finlab

SKILL.md

FinLab Quantitative Trading Package

Prerequisites

Before running any FinLab code, verify these in order:

uv is installed (Python package manager):

uv --version

If uv is not installed, tell the user to install it.

After installing, ensure uv is on PATH:

source $HOME/.local/bin/env 2>/dev/null  # Add uv to current shell

FinLab is installed via uv (requires >= 2.0.0):

uv python install 3.12  # Ensure Python is available (skip if already installed)
uv pip install --system "finlab>=2.0.0" 2>/dev/null || uv pip install "finlab>=2.0.0"

Or use uv run for zero-setup execution (recommended for one-off scripts):

uv run --with "finlab" python3 script.py

uv run --with auto-creates a temporary environment with dependencies — no venv management needed.

API Token is set (required - finlab will fail without it):

If no token, use finlab's built-in login (available in >= 1.5.9, improved Firebase flow in v1.5.11):

import finlab
finlab.login()  # Opens browser for Google OAuth, saves token automatically

This handles the full OAuth flow (browser login, token retrieval, .env storage) automatically.

Language

Respond in the user's language. If user writes in Chinese, respond in Chinese. If in English, respond in English.

Market Support

FinLab supports TW (default), US, KR, JP, and HK markets. The rest of this file plus dataframe-reference.md, backtesting-reference.md, best-practices.md, factor-analysis-reference.md, and machine-learning-reference.md are market-agnostic — the APIs behave the same across markets.

For US-market work — whether single-name equities (data.set_market('us')) or ETFs/funds (data.set_market('us_fund')) — read us-market.md first. Queries that should trigger it include: US equity, S&P 500, NASDAQ 100, 美股, SPY / QQQ, sector SPDRs, leveraged / inverse ETFs, ETF rotation, us_price:*, us_fund_price:*, data.us_universe(...), or us_income_statement:* / us_cash_flow:* / us_balance_sheet:*. It documents:

Which US data tables are safe for backtesting versus current-snapshot-only (analyst consensus, ratios, DCF are live-only — do not use them historically)

Filing-date-aligned quarterly fundamentals (key_date == filing_date) — no .shift() workaround needed

Report API names on US (creturn / daily_creturn / get_stats(); no get_equity())

US backtest defaults for both markets: USMarket (fee_ratio=0, tax_ratio=0, trade_at_price='close') and USFundMarket for ETF/fund backtests

How data.set_market(...) is the session-scope switch (there is no market= kwarg on data.get())

Dollar-volume-top-N universe construction (works back to 2016), S&P 500 / NASDAQ 100 membership via data.us_universe(index='S&P 500' | 'NASDAQ 100') with its 2022-11 history-start caveat, quality gates, and sector-exclusion rationale

Lookahead-bias checklist specific to US data (rolling-window universe filters, survivorship avoidance)

ETF / sector-rotation backtesting via USFundMarket and us_fund_price:*

Other-market queries can skip that file.

API Token Tiers & Usage

Token Tiers

Tier
Daily Limit
Token Pattern

Free
500 MB
ends with #free

VIP
5000 MB
no suffix

Usage Reset

Resets daily at 8:00 AM UTC+8

When limit exceeded, user must wait for reset or upgrade to VIP

Quick Start Example

from finlab import data
from finlab.backtest import sim

# 1. Fetch data
close = data.get("price:收盤價")
vol = data.get("price:成交股數")
pb = data.get("price_earning_ratio:股價淨值比")

# 2. Create conditions
cond1 = close.rise(10)  # Rising last 10 days
cond2 = vol.average(20) > 1000*1000  # High liquidity
cond3 = pb.rank(axis=1, pct=True) < 0.3  # Low P/B ratio

# 3. Combine conditions and select stocks
position = cond1 & cond2 & cond3
position = pb[position].is_smallest(10)  # Top 10 lowest P/B

# 4. Backtest
report = sim(position, resample="M", upload=False)

# 5. Print metrics - Two equivalent ways:

# Option A: Using metrics object
print(report.metrics.annual_return())
print(report.metrics.sharpe_ratio())
print(report.metrics.max_drawdown())

