AlphaInspect

https://github.com/wukan1986/alphainspect
Admin
AlphaInspect is a factor performance visualization and analysis tool for quantitative trading that...
Tokens:11,371
Snippets:65
Trust Score:9.5
License:MIT
Update:6 months ago
Show doc for...
Tokens:
Raw
### Install AlphaInspect

Source: https://github.com/wukan1986/alphainspect/blob/main/README.md

Install AlphaInspect using pip. Use the official PyPI source or a domestic mirror.

```commandline
pip install -i https://pypi.org/simple --upgrade alphainspect  # 官方源
```

```commandline
pip install -i https://pypi.tuna.tsinghua.edu.cn/simple --upgrade alphainspect  # 国内镜像源
```

--------------------------------

### Clone AlphaInspect Repository

Source: https://github.com/wukan1986/alphainspect/blob/main/README.md

Clone the AlphaInspect repository from GitHub and install it in editable mode.

```commandline
git --clone https://github.com/wukan1986/alphainspect.git
cd alphainspect
pip install -e .
```

--------------------------------

### Data Preparation Workflow

Source: https://context7.com/wukan1986/alphainspect/llms.txt

Prepares raw OHLC data into the required AlphaInspect format, including forward returns and factor calculations.

```python
import numpy as np
import pandas as pd
import polars as pl

# 生成模拟数据
_N = 250 * 10  # 10年日数据
_K = 500       # 500只股票

asset = [f's_{i:04d}' for i in range(_K)]
date = pd.date_range('2018-1-1', periods=_N)

np.random.seed(42)

# 创建OHLC数据
df = pd.DataFrame({
    'OPEN': np.cumprod(1 + np.random.uniform(-0.01, 0.01, size=(_N, _K)), axis=0).reshape(-1),
    'HIGH': np.cumprod(1 + np.random.uniform(-0.01, 0.01, size=(_N, _K)), axis=0).reshape(-1),
    'LOW': np.cumprod(1 + np.random.uniform(-0.01, 0.01, size=(_N, _K)), axis=0).reshape(-1),
    'CLOSE': np.cumprod(1 + np.random.uniform(-0.01, 0.01, size=(_N, _K)), axis=0).reshape(-1),
}, index=pd.MultiIndex.from_product([date, asset], names=['date', 'asset'])).reset_index()

df = pl.from_pandas(df)

# 计算远期收益率（需要shift到当前位置）
# RETURN_OO_01: T+1开盘买入，T+2开盘卖出
# RETURN_CC_01: T+0收盘买入，T+1收盘卖出
df = df.with_columns([
    (pl.col('OPEN').shift(-2) / pl.col('OPEN').shift(-1) - 1).over('asset').alias('RETURN_OO_01'),
    (pl.col('CLOSE').shift(-1) / pl.col('CLOSE') - 1).over('asset').alias('RETURN_CC_01'),
])

# 计算因子
df = df.with_columns([
    pl.col('CLOSE').rolling_mean(10).over('asset').alias('SMA_010'),
    pl.col('CLOSE').rolling_std(10).over('asset').alias('STD_010'),
])

# 保存数据
df.write_parquet('data/data.parquet')
```

--------------------------------

### Create Portfolio Sheet with Asset-Specific Returns

Source: https://github.com/wukan1986/alphainspect/blob/main/deprecated/deprecated.txt

Generates plots for individual asset cumulative returns and an overall portfolio equal-weighted return. Assumes weights are provided externally and capital is isolated per asset.

```python
def create_portfolio2_sheet(df: pl.DataFrame,
                            fwd_ret_1: str,
                            *, 
                            axvlines=()) -> None:
    """分资产收益。权重由外部指定，资金是隔离"""
    # 各资产收益，如果资产数量过多，图会比较卡顿
    df_cum_ret = calc_cum_return_weights(df, fwd_ret_1, 1)

    fig, axes = plt.subplots(2, 1, figsize=(12, 9), squeeze=False)
    axes = axes.flatten()
    # 分资产收益
    plot_quantile_portfolio(df_cum_ret, fwd_ret_1, axvlines=axvlines, ax=axes[0])

    # 资产平均收益，相当于等权
    s = df_cum_ret.mean(axis=1)
    s.name = 'portfolio'
    plot_quantile_portfolio(s, fwd_ret_1, axvlines=axvlines, ax=axes[1])
    fig.tight_layout()
```

