alpha-features-core 0.1.7

Fast C++ implementation of Alpha191 financial factors

alpha-features-core

Fast C++ implementation of the Alpha191 quantitative finance factor library, with a pure-Python/pandas fallback.

Installation

pip install alpha-features-core

Building from source requires CMake ≥ 3.15 and a C++17 compiler (MSVC on Windows, GCC/Clang on Linux/macOS).

Quick start

import pandas as pd
from alpha_features_core.bridge import compute_alphas_cpp   # C++ backend
from alpha_features_core.alpha191 import Alphas191          # pure-Python backend

# df is a long-format OHLCV DataFrame with columns:
#   ticker, date, open, high, low, close, volume, amount, past_return

# --- C++ backend (fast) ---
cpp_result = compute_alphas_cpp(df)          # all 191 factors
cpp_result = compute_alphas_cpp(df, nums=[1, 5, 10])  # subset

# --- Pure-Python backend ---
py_result = Alphas191(df).calculate_all_alphas(return_long=True)

Both functions return a long-format pd.DataFrame with columns [ticker, date, alpha001, alpha002, ...].

Performance

The C++ backend compiles all factors in a single pass over the data and significantly outperforms the pandas/numba implementation at scale.

from alpha_features_core.bridge import compute_alphas_cpp
from alpha_features_core.alpha191 import Alphas191
import numpy as np, pandas as pd, time, matplotlib.pyplot as plt

EXPERIMENTS = [
    (5,  40), (10, 50), (20, 50), (30, 70), (50, 80),
    (80, 100), (100, 120), (150, 150), (200, 180),
    (250, 200), (300, 250), (350, 300), (500, 500),
]
ALL_NUMS = list(range(1, 192))
times_cpp, times_py, n_rows = [], [], []

for T, D in EXPERIMENTS:
    rng = np.random.default_rng(42)
    tickers = [f"T{i:04d}" for i in range(T)]
    dates   = pd.date_range("2020-01-01", periods=D, freq="B")
    idx     = pd.MultiIndex.from_product([tickers, dates], names=["ticker", "date"])
    n       = len(idx)
    close   = 100 + np.cumsum(rng.normal(0, 0.5, n))
    df = pd.DataFrame({
        "close":       close,
        "open":        close + rng.normal(0, 0.2, n),
        "high":        close + np.abs(rng.normal(0, 0.5, n)),
        "low":         close - np.abs(rng.normal(0, 0.5, n)),
        "volume":      np.abs(rng.normal(1e6, 1e5, n)),
        "past_return": rng.normal(0, 0.01, n),
    }, index=idx).reset_index()
    df["amount"] = df["close"] * df["volume"]
    df = df.sort_values(["ticker", "date"]).reset_index(drop=True)

    t0 = time.perf_counter()
    compute_alphas_cpp(df, nums=ALL_NUMS, verbose=False)
    times_cpp.append(time.perf_counter() - t0)

    t0 = time.perf_counter()
    Alphas191(df).calculate_all_alphas(return_long=True)
    times_py.append(time.perf_counter() - t0)

    n_rows.append(n)
    print(f"{T}×{D} = {n:>7,} rows  C++: {times_cpp[-1]:.2f}s  "
          f"Python: {times_py[-1]:.2f}s  speedup: {times_py[-1]/times_cpp[-1]:.1f}x")

fig, ax = plt.subplots(figsize=(10, 6))
ax.plot(n_rows, times_cpp, "o-", color="steelblue",  lw=2, ms=6, label="C++ backend")
ax.plot(n_rows, times_py,  "s-", color="darkorange", lw=2, ms=6, label="Pure Python (numba)")
ax.set_xlabel("Number of rows (tickers × dates)")
ax.set_ylabel("Execution time (s)")
ax.set_title("Alpha191: C++ vs Pure Python — scaling benchmark")
ax.legend()
plt.tight_layout()
plt.savefig("benchmark.png", dpi=150)

Numerical differences between backends

The C++ and Python backends produce identical results for the vast majority of factors. Small numerical differences exist for factors that combine Corr and Rank operations (e.g. alpha001, alpha016, alpha090).

The root cause is a floating-point edge case in pandas' rolling().corr(): when one input series is constant within the rolling window, pandas internally produces ±inf or NaN depending on accumulated floating-point rounding errors that vary per window. The C++ backend applies a deterministic kEps threshold and cannot replicate this non-deterministic behaviour exactly.

In practice the affected factors differ by at most 1–2 rank positions out of 100 tickers per date, which has negligible impact on factor signal quality.

The following factors are affected: alpha001, alpha005, alpha016, alpha036, alpha054, alpha056, alpha061, alpha064, alpha073, alpha077, alpha083, alpha090, alpha091, alpha092, alpha099, alpha101, alpha113, alpha115, alpha119, alpha121, alpha123, alpha130, alpha131, alpha138, alpha141, alpha148, alpha170, alpha176, alpha179, alpha191.

All remaining 160+ factors match the Python backend within floating-point tolerance (atol=1e-5).

Input DataFrame format

Column	Type	Description
ticker	str	Asset identifier
date	datetime	Trading date
open	float	Opening price
high	float	High price
low	float	Low price
close	float	Closing price
volume	float	Trading volume
amount	float	Traded amount (close × volume)
past_return	float	Previous period return

The DataFrame must be sorted by [ticker, date]. Column names can be customised via keyword arguments to compute_alphas_cpp().

Performance boost in using C++ vs pure Python:

License

MIT