# Option B: Using get_stats() dictionary (different key names!)
stats = report.get_stats()
print(f"CAGR: {stats['cagr']:.2%}")
print(f"Sharpe: {stats['monthly_sharpe']:.2f}")
print(f"MDD: {stats['max_drawdown']:.2%}")

report

Core Workflow: 5-Step Strategy Development

Step 1: Fetch Data

Use data.get("<TABLE>:<COLUMN>") to retrieve data:

from finlab import data

# Price data
close = data.get("price:收盤價")
volume = data.get("price:成交股數")

# Financial statements
roe = data.get("fundamental_features:ROE稅後")
revenue = data.get("monthly_revenue:當月營收")

# Valuation
pe = data.get("price_earning_ratio:本益比")
pb = data.get("price_earning_ratio:股價淨值比")

# Institutional trading
foreign_buy = data.get("institutional_investors_trading_summary:外陸資買賣超股數(不含外資自營商)")

# Technical indicators
rsi = data.indicator("RSI", timeperiod=14)
macd, macd_signal, macd_hist = data.indicator("MACD", fastperiod=12, slowperiod=26, signalperiod=9)

Filter by market/category using data.universe():

# Limit to specific industry
with data.universe(market='TSE_OTC', category=['水泥工業']):
    price = data.get('price:收盤價')

# Set globally
data.set_universe(market='TSE_OTC', category='半導體')

Use data.search('keyword', market='<market>') to discover available datasets. Supported markets: tw, us, kr, jp, hk. Use keywords in the dataset's native language (e.g. data.search('營收', market='tw'), data.search('revenue', market='us')).

Step 2: Create Factors & Conditions

Use FinLabDataFrame methods to create boolean conditions:

# Trend
rising = close.rise(10)  # Rising vs 10 days ago
sustained_rise = rising.sustain(3)  # Rising for 3 consecutive days

# Moving averages
sma60 = close.average(60)
above_sma = close > sma60

# Ranking
top_market_value = data.get('etl:market_value').is_largest(50)
low_pe = pe.rank(axis=1, pct=True) < 0.2  # Bottom 20% by P/E

# Industry ranking
industry_top = roe.industry_rank() > 0.8  # Top 20% within industry

See dataframe-reference.md for all FinLabDataFrame methods.

Step 3: Construct Position DataFrame

Combine conditions with & (AND), | (OR), ~ (NOT):

# Simple position: hold stocks meeting all conditions
position = cond1 & cond2 & cond3

# Limit number of stocks
position = factor[condition].is_smallest(10)  # Hold top 10

# Entry/exit signals with hold_until
entries = close > close.average(20)
exits = close < close.average(60)
position = entries.hold_until(exits, nstocks_limit=10, rank=-pb)

Important: Position DataFrame should have:

Index: DatetimeIndex (dates)

Columns: Stock IDs (e.g., '2330', '1101')

Values: Boolean (True = hold) or numeric (position size)

Step 4: Backtest

from finlab.backtest import sim

# Basic backtest
report = sim(position, resample="M")

# With risk management
report = sim(
    position,
    resample="M",
    stop_loss=0.08,
    take_profit=0.15,
    trail_stop=0.05,
    position_limit=1/3,
    fee_ratio=1.425/1000/3,
    tax_ratio=3/1000,
    trade_at_price='open',
    upload=False
)

# Extract metrics - Two ways:
# Option A: Using metrics object
print(f"Annual Return: {report.metrics.annual_return():.2%}")
print(f"Sharpe Ratio: {report.metrics.sharpe_ratio():.2f}")
print(f"Max Drawdown: {report.metrics.max_drawdown():.2%}")

# Option B: Using get_stats() dictionary (note: different key names!)
stats = report.get_stats()
print(f"CAGR: {stats['cagr']:.2%}")           # 'cagr' not 'annual_return'
print(f"Sharpe: {stats['monthly_sharpe']:.2f}") # 'monthly_sharpe' not 'sharpe_ratio'
print(f"MDD: {stats['max_drawdown']:.2%}")     # same name

See backtesting-reference.md for complete sim() API.