--------------------------------

### LightGBM Model Visualization

Source: https://context7.com/wukan1986/alphainspect/llms.txt

Visualizes feature importance and training metrics for LightGBM models, and exports decision trees to HTML.

```python
import matplotlib.pyplot as plt
from alphainspect.dtree import plot_importance_box, plot_metric_errorbar, tree_to_html

# 假设已训练好多个LightGBM模型
# models = [model1, model2, model3, ...]

fig, axes = plt.subplots(1, 2, figsize=(14, 6))

# 多模型特征重要性箱线图
plot_importance_box(
    models,
    plot_top_k=20,                        # 显示前20个特征
    importance_type='gain',               # 或 'split'
    ax=axes[0]
)

# 训练过程评估指标图
plot_metric_errorbar(
    models,
    metric='auc',                         # 评估指标名
    ax=axes[1]
)

plt.tight_layout()
plt.show()

# 将决策树导出为HTML
html_content = tree_to_html(
    models[-1],
    tree_index=0,
    show_info=['split_gain', 'leaf_count', 'data_percentage'],
    precision=3,
    orientation='horizontal'
)

with open('tree.html', 'w', encoding='utf-8') as f:
    f.write(html_content)
```

--------------------------------

### Generate Portfolio Performance Sheet

Source: https://context7.com/wukan1986/alphainspect/llms.txt

Creates a visualization of layered cumulative returns, including net value curves and monthly return heatmaps.

```python
import polars as pl
import matplotlib.pyplot as plt
from alphainspect.portfolio import create_portfolio_sheet
from alphainspect.utils import with_factor_quantile
from alphainspect import _DATE_

df = pl.read_parquet('data/data.parquet')
df = with_factor_quantile(df, 'SMA_010', quantiles=9, by=[_DATE_], factor_quantile='_fq_1')

# 生成分层绩效图表
create_portfolio_sheet(
    df,
    fwd_ret_1='RETURN_OO_05',
    factor_quantile='_fq_1',
    axvlines=('2020-01-01', '2024-01-01')
)

# 生成的图表：
# - 上半部分：分层净值曲线（含多空对冲）
# - 下半部分：最低层和最高层的月度收益热力图

plt.show()
```

--------------------------------

### Multi-Factor IC Comparison

Source: https://context7.com/wukan1986/alphainspect/llms.txt

Analyze multiple factors across different holding periods using correlation matrices and heatmaps.

```python
import polars as pl
import matplotlib.pyplot as plt
from alphainspect.ic import create_ic2_sheet

df = pl.read_parquet('data/data.parquet')

# 多因子多收益率IC分析
df_ic = create_ic2_sheet(
    df,
    factors=['SMA_005', 'SMA_010', 'SMA_020'],
    forward_returns=['RETURN_OO_01', 'RETURN_OO_05', 'RETURN_OO_10'],
    axvlines=('2020-01-01',)
)

plt.show()
```

--------------------------------

### Daily Simple Returns Calculation

Source: https://github.com/wukan1986/alphainspect/blob/main/README.md

Formulas for calculating daily simple returns for long and short positions, adjusted for different trading strategies.

```python
FWD_RET_LONG = (卖出平仓价 / 买入开仓价) ** (1 / D) - 1
FWD_RET_SHORT = (2 - 买入平仓价 / 卖出开仓价) ** (1 / D) - 1
```

```python
FWD_RET = (OPEN[-1] / CLOSE) ** (1 / 1) - 1  # 收盘前买入，第二天开盘卖出
```

```python
FWD_RET = (OPEN[-1] / HIGH[1]) ** (1 / 1) - 1  # 突破昨高买入，第二天开盘卖出
```

```python
FWD_RET = (OPEN[-1] / max_(OPEN, CLOSE)[1]) ** (1 / 1) - 1  # 突破昨实体上沿买入，第二天开盘卖出
```

