2.0.0
FinLab 2.0.0 turns strategy research into a faster feedback loop: rebalanced strategies can compute only the dates they need, datasets can load together, factor tools are richer, backtest internals are safer, and wheels are ready for desktop, Linux, Windows, macOS, and Pyodide.
For monthly rebalancing, the engine does not need to calculate every daily intermediate result. Add `.lazy()` to the source dataframe and keep the rest of the strategy readable.
from finlab import data
from finlab.backtest import sim
close = data.get("price:收盤價")
position = (
close.rank(axis=1, pct=True)
[close.average(20) > close.average(60)]
.is_largest(20)
)
report = sim(position, resample="M", upload=False)from finlab import data
from finlab.backtest import sim
close = data.get("price:收盤價").lazy()
position = (
close.rank(axis=1, pct=True)
[close.average(20) > close.average(60)]
.is_largest(20)
)
report = sim(position, resample="M", upload=False)Test-style synthetic benchmark: 2,500 business days x 1,200 symbols, rolling 20/60 averages, percentile rank, boolean mask, top 20, monthly rebalance. Median of 5 runs on this local checkout.
3.5x faster. Results were checked to match; only the dataframe container type is different.
Rebalanced strategies do not need every intermediate daily row. The same expression is evaluated at target rebalance dates, while preserving familiar operations: rolling, shift, rank, masks, arithmetic, and top-N selection.
2.0.0 adds a batch fetch path with one metadata request, parallel dataset materialization, and pip-style progress bars in terminal or HTML progress in notebooks.
close = data.get("price:收盤價")
volume = data.get("price:成交股數")
revenue = data.get("monthly_revenue:當月營收")close, volume, revenue = data.gets(
"price:收盤價",
"price:成交股數",
"monthly_revenue:當月營收",
)Measured with the integration-test datasets `security_categories` and `security_industry_themes`, `force_download=True`, isolated in-memory storage, median of 3 runs.
2.0x faster on this measured path. Larger independent downloads can benefit more from fetching together.
The dataframe API now covers more of the daily quant workflow: cross-sectional transforms, sector-aware transforms, and portfolio weight construction live next to the data they operate on.
`rank()`, `winsorize()`, `bucket()`, `zscore()`, and `demean()` for every date across the stock universe.
Example: turn a raw factor into a clean cross-sectional score before selecting names.
score = (
pe.cs.winsorize(0.02, 0.98)
.cs.zscore()
.cs.rank()
)
signal = score.is_smallest(20)Aggregate and transform within sectors using `mean`, `std`, `median`, `sum`, `min`, `max`, `count`, and sector ranks.
Example: rank profitability within each industry, then keep only sector leaders.
quality = (
roe.sector.winsorize(0.05, 0.95)
.sector.zscore()
.sector.rank()
)
signal = quality > 0.8Cap industries, clip by liquidity, inverse-vol weight, risk parity, target volatility, turnover limits, and drawdown control.
Example: compose sizing, risk, liquidity, and turnover controls in one readable chain.
weights = (
signal.is_largest(20)
.weight.inverse_volatility(60)
.weight.cap_industry(0.30)
.weight.clip_by_volume(1e8)
.weight.limit_turnover(0.20)
)2.0.0 puts production-facing pieces behind clearer contracts: structured exceptions, safer parsers, a staged backtest pipeline, isolated data context, stronger lint rules, and CI coverage gates.
The package was validated locally, merged to `main`, merged to `stable`, and compiled through GitHub Actions. Wheels were downloaded into `wheelhouse/` for release handling.
Ruff passed on edited files and pre-commit hooks. Integration path: `uv run pytest -m "not unit and not real_order"`.
`wheelhouse/` contains 26 wheels for CPython 3.9-3.14 across macOS universal2, Linux x86_64, Linux aarch64, Windows amd64, plus Pyodide wasm32 wheels.
Keep your strategy readable, fetch the data together, and let the engine do less unnecessary work.