: popular libraries disagree on identical input, one silently discards the risk-free rate you pass it, and an absurdly negative Sharpe has one known cause. SKIP for whether the result survives multiple testing (backtest-validation).
quant-stack-router
Entry router for Python quantitative finance: names the right library and flags where the model's training prior is stale. TRIGGER - read this FIRST when a quant task names no specific sub-domain, when asked "which library should I use" for anything financial, or when a plan spans several of data, backtest, risk and execution. Also read it before importing any finance package whose status you are assuming from memory - several widely used ones are dead, relicensed, or have flipped a default since training. SKIP when the task already names its own domain - go straight to market-data-sourcing, backtesting-engines, broker-execution-apis, portfolio-and-risk, factor-and-timeseries-research, china-ashare-data or crypto-data-and-execution rather than routing through here.
regime-detection
Detect and label market regimes without letting the labels see the future, and state regime coverage in the form the result gate demands. TRIGGER - detect market regimes, regime detection, bull bear regime labels, volatility regime, high-vol low-vol state, risk-on risk-off; hidden markov model on returns, HMM, hmmlearn, markov switching, MarkovRegression, smoothed vs filtered probabilities; change point detection, ruptures, structural break; turbulence index, Mahalanobis distance; "my strategy only works in one regime", "does it survive 2008 or 2020"; "result_manifest says no regime coverage", regimes_covered. SKIP for forecasting volatility itself with GARCH or arch (factor-and-timeseries-research), for whether a regime-conditional result survives the trials behind it (backtest-validation), for the full pre-report audit (research-integrity-guards), and for RL or deep-learning state models (rl-and-ml-trading).
research-integrity-guards
Second-pass audit that decides whether a finance result is real, applied after the work exists. TRIGGER - about to REPORT, publish or act on a backtest, factor test or model score; a result that looks good ("Sharpe 3.5", "beats SPY", "85% accuracy") and needs challenging; asked to validate, verify, sanity-check or critique a research design; asked "what should I check". Covers five gates: universe survivorship, availability timestamps, label leakage, cost realism, trial count. SKIP when the task is to BUILD something rather than judge it - go to the domain skill first (market-data-sourcing, backtesting-engines, factor-and-timeseries-research) and return here before reporting a number.
after-tax-backtesting
Attach lot matching, wash sales and section 1256 to an existing backtest and report after-tax Sharpe beside pre-tax - and refuse to report one that does not state its rate, jurisdiction and lot method. TRIGGER - after-tax return, after tax Sharpe, tax drag, tax-aware backtest, tax alpha, tax-managed strategy, turnover penalty, "what does this strategy return after tax", tax cost of rebalancing, taxable account backtest, tax-efficient turnover, pre-tax vs post-tax performance, capital gains netting, loss carryforward, "should I hold this in an IRA". Modelling assumptions for backtests, not tax advice. SKIP for the lot rules themselves (tax-lot-matching-and-cost-basis), for the wash-sale mechanics (wash-sale-rules), for futures and index options (section-1256-and-derivatives-tax), for A-share taxes (china-ashare-trading-taxes), and for transaction costs, which are a different drag entirely (execution-cost-analysis).
alpha-combination-and-neutralization
Score several alphas, combine them, and strip the exposures you did not mean to take. TRIGGER - information coefficient, IC, rank IC, ICIR, IC decay, "is my IC good", IC t-stat, Newey-West on IC, overlapping forward returns; combining alphas, blending signals, alpha weighting, z-score or rank combination, multi-factor signal; sector neutral, beta neutral, market neutral signal, industry neutralization, residualize the alpha, cross-sectional regression residuals, orthogonalize signals; winsorize, clip outliers, cross-sectional standardization; turnover-aware combination, signal smoothing, "my alpha dies after costs". SKIP for one time-series trend signal (trend-following-models), for the factor library and the alphalens forward-return convention (factor-and-timeseries-research, lib-alphalens), for weights under constraints (portfolio-and-risk), and for whether the survivor is real (backtest-validation).