--------------------------------

### Generate 2x2 Quick Analysis Report

Source: https://context7.com/wukan1986/alphainspect/llms.txt

Produces a concise 2x2 grid of charts including IC time series, IC heatmaps, factor histograms, and cumulative returns.

```python
import polars as pl
import matplotlib.pyplot as plt
from alphainspect.reports import create_2x2_sheet
from alphainspect.utils import with_factor_quantile
from alphainspect import _DATE_

df = pl.read_parquet('data/data.parquet')
df = with_factor_quantile(df, 'SMA_010', quantiles=9, by=[_DATE_], factor_quantile='_fq_1')

# 生成2x2分析图表
fig, ic_dict, hist_dict, cum, avg, std = create_2x2_sheet(
    df,
    factor='SMA_010',
    fwd_ret_1='RETURN_OO_05',
    factor_quantile='_fq_1',
    figsize=(12, 9),
    axvlines=('2020-01-01', '2024-01-01')
)

# 返回值包含统计指标
print(f"IC统计: {ic_dict}")
# {'mean': 0.0234, 'zscore': 1.23, 'ratio': 0.05, 't_stat': 3.45, 'p_value': 0.001}

print(f"直方图统计: {hist_dict}")
# {'std': 0.0156, 'skew': 0.12, 'kurt': 3.45, 'mean_': 0.0023}

plt.show()
```

--------------------------------

### Top-K Factor Stratification

Source: https://context7.com/wukan1986/alphainspect/llms.txt

Use with_factor_top_k for scenarios with limited cross-sectional stock counts, outputting three layers: short, hedge, and long.

```python
import polars as pl
from alphainspect.utils import with_factor_top_k
from alphainspect import _DATE_

df = pl.read_parquet('data/data.parquet')

# 使用Top-K分层，选取因子值最高和最低的各20只股票
df = with_factor_top_k(
    df,
    factor='SMA_010',
    top_k=20,                   # 前20和后20
    by=[_DATE_],
    factor_quantile='_fq_topk'
)

# 结果：0=做空组, 1=对冲组, 2=做多组
print(df.group_by('_fq_topk').count())
```

--------------------------------

### Calculate cumulative returns

Source: https://github.com/wukan1986/alphainspect/blob/main/deprecated/utils.txt

Estimates portfolio performance by splitting capital into sub-portfolios. Requires returns and weights as numpy arrays.

```python
def cumulative_returns(returns: np.ndarray, weights: np.ndarray,
                       funds: int = 3, freq: int = 3,
                       benchmark: np.ndarray = None,
                       ret_mean: bool = True,
                       init_cash: float = 1.0,
                       risk_free: float = 1.0,  # 1.0 + 0.025 / 250
                       ) -> np.ndarray:
    """累积收益

    精确计算收益是非常麻烦的事情，比如考虑手续费、滑点、涨跌停无法入场。考虑过多也会导致计算量巨大。
    这里只做估算，用于不同因子之间收益比较基本够用。更精确的计算请使用专用的回测引擎

    需求：因子每天更新，但策略是持仓3天
    1. 每3天取一次因子，并持有3天。即入场时间对净值影响很大。净值波动剧烈
    2. 资金分成3份，每天入场一份。每份隔3天做一次调仓，多份资金不共享。净值波动平滑

    本函数使用的第2种方法，例如：某支股票持仓信息如下
    [0,1,1,1,0,0]
    资金分成三份，每次持有三天，
    [0,0,0,1,1,1] # 第0、3、6...位，fill后两格
    [0,1,1,1,0,0] # 第1、4、7...位，fill后两格
    [0,0,1,1,1,0] # 第2、5、8...位，fill后两格

    Parameters
    ----------
    returns: np.ndarray
        1期简单收益率。自动记在出场位置。
    weights: np.ndarray
        持仓权重。需要将信号移动到出场日期。权重绝对值和
    funds: int
        资金拆成多少份
    freq:int
        再调仓频率
    benchmark: 1d np.ndarray
        基准收益率
    ret_mean: bool
        返回多份资金合成曲线
    init_cash: float
        初始资金
    risk_free: float
        无风险收益率。用在现金列。空仓时，可以给现金提供利息

    Returns
    -------
    np.ndarray

    References
    ----------
    https://github.com/quantopian/alphalens/issues/187

    """
    # 一维修改成二维，代码统一
    if returns.ndim == 1:
        returns = returns.reshape(-1, 1)
    if weights.ndim == 1:
        weights = weights.reshape(-1, 1)

    # 形状
    m, n = weights.shape

    # 现金权重
    weights_cash = 1 - np.round(np.nansum(np.abs(weights), axis=1), 5)
    # TODO 也可以添加两列现金，一列有利息，一列没利息。细节要按策略进行定制
    returns = np.concatenate((np.ones(shape=(m, 1), dtype=returns.dtype), returns), axis=1)
    weights = np.concatenate((np.zeros(shape=(m, 1), dtype=weights.dtype), weights), axis=1)
    # 添加第0列做为现金，用于处理CTA空仓的问题
    weights[:, 0] = weights_cash
    # 可以考虑给现金指定一个固定收益
    returns[:, 0] = risk_free

    # 修正数据中出现的nan
    returns = np.where(returns == returns, returns, 1.0)
    # 权重需要已经分配好，绝对值和为1
    weights = np.where(weights == weights, weights, 0.0)

    # 新形状
    m, n = weights.shape

    #  记录每份资金每期收益率
    out = _sub_portfolio_returns(m, n, weights, returns, funds, freq, init_cash)
    if ret_mean:
        if benchmark is None:
            # 多份净值直接叠加后平均
            return out.mean(axis=1)
        else:
            # 有基准，计算超额收益
            return out.mean(axis=1) - (benchmark + 1).cumprod()
    else:
        return out
```

