High-performance quantitative finance algorithm library, implemented in Rust with Python bindings (PyO3).

Provides efficient rolling-window calculations commonly used in factor-based quantitative trading.

Benchmarked on Alpha 101, 4000 stocks x 261 trading days (1,044,000 data points per factor):

See COMPARISON.md for per-factor timing and correctness analysis.

pip install py-alpha-lib

Control computation behavior via alpha.set_ctx():

• groups — Number of securities in the data array. Each group is processed independently and in parallel. Required for cross-sectional operations like RANK.

• start — Starting index for calculation (default: 0).

• end — Ending index for calculation (default: len(data)). end can be used when you want to calculate only a part of the data. for example, when back test iteratively.

• flags — Bitwise flags:

  • FLAG_SKIP_NAN (1): Skip NaN values in rolling windows.
  • FLAG_STRICTLY_CYCLE (2): Return NaN until window is full (matches pandas rolling() default).
  • Combine with |: flags=FLAG_SKIP_NAN | FLAG_STRICTLY_CYCLE

  import alpha
  import numpy as np
  
  data = np.array([1, 2, 3, 4, 5, 6, 7, 8, 9, 10], dtype=np.float64)
  
  # 3-period moving average (partial results during warm-up)
  result = alpha.MA(data, 3)
  # [1.  1.5 2.  3.  4.  5.  6.  7.  8.  9.]
  
  # Strict mode: NaN until window is full
  alpha.set_ctx(flags=alpha.FLAG_STRICTLY_CYCLE)
  result = alpha.MA(data, 3)
  # [nan nan 2.  3.  4.  5.  6.  7.  8.  9.]
  
  # Skip NaN values
  alpha.set_ctx(flags=alpha.FLAG_SKIP_NAN)
  data_nan = np.array([1, 2, np.nan, 4, 5, 6, 7, 8, 9, 10], dtype=np.float64)
  result = alpha.MA(data_nan, 3)
  #[1.    1.5     nan 2.333 3.667 5.    6.    7.    8.    9.   ]

import alpha
from alpha.context import ExecContext

# ExecContext auto-infers groups from securityid/tradetime columns
# and calls alpha.set_ctx(groups=...) automatically
data = pl.read_csv("data.csv").sort(["securityid", "tradetime"])
ctx = ExecContext(data)

# Call operators directly on numpy arrays
close = data["close"].to_numpy()
ma20 = alpha.MA(close, 20)
rank = alpha.RANK(close)       # cross-sectional rank (groups auto-configured)
corr = alpha.CORR(close, data["vol"].to_numpy().astype(float), 10)

Data layout: flat 1D array [stock1_day1, stock1_day2, ..., stockN_dayM], sorted by security then time. The groups parameter tells the library where each stock's data begins.

Convert factor expressions to Python code, then run:

python -m alpha.lang examples/wq101/alpha101.txt

# 3. Use generated code
from alpha.context import ExecContext
from factors import alpha_001

data = pl.read_csv("data.csv").sort(["securityid", "tradetime"])
ctx = ExecContext(data)  # auto-infers groups
result = alpha_001(ctx)

You can convert factor expressions to Python code using the lang module. For example:

python -m alpha.lang examples/wq101/alpha101.txt

This will read the factor expressions from examples/wq101/alpha101.txt and generate corresponding Python code using alpha-lib functions.

After generating the code, you may need to adjust the code

• Fix type conversions between float and bool.
• Add context settings if needed.

Implementation of 190/191 factors from the GTJA (国泰君安) Alpha 191 factor set in examples/gtja191/:

python -m examples.gtja191.al 143     # run specific factor
python -m examples.gtja191.al          # run all factors

Full implementation of 101 Formulaic Alphas in examples/wq101/:

• al/ — alpha-lib implementation (Rust backend)
• pd_/ — pandas reference (DolphinDB port)
• pl_/ — polars_ta reference

examples/wq101/main.py --with-al 1 2 3 4 # Run specific factors
examples/wq101/main.py --with-al -s 1 -e 102 # Run all factors
examples/wq101/main.py --with-pd --with-al -s 1 -e 15 # Compare with pandas

Benchmark scripts in benchmarks/.

Naming Rules:

• Function starts with CC_ means it is a cross-commodity/cross-security/cross-group operation.
• Function without prefix means it is a rolling window operation.

Full function signatures: python/alpha/algo.md

This project is Python based originally, but for some case Python is not available, for example

• You build a stock web app, you want draw an indicator on the chart.
• Your application will be deployed to customer's computer, it is not easy to install Python on the customer's computer. But lua can be embeded in your app.

So we add some other language bindings for this project.

• lua-binding: Provide a UserData named NumArray to do numpy.ndarray like operations, and register all ta funtions to lua. So you can use it in your lua application.
• js-binding: Provide a MLang runtime to execute mlang code on the browser. And use canvas to draw the chart.

Requirements:

• Rust (latest stable)
• Python 3.11+
• maturin

# Build and install in development mode
maturin develop --release

# Run tests
cargo test

When adding new algorithms with LLM assistance, provide the function list as context. Use the skill add_algo.md for guided implementation.

This project is a co-created by Gemini (through Antigravity) and Claude (from tic-top).