bet-sizing
Turn a predicted probability into a position - the 2*Phi(z)-1 size curve, averaging concurrent bets instead of adding them, discretising to buy turnover, and the concurrency budget whose divisor is usually a look-ahead. TRIGGER - bet sizing, getSignal, get_signal, getBetSize, bet size from probability, avgActiveSignals, average active bets, discreteSignal, discrete signal, step size, position from predict_proba, "how big should this trade be", "my positions flip every bar", turnover from a probability, concurrent bets leverage, budgeting bets, Lopez de Prado chapter 10, AFML bet sizing. SKIP for the Kelly fraction and how large the book should be overall (position-sizing-kelly - it owns Kelly, do not repeat it), for the secondary model that produces the probability (meta-labeling), for volatility targeting and risk budgets (portfolio-and-risk), and for the execution schedule once the size is chosen (execution-algorithms).
bond-conventions-and-accrued
Compute accrued interest, clean and dirty prices and day-count year fractions on a bond without silently picking the wrong convention. TRIGGER - accrued interest, day count, daycount, ACT/ACT ICMA vs ISMA vs ISDA, ACT/365F, ACT/360, 30/360, 30E/360, 30E/360 ISDA, Thirty360 BondBasis vs USA vs European vs NASD, year fraction, yearFraction, ActualActual(ISMA), "my accrued interest is off by a few hundred dollars", "which 30/360 is this", clean price vs dirty price vs invoice price, settlement amount, T+1 settlement, quasi-coupon date, first and last stub period. SKIP for negative accrued inside a gilt ex-dividend window (ex-dividend-and-rebate-interest), for turning a price into a yield (yield-measures-and-bill-quotes), for duration and DV01 (duration-convexity-and-dv01), for discount-curve conventions and compounding (../../../fin-models/skills/term-structure-models), and for QuantLib's evaluationDate global (../../../fin-libraries/skills/lib-quantlib).
cds-mechanics-and-upfront
Turn a CDS quote into the cash that actually changes hands - standard coupons, points upfront, the risky annuity, the IMM roll and the accrual rebate. TRIGGER - points upfront, upfront payment on a CDS, convert spread to upfront, conventional spread, quoted spread vs par spread, ISDA CDS Standard Model, cdsmodel.com, CDS converter, flat hazard quoting convention, RPV01, risky PV01, risky annuity, "upfront = spread difference times duration", 100bp or 500bp coupon, SNAC, IMM dates, CDS roll, 20 Mar/Jun/Sep/Dec, "why is my 5y CDS maturing in June", accrual on default, accrual rebate, accrued coupon on a CDS, cdsMaturity, IsdaCdsEngine, MidPointCdsEngine, ISDA standard rate curve, rfr.spglobal.com. SKIP for hazard-rate and default-probability modelling, Merton and recovery sensitivity (credit-risk-models), for bond Z-spreads, G-spreads and OAS (credit-spread-measures), for the bond tape and TRACE (corporate-bond-data-and-trace), and for building the discount curve itself (term-structure-models).
china-ashare-trading-taxes
A-share stamp duty is charged to the seller only and halved on 2023-08-28, and dividend tax is a step function of holding period - a turnover penalty written into the tax code that a flat symmetric cost model cannot express. TRIGGER - A股印花税, 证券交易印花税, stamp duty, stamp tax, seller side only, 单边征收, 印花税减半, 0.05%, 2023-08-28, A股股息红利税, 股息红利差别化, dividend tax by holding period, 持股期限, 一个月, 一年, 财税〔2015〕101号, dividend capture China, A-share transaction cost model, "what does an A-share round trip actually cost", backtesting Chinese equities net of tax. Modelling assumptions for backtests, not tax advice, and for an individual resident investor only. SKIP for US wash sales and lot matching (wash-sale-rules, tax-lot-matching-and-cost-basis), for reporting an after-tax Sharpe (after-tax-backtesting), and for A-share data, T+1 and price limits (china-trading-stack).
copulas-and-dependence
Separate the marginals from the dependence - Gaussian, Student t, Clayton and Gumbel copulas, Kendall's tau, tail dependence coefficients, and what fitting the wrong family costs in the joint tail. TRIGGER - copula, Gaussian copula, t copula, Student t copula, Clayton copula, Gumbel copula, Archimedean copula, Sklar's theorem; tail dependence, lower tail dependence, upper tail dependence, lambda_U, "correlation is not dependence", "correlations go to one in a crisis", joint tail probability, joint exceedance; Kendall's tau, Spearman rho, rank correlation, pseudo-observations, inversion of Kendall's tau, copula MLE, copulas python, copulae, statsmodels copula; diversification benefit, "my VaR says the portfolio is safe". SKIP for estimating a covariance matrix and shrinkage (covariance-and-risk-models), for VaR/CVaR methods and their backtests (risk-measures-var-cvar), and for GARCH marginals (volatility-models).