--------------------------------

### Single Factor IC Visualization

Source: https://context7.com/wukan1986/alphainspect/llms.txt

Generate a comprehensive sheet of IC analysis charts for a single factor, including time-series, histograms, and heatmaps.

```python
import polars as pl
import matplotlib.pyplot as plt
from alphainspect.ic import create_ic1_sheet

df = pl.read_parquet('data/data.parquet')

# 生成单因子多收益率的IC图表
df_ic = create_ic1_sheet(
    df,
    factor='SMA_010',
    forward_returns=['RETURN_OO_01', 'RETURN_OO_05', 'RETURN_OO_10'],
    axvlines=('2020-01-01', '2024-01-01'),  # 垂直参考线
    method='rank_ic'
)

plt.show()
```

--------------------------------

### Stratified Cumulative Return Calculation

Source: https://context7.com/wukan1986/alphainspect/llms.txt

Calculate cumulative returns for each factor quantile to assess monotonicity and profitability.

```python
import polars as pl
from alphainspect.portfolio import calc_cum_return_by_quantile
from alphainspect.utils import with_factor_quantile
from alphainspect import _DATE_

df = pl.read_parquet('data/data.parquet')
df = with_factor_quantile(df, 'SMA_010', quantiles=5, by=[_DATE_], factor_quantile='_fq')

# 计算分层累计收益
ret, cum, avg, std = calc_cum_return_by_quantile(
    df,
    fwd_ret_1='RETURN_OO_01',    # 1期远期收益率
    factor_quantile='_fq'
)

print(cum.tail())
```

--------------------------------

### Factor Quantile Stratification

Source: https://context7.com/wukan1986/alphainspect/llms.txt

Use with_factor_quantile to divide factor values into groups based on quantiles, which is a foundational step for single-factor analysis.

```python
import polars as pl
from alphainspect.utils import with_factor_quantile
from alphainspect import _DATE_

# 加载数据
df = pl.read_parquet('data/data.parquet')

# 对因子进行9分位分层，按日期分组
df = with_factor_quantile(
    df,
    factor='SMA_010',           # 要分层的因子名
    quantiles=9,                # 分成9层
    by=[_DATE_],                # 按日期分组
    factor_quantile='_fq_1'     # 输出的分层列名
)

# 查看分层结果
print(df.select(['date', 'asset', 'SMA_010', '_fq_1']).head(10))
```