Step 5: Execute Orders (Optional)

Convert backtest results to live trading:

from finlab.online.order_executor import Position, OrderExecutor
from finlab.online.sinopac_account import SinopacAccount

# 1. Convert report to position
position = Position.from_report(report, fund=1000000)

# 2. Connect broker account
acc = SinopacAccount()

# 3. Create executor and preview orders
executor = OrderExecutor(position, account=acc)
executor.create_orders(view_only=True)  # Preview first

# 4. Execute orders (when ready)
executor.create_orders()

See trading-reference.md for complete broker setup and OrderExecutor API.

Reference Files

File
Content

backtesting-reference.md
sim() 參數、stop-loss、rebalancing

trading-reference.md
券商設定、OrderExecutor、Position

factor-examples.md
60+ 策略範例

dataframe-reference.md
FinLabDataFrame 方法

factor-analysis-reference.md
IC、Shapley、因子分析

best-practices.md
常見錯誤、lookahead bias

machine-learning-reference.md
ML 特徵工程

us-market.md
US market specifics: data map, quarterly alignment, defaults, universe construction

What's New (since v1.5.8)

Short version pointers for features added in recent releases. Each reference file tags the exact API with (vX.Y.Z).

v2.0.0 (2026-04-04) — major release

finlab.exceptions: structured error hierarchy (FinlabError, DataError, BacktestError, ...) — see backtesting-reference.md

data.get(lazy=True) / data.gets(..., lazy=True): batch fetch + deferred compute; data.override() / DataContext for scoped global state

df.cs / df.sector / df.weight accessors; rolling().std/var/skew/kurt/median — see dataframe-reference.md

PositionStreamMixin for realtime position streaming — see trading-reference.md

from finlab import FinlabDataFrame top-level export

backtest.sim() refactored into 5 testable stages; eval() removed from optimize.combinations

v1.5.13 (2026-03-22)

universe(index=...) / us_universe(index=...): filter US stocks by S&P 500 / NASDAQ 100

New market code TW_CB (TW convertible bonds)

v1.5.11 (2026-03-11)

data.get_role() / data.is_vip(): query user quota tier

Report migration to canonical Firestore flow (transparent to users)

v1.5.9

finlab.schemas: typed PositionEntry, OrderEntry, PortfolioData contracts

OrderExecutor.generate_orders(as_entries, quantity_type) and generate_order_entries()

PortfolioSyncManager.get_data_typed() / set_data_typed()

data.get() 80% quota usage warning

sim() uses market-specific default fee_ratio / tax_ratio (no longer hardcoded TW values)

v1.5.8 (baseline)

verify_strategy(): automated lookahead-bias detector

report.to_terminal(): ASCII report for non-Jupyter runs

Overall strategy execution 3.4x faster

Prevent Lookahead Bias

Critical: Avoid using future data to make past decisions:

# ✅ GOOD: Use shift(1) to get previous value
prev_close = close.shift(1)

# ❌ BAD: Don't use iloc[-2] (can cause lookahead)
# prev_close = close.iloc[-2]  # WRONG

# ✅ GOOD: Leave index as-is even with strings like "2025Q1"
# FinLabDataFrame aligns by shape automatically

# ❌ BAD: Don't manually assign to df.index
# df.index = new_index  # FORBIDDEN

See best-practices.md for more anti-patterns.

Performance Defaults

Pass lazy=True by default; drop to eager pandas only when debugging. data.get(..., lazy=True) and data.gets(..., lazy=True) (v2.0.0) return lazy FinlabDataFrames that defer the compute graph until a terminal call materializes it — chained ops avoid redundant passes (single-CPU). Omit lazy=True when you need to print/inspect intermediate values interactively.

# ✅ Default: fetch lazy directly
price, volume, pe = data.gets(
    'price:收盤價', 'price:成交股數', 'price_earning_ratio:本益比',
    lazy=True,
)

# ✅ Debug: eager pandas for row-level inspection
close = data.get('price:收盤價')
print(close.loc['2024-01-15', '2330'])

Feedback

Direct users to open an issue on GitHub: https://github.com/koreal6803/finlab-ai/issues

Notes

Some data columns use Chinese names — this is expected, use them as-is in data.get() calls

Data frequency varies: daily (price), monthly (revenue), quarterly (financial statements)

Always use sim(..., upload=False) for experiments, upload=True only for final production strategies