corporate-bond-data-and-trace
Use FINRA TRACE corporate bond data without inheriting the two things it does not tell you - the 15-minute reporting window and the size caps that censor volume. TRIGGER - TRACE, FINRA trade reporting, corporate bond tape, bond transaction data, "how fast are bond trades reported", one-minute TRACE reporting, Rule 6730, 15-minute reporting, dissemination cap, "5MM+", "1MM+", capped trade size, bond volume, TRACE academic or historic files, bond turnover, Amihud illiquidity on bonds, bond VWAP, "why is my bond volume so low", corporate bond liquidity screen, WRDS bond data, PyBondLab. SKIP for turning a bond price into a spread (credit-spread-measures), for CDS quotes and upfronts (cds-mechanics-and-upfront), for the rating that put the bond in an index (ratings-transitions-and-migration), for equity tick data and TAQ (intraday-microstructure), and for choosing a market data vendor or API in general (market-data-sourcing).
covariance-and-risk-models
Estimate a covariance matrix an optimizer can actually invert, and report how much variance it hides. TRIGGER - covariance matrix estimation, sample covariance singular, "matrix is not positive definite", np.cov more assets than observations, N > T, condition number, Ledoit-Wolf shrinkage, sklearn LedoitWolf, CovarianceShrinkage, shrinkage intensity or delta, RiskMetrics EWMA, lambda 0.94 or 0.97, exponentially weighted covariance, exp_cov span, PCA or statistical factor risk model, Marchenko-Pastur, Barra fundamental factor model, specific risk, predicted vs realized volatility, risk model bias test; "my minimum-variance portfolio has 90x leverage", "the optimizer says 0% risk". SKIP for turning a covariance into weights and the optimizers themselves (portfolio-optimizers), for VaR, Expected Shortfall and their backtests (risk-measures-var-cvar), for GARCH and univariate volatility forecasting (volatility-models), and for Sharpe and drawdown conventions (portfolio-and-risk).
credit-risk-models
Estimate a default probability and price credit, and keep the two probabilities apart - the risk-neutral one that prices and the physical one that forecasts. TRIGGER - Merton model, structural credit model, KMV, distance to default, asset value and asset volatility from equity, N(-d2), solve the two Merton equations; hazard rate, intensity, reduced form, survival probability, constant hazard, credit curve bootstrapping; CDS par spread, premium leg, protection leg, risky PV01, RPV01, accrual on default, "spread = lambda times one minus recovery", implied hazard from a CDS spread, recovery assumption, 40% recovery; risk-neutral vs physical default probability, rating agency default table, "my CDS spread is too low", credit spread from a bond price, expected loss, CVA default probability. SKIP for option pricing and Greeks (option-pricing-models, derivatives-pricing), for interest-rate curves and short-rate models (term-structure-models), and for portfolio risk and VaR (portfolio-and-risk).
credit-spread-measures
Work out which spread a corporate bond quote actually is and what it was measured against, so two "spreads" on the same bond stop disagreeing. TRIGGER - Z-spread, I-spread, G-spread, benchmark spread, spread to Treasuries, asset swap spread, ASW, par/par asset swap, discount margin, DM on a floater, quoted margin, OAS, option-adjusted spread, option cost, static spread, zero-volatility spread, "my Z-spread and my G-spread disagree", "is this spread over Treasuries or over swaps", "YTM minus the 5-year Treasury", BondFunctions.zSpread, "why is my OAS lower than my Z-spread", spread on a callable bond. SKIP for CDS spreads, points upfront and the ISDA model (cds-mechanics-and-upfront), for hazard rates, Merton and default probability (credit-risk-models), for building, bootstrapping or interpolating the underlying curve (term-structure-models), and for where the price and the trade came from (corporate-bond-data-and-trace).