--------------------------------

### Generate 3x2 Comprehensive Analysis Report

Source: https://context7.com/wukan1986/alphainspect/llms.txt

Produces a detailed 3x2 grid of charts, extending the 2x2 report with IC histograms and turnover analysis.

```python
import polars as pl
import matplotlib.pyplot as plt
from alphainspect.reports import create_3x2_sheet
from alphainspect.utils import with_factor_quantile
from alphainspect import _DATE_

df = pl.read_parquet('data/data.parquet')
df = with_factor_quantile(df, 'SMA_010', quantiles=9, by=[_DATE_], factor_quantile='_fq_1')

# 生成3x2完整分析图表
fig, ic_dict, hist_dict, cum, avg, std = create_3x2_sheet(
    df,
    factor='SMA_010',
    fwd_ret_1='RETURN_OO_05',
    factor_quantile='_fq_1',
    periods=[1, 5, 10, 20],     # 换手率周期
    figsize=(12, 14),
    axvlines=('2020-01-01', '2024-01-01')
)

# 图表布局：
# [0,0] IC时序图      [0,1] IC直方图
# [1,0] IC月度热力图  [1,1] 分层累计收益
# [2,0] 因子自相关    [2,1] 换手率图

plt.show()
```

--------------------------------

### Batch HTML Report Generation

Source: https://context7.com/wukan1986/alphainspect/llms.txt

Generates multiple HTML reports for different factor groups using multiprocessing to improve efficiency.

```python
import polars as pl
from multiprocessing import get_context
from alphainspect.reports import report_html

INPUT_PATH = 'data/data.parquet'
OUTPUT_PATH = 'output'

def generate_report(kv):
    name, factors = kv
    df = pl.read_parquet(INPUT_PATH)

    report_html(
        name=name,                          # 报表名称
        factors=factors,                    # 因子列表
        df=df,
        output=OUTPUT_PATH,
        fwd_ret_1='RETURN_OO_05',          # 用于计算收益的远期收益率
        quantiles=9,                        # 分层数
        top_k=0,                           # 若>0则使用top_k分层
        axvlines=('2020-01-01', '2024-01-01')
    )
    return 0

if __name__ == '__main__':
    # 定义因子分组
    factors_kv = {
        "SMA": ['SMA_005', 'SMA_010', 'SMA_020'],
        "STD": ['STD_005', 'STD_010', 'STD_020'],
    }

    # 多进程并行生成
    with get_context("spawn").Pool(2) as pool:
        results = list(pool.map(generate_report, factors_kv.items()))

    # 输出：output/SMA.html, output/STD.html
```

--------------------------------

### Calculate Cumulative Returns by Quantile

Source: https://github.com/wukan1986/alphainspect/blob/main/deprecated/deprecated.txt

Calculates cumulative returns for each quantile group, assuming equal weighting within each group. Useful for analyzing performance across different factor-based tiers.

```python
def _calc_cum_return_by_quantile(df: pl.DataFrame, fwd_ret_1: str, period: int = 5, factor_quantile: str = _QUANTILE_) -> pd.DataFrame:
    """分层计算收益。分成N层，层内等权"""
    q_max = df.select(pl.max(factor_quantile)).to_series(0)[0]
    rr = df.pivot(index=_DATE_, columns=_ASSET_, values=fwd_ret_1, aggregate_function='first', sort_columns=True).sort(_DATE_)
    qq = df.pivot(index=_DATE_, columns=_ASSET_, values=factor_quantile, aggregate_function='first', sort_columns=True).sort(_DATE_)

    out = pd.DataFrame(index=rr[_DATE_])
    rr = rr.select(pl.exclude(_DATE_)).to_numpy() + 1  # 日收益
    qq = qq.select(pl.exclude(_DATE_)).to_numpy()  # 分组编号
    # logger.info('累计收益准备数据,period={}', period)

    np.seterr(divide='ignore', invalid='ignore')
    for i in range(int(q_max) + 1):
        # 等权
        b = qq == i
        w = b / b.sum(axis=1).reshape(-1, 1)
        w[(w == 0).all(axis=1), :] = np.nan
        # 权重绝对值和为1
        out[f'G{i}'] = cumulative_returns(rr, w, funds=period, freq=period)
    # !!!直接减是错误的，因为两资金是独立的，资金减少的一份由于资金不足对冲比例已经不再是1:1
    # out['spread'] = out[f'G{q_max}'] - out[f'G0']
    logger.info('累计收益计算完成,period={}\n{}', period, out.tail(1).to_string())
    return out
```