duration-convexity-and-dv01
Get the right duration number and the right DV01, for a bond, a floater or a hedge ratio. TRIGGER - Macaulay vs modified duration, effective duration, spread duration, key rate duration, DV01, PV01, BPV, dollar duration, basis point value, convexity, "my hedge ratio is off by a few percent", "duration says the price should be X but it is Y", duration times spread, floating rate note duration, FRN duration, "why is my floater duration almost zero", BondFunctions.duration, Duration.Macaulay vs Duration.Modified vs Duration.Simple, basisPointValue sign. SKIP for turning a price into a yield in the first place (yield-measures-and-bill-quotes), for accrued interest and day counts (bond-conventions-and-accrued), for the swap annuity and PV01 under OIS discounting (ois-discounting-and-multi-curve), and for portfolio VaR and risk aggregation (../../../fin-core/skills/portfolio-and-risk).
ex-dividend-and-rebate-interest
Handle bonds that trade ex-dividend, where accrued interest goes negative and the buyer is paid rebate interest instead of paying it. TRIGGER - gilt, UK gilt, ex-dividend, ex-div, ex-coupon, exCouponPeriod, rebate interest, negative accrued interest, "my accrued interest is negative", "accrued should be negative but isn't", seven business days before the coupon, quasi-coupon date, DMO formulae, "Formulae for Calculating Gilt Prices from Yields", ql.FixedRateBond exCouponPeriod, ql.Period(-7, ql.Days), record date vs ex-date on a bond, 3.5% War Loan, JGB and gilt settlement. SKIP for ordinary positive accrued and day-count choice (bond-conventions-and-accrued), for price-to-yield solving in general (yield-measures-and-bill-quotes), for index-linked gilt indexation lags (../../../fin-models/skills/term-structure-models), and for QuantLib's evaluationDate global (../../../fin-libraries/skills/lib-quantlib).
execution-algorithms
Build the schedule that works an order - VWAP, TWAP, POV, Almgren-Chriss - and know what each one is optimizing. TRIGGER - VWAP algo, TWAP algo, POV, percentage of volume, participation rate, child order slicing, execution schedule, order scheduling, "how should I work this order", intraday volume profile, U-shaped volume curve; Almgren-Chriss, optimal execution, optimal liquidation, trading trajectory, efficient frontier of execution, risk aversion lambda, trade half-life, kappa, market impact model, temporary vs permanent impact, square root law; implementation shortfall, Perold, arrival price, decision price, delay cost, opportunity cost, unfilled shares. SKIP for measuring fills you already have and choosing a benchmark after the fact (execution-cost-analysis), for spreads, trade classification and order-flow measures (intraday-microstructure), for quoting rather than taking (market-making-models), and for broker order types and routing (broker-execution-apis).
factor-models
Build long-short factor portfolios from a characteristic panel and test the alpha with standard errors that survive serial correlation. TRIGGER - factor model, Fama-French, Fama-MacBeth, cross-sectional regression, decile or quintile long-short sort, 2x3 sort, SMB and HML, value- weight vs equal-weight portfolio, characteristic panel, alpha t-stat, Newey-West, HAC standard errors, cov_type="HAC" maxlags, Ken French Data Library, F-F_Research_Data_Factors, book-to- market, 11-1 momentum; "my factor has a t-stat of 15", "should I lag the signal", "my HML does not match Ken French", "joining monthly factors to daily returns". SKIP for scoring one alpha signal with alphalens, IC decay or GARCH (factor-and-timeseries-research), for the covariance matrix a factor model implies (covariance-and-risk-models), for turning expected returns into weights (portfolio-optimizers), and for counting the specifications you tried (backtest-validation).
feature-importance-financial
Rank features without believing MDI - it is in-sample, it favours columns with many distinct values, and it splits credit between substitutable features; MDA under-states collinear pairs and leaks outright on a shuffled k-fold. TRIGGER - feature_importances_, feature importance, MDI, mean decrease impurity, Gini importance, MDA, mean decrease accuracy, permutation importance, permutation_importance, single feature importance, SFI, clustered feature importance, "which features matter", "my random forest says this random column is important", "importance changes every run", correlated features importance, substitution effect, Lopez de Prado chapter 8, AFML feature importance. SKIP for purged and embargoed cross-validation itself (lib-purgedcv), for the overlapping-label weights (sample-weights-and-uniqueness), for feature construction and causality (signal-construction), for factor-return attribution (factor-models), and for the trials a feature search spends (backtest-validation).
fin-skills-as-tools
How to hand this library to an agent as TOOLS rather than as reading - the MCP server, the exported Anthropic and OpenAI tool definitions, the JSON payload conventions, and the four guards that cannot cross a JSON boundary. TRIGGER - expose fin-skills to an LLM agent; run the fin-skills MCP server; "python -m fin_skills.mcp"; fin-skills-mcp; import the guards as function-calling tools; generate tool schemas for the Messages API or the OpenAI Responses API; send a pandas Series or DataFrame through a tool call; check_backtest, bundle_coverage, describe_guard, read_skill; why assert_causal or warmup_probe is missing from the tool list; payload size limits for a tool call; 把 guard 挂成 agent 工具. SKIP for how to design the agent pipeline itself (finance-agent-architectures), which third-party finance MCP servers exist and which can move money (finance-mcp-servers), and whether an LLM strategy makes money at all (llm-finance-agents).
fractional-differentiation
Make a price series stationary without throwing away the memory a model needs - the weight recursion, the fixed-width window, and the scan for the smallest d that passes ADF. TRIGGER - fractional differentiation, fractional differencing, fracdiff, frac_diff, frac_diff_ffd, get_weights_ffd, plotMinFFD, plot_min_ffd, "minimum d", ARFIMA, long memory, (1-B)^d, binomial weights, fixed-width window fracdiff; "my model only sees returns", "prices are non-stationary so I differenced them", "ADF says my feature is non-stationary", "should I feed prices or returns to the model", Lopez de Prado chapter 5, AFML fracdiff. SKIP for the cointegration ADF table and the fitted-residual null (stat-arb-cointegration), for unit-root and ARIMA model selection generally (time-series-forecasting-models), for making a feature causal and its warm-up (signal-construction), and for labels rather than features (triple-barrier-labeling).
gdp-nowcasting-dynamic-factor
Nowcast the quarter you are in from monthly data with a ragged edge, using statsmodels' DynamicFactorMQ - and score it against the benchmarks it has to beat. TRIGGER - nowcast, nowcasting, GDPNow, Atlanta Fed GDP tracker, New York Fed Staff Nowcast, DynamicFactorMQ, endog_quarterly, k_endog_monthly, fit_em, dynamic factor model, mixed frequency, monthly and quarterly in one model, ragged edge, jagged edge, unbalanced panel, Mariano-Murasawa, Banbura Modugno, bridge equation, MIDAS, "how do I combine monthly indicators into a GDP forecast", news decomposition of a data release. SKIP for the Kalman filter and smoother themselves (state-space-and-kalman), for univariate forecasting and its baselines (time-series-forecasting-models), for release timestamps (macro-release-calendar-and-embargo), and for vintages of the inputs (real-time-macro-backtesting).
hawkes-processes
Fit and test a self-exciting point process for clustered order arrivals - exponential-kernel Hawkes intensity, Ogata thinning, maximum likelihood, the branching ratio, and the random time change that tests the fit. TRIGGER - Hawkes process, self-exciting point process, order arrival clustering, trade clustering, order flow clustering, mutually exciting, branching ratio, alpha over beta, criticality, endogeneity of market activity; Ogata thinning, simulate a point process, tick, hawkeslib, conditional intensity; Hawkes MLE, log-likelihood recursion, random time change, residual analysis, "are my arrivals Poisson", overdispersed counts, Fano factor, "my Poisson confidence interval is too narrow". SKIP for the birth-death queue model of a single price level (limit-order-book-models), for measuring realised activity from a tape (intraday-microstructure), for GARCH and volatility clustering in returns rather than arrivals (volatility-models), and for regime switching (regime-detection).
implied-vol-surface
Build a volatility surface that is not silently arbitrageable - invert prices to implied vols, fit a smile, check butterfly and calendar arbitrage, and interpolate between maturities. TRIGGER - implied volatility solver, Newton diverges, bisection bracket, "implied vol returns 0.001", "no implied volatility for this option", price below intrinsic, BelowIntrinsicException; SVI, raw SVI, Gatheral, svi calibration, a b rho m sigma, SviSmileSection; volatility smile, skew, surface fitting, total variance, log-moneyness; butterfly arbitrage, negative implied density, Durrleman g(k), calendar spread arbitrage, static arbitrage check, Gatheral and Jacquier 2014; interpolating the vol surface, "my interpolated surface has arbitrage", "vol interpolation between expiries". SKIP for pricing one option and the models themselves - Heston, CRR, SABR, Monte Carlo (option-pricing-models), for library choice, Greek units and licences (derivatives-pricing), and for option chain data and historical chains (options-backtesting).
backtest-overfitting
Decide whether an edge that passed every mechanical check is still just the best of N tries. TRIGGER - PBO, probability of backtest overfitting, CSCV, combinatorially symmetric cross validation, logit of the out-of-sample rank; minimum backtest length, MinBTL, "how much history do I need", "how many parameter sets can I try on N years of data"; a grid search, Optuna or AutoML picked a winner and it decayed; "in-sample Sharpe 2, live Sharpe 0", "the best parameter set stopped working", "is this peak on my parameter surface real"; pypbo, RiskLabAI CSCV. SKIP for the deflated and probabilistic Sharpe ratios, the trial ledger and SPA/StepM/MCS against a benchmark (backtest-validation); for family-wise error and false-discovery control over a whole research programme (multiple-testing-ledger); for the purged and combinatorial CV splitters themselves (lib-purgedcv); for the mechanical leakage gates that come first (research-integrity-guards); and for regime coverage of the test window (regime-detection).
choosing-a-data-vendor
Decide whether a data source may legally and factually serve a research question, before any fetch code is written. TRIGGER - need delisted, survivorship-free or point-in-time data and are choosing where to get it; may I store, cache, redistribute or publish what I fetched; is the free tier enough and what does a key cost; comparing vendor terms, licences, rate limits or paid tiers; "there are no delisted names on the free tier"; picking between yfinance, Tiingo, Alpha Vantage, stooq, EODHD, Norgate, CRSP or Polygon. Automated by `python -m fin_skills.data advise`. SKIP once the source is chosen and the question is how to CALL it - adjustment, timezone, calendar and off-by-one traps (market-data-sourcing); storing, partitioning or as-of joining data you already hold (market-data-engineering); EDGAR, XBRL and macro vintages (fundamental-and-macro-data); A-share sources and 退市 lists (china-ashare-data); exchange OHLCV and venue limits (crypto-data-and-execution).
combining-data-sources
Combine information of DIFFERENT kinds into one research view whose every number can be traced back to what was knowable when. TRIGGER - joining prices to fundamentals, earnings or a macro series; building a signal from a filing plus a price; merging two vendors into one series; "which vendor do I trust"; combine_first / fillna / bfill across sources; assembling a company dossier, profile or research view from several places; a field that must carry its provenance; "when did this number become knowable"; period_end vs filed vs available; a ticker that changed company mid-sample; filling a ResultCard's data block. SKIP for the mechanics of ONE as-of join (safe_asof, in market-data-engineering); for reconciling two price series from two vendors (reconcile_sources, in market-data-sourcing); for auditing a finished backtest (research-integrity-guards); and for resolving a ticker to a permanent id (security-master-and-symbology).
congressional-trading-disclosures
Build a congressional-trading signal on the disclosure date instead of the transaction date, and price what the amount brackets cost you. TRIGGER - congressional trading, congress stock trades, STOCK Act, Periodic Transaction Report, PTR, House Clerk financial disclosure, disclosures-clerk.house.gov, efdsearch.senate.gov, Senate eFD, senator trade trackers, "copy the trades Congress makes", "backtest congressional trades", "what did Congress buy", amount ranges, "$1,001 - $15,000", "how do I weight a trade disclosed as a bracket", the 45-day filing deadline, late PTR filings, scraping the House or Senate disclosure sites. SKIP for corporate insiders and Form 4 transaction codes (insider-form-4), for fund holdings and 13F staleness (institutional-13f), for social posts and influencer panels (social-and-influencer-feeds), for EDGAR endpoints and XBRL point-in-time fundamentals (fundamental-and-macro-data), and for the general availability rule and the five-gate audit (research-integrity-guards).
finding-and-searching-data
Find the identifier before you fetch it, and know which free sources can actually search. TRIGGER - "what is the FRED series id for X", "which ticker is this company", "find the CIK for", searching for a series, ticker, symbol, contract, crypto pair or filing by name or phrase; fred/series/search, company_tickers.json, EDGAR full-text search, efts.sec.gov; a guessed identifier returning an empty frame; retrieving filings by form and date range; when a document became PUBLIC versus when it was filed; earnings-call transcript timestamps. Load BEFORE guessing an identifier - three of the four free searches are current snapshots and cannot answer as of a past date. SKIP for resolving or dating an identifier you already hold (security-master-and-symbology), for choosing a price vendor (market-data-sourcing), for the point-in-time join once the data is yours (fundamental-and-macro-data), for auditing a finished result (research-integrity-guards), and for crypto venue mechanics (crypto-data-and-execution).
insider-form-4
Filter Form 4 to open-market purchases by transaction code, then key the signal to the first session that can trade the acceptance timestamp. TRIGGER - Form 4, Forms 3/4/5, Section 16, insider buying, insider selling, corporate insider trades, "officer bought shares", transaction codes P S A M F G X C, code P vs code A, 10b5-1 plan, EDGAR ownership XML, ownershipDocument, transactionCode, two business days, "insider trading signal", "should I follow insider buys", openinsider-style screens, "why does my insider backtest look amazing", stripping option exercises and tax withholding out of insider data. SKIP for congressional trades and the STOCK Act (congressional-trading-disclosures), for fund holdings and 13F staleness (institutional-13f), for social posts and influencer panels (social-and-influencer-feeds), for EDGAR endpoints, edgartools and XBRL fundamentals (fundamental-and-macro-data), and for the general availability rule (research-integrity-guards).
institutional-13f
Clone or study 13F holdings without the quarter-end look-ahead, and report the full age distribution of the positions instead of the 45-day deadline. TRIGGER - 13F, Form 13F, 13F-HR, institutional holdings, "what does Berkshire own", hedge fund holdings, whale watching, cloning a manager's portfolio, guru portfolios, 13F information table XML, infoTable, $100 million threshold, 45 days after quarter end, confidential treatment request, "positions omitted from the 13F", "backtest a 13F clone", quarterly holdings turnover, WhaleWisdom-style data, reconstructing a fund's book from filings. SKIP for corporate insiders and Form 4 codes (insider-form-4), for congressional trades and the STOCK Act (congressional-trading-disclosures), for social posts and influencer panels (social-and-influencer-feeds), for EDGAR endpoints and XBRL fundamentals (fundamental-and-macro-data), and for the general availability rule (research-integrity-guards).
signal-construction
Compute technical indicators and engineered features without leaking the future. TRIGGER - RSI, MACD, moving average, Bollinger, ATR, ADX, Ichimoku, PSAR, stochastic or any named technical indicator; TA-Lib, pandas-ta, pandas-ta-classic, ta, talipp, finta; choosing an indicator library or reconciling two that disagree; "does this indicator repaint"; warm-up, unstable period, or an indicator differing between backtest and live; zigzag, fractals, swing highs. SKIP for judging whether a finished signal predicts returns (factor-and-timeseries-research) and for the backtest that consumes it (backtesting-engines).
us-market-rules
US trading rules that decide whether a strategy is executable at all - short-sale restrictions, margin, settlement, day-trading limits, wash sales, and what a data licence lets you keep. TRIGGER - can I short this, locate, hard to borrow, borrow fee, short interest, Reg SHO, uptick rule, SSR, short sale restricted; PDT, pattern day trader, day trade limit; Reg T, initial or maintenance margin, margin call, buying power, leverage limit; T+1, settlement, cash account; wash sale, tax-loss harvesting, after-tax returns; "can I redistribute this data"; may_cache, may_redistribute, market data licence; presenting or publishing backtested performance, Marketing Rule. Two of the most-cited rules moved in 2024-2026, so a training-prior answer is usually stale. US ONLY - SKIP for short-selling bans or calendars in Asia (asia-pacific-markets), for A-share T+1 and price limits (china-trading-stack) and for the order-safety mechanics of actually sending an order (broker-execution-apis).
rl-and-ml-trading
Reinforcement learning and deep learning for trading: what installs, and what the evidence says. TRIGGER - reinforcement learning, RL agent, trading gym, custom trading environment, reward function, PPO or DQN for trading; FinRL, FinRL-Meta, ElegantRL, stable-baselines3, gymnasium, gym, gym-anytrading, TensorTrade; LSTM, GRU, TCN or Transformer for return prediction; "does deep learning beat a linear model"; FinRL or gym failing to import with ModuleNotFoundError. FinRL's published wheel declares no dependencies and cannot import, and linear models beat deep ones in the published return-prediction comparisons. SKIP for LLM agents specifically (llm-finance-agents).
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.