--------------------------------

### Calculate Cumulative Return Spread (Long-Short)

Source: https://github.com/wukan1986/alphainspect/blob/main/deprecated/deprecated.txt

Calculates the spread between the top and bottom quantiles, representing a long-short portfolio return. It handles weight adjustments for a balanced long-short strategy.

```python
def calc_cum_return_spread(df: pl.DataFrame, fwd_ret_1: str, period: int = 5, factor_quantile: str = _QUANTILE_) -> pd.DataFrame:
    """分层计算收益。分成N层，层内等权。
    取Top层和Bottom层。比较不同的计算方法多空收益的区别"""

    q_max = df.select(pl.max(factor_quantile)).to_series(0)[0]
    rr = df.pivot(index=_DATE_, columns=_ASSET_, values=fwd_ret_1, aggregate_function='first', sort_columns=True).sort(_DATE_).fill_nan(0)
    qq = df.pivot(index=_DATE_, columns=_ASSET_, values=factor_quantile, aggregate_function='first', sort_columns=True).sort(_DATE_).fill_nan(-1)

    out = pd.DataFrame(index=rr[_DATE_])
    rr = rr.select(pl.exclude(_DATE_)).to_numpy() + 1  # 日收益
    qq = qq.select(pl.exclude(_DATE_)).to_numpy()  # 分组编号
    logger.info('多空收益准备数据,period={}', period)

    np.seterr(divide='ignore', invalid='ignore')

    # 等权
    w0 = qq == 0
    w9 = qq == q_max
    w0 = w0 / w0.sum(axis=1).reshape(-1, 1)
    w0 = np.where(w0 == w0, w0, 0)
    w9 = w9 / w9.sum(axis=1).reshape(-1, 1)
    w9 = np.where(w9 == w9, w9, 0)
    ww = (w9 - w0) / 2  # 除2，权重绝对值和一定要调整为1，否则后面会计算错误

    # 整行都为0，将其设成nan，后面计算时用于判断是否为0
    ww[(ww == 0).all(axis=1), :] = np.nan
    w0[(w0 == 0).all(axis=1), :] = np.nan
    w9[(w9 == 0).all(axis=1), :] = np.nan

    # 曲线的翻转
    out['1-G0,w=+1'] = 1 - cumulative_returns(rr, w0, funds=period, freq=period)
    # 权重的翻转。资金发生了变化。如果资金不共享，无法完全对冲
    out['G0-1,w=-1'] = cumulative_returns(rr, -w0, funds=period, freq=period) - 1

    out[f'G{q_max},w=+1'] = cumulative_returns(rr, w9, funds=period, freq=period)
    # 资金是共享的，每次调仓时需要将资金平分成两份
    out[f'G{q_max}~G0,w=+.5/-.5'] = cumulative_returns(rr, ww, funds=period, freq=period, init_cash=1.0)
    logger.info('多空收益计算完成,period={}\n{}', period, out.tail(1).to_string())
    return out
```

--------------------------------

### Calculate Cumulative Returns with Specified Weights

Source: https://github.com/wukan1986/alphainspect/blob/main/deprecated/deprecated.txt

Calculates cumulative returns for assets based on externally provided weights. This function assumes separate capital for each asset and is suitable for daily rebalancing.

```python
def calc_cum_return_weights(df: pl.DataFrame, fwd_ret_1: str, period: int = 1) -> pd.DataFrame:
    """指定权重计算收益。不再分层计算。资金也不分份"""
    rr = df.pivot(index=_DATE_, columns=_ASSET_, values=fwd_ret_1, aggregate_function='first', sort_columns=True).sort(_DATE_)
    ww = df.pivot(index=_DATE_, columns=_ASSET_, values=_WEIGHT_, aggregate_function='first', sort_columns=True).sort(_DATE_)

    out = pd.DataFrame(index=rr[_DATE_], columns=rr.columns[1:])
    rr = rr.select(pl.exclude(_DATE_)).to_numpy()  # 日收益
    ww = ww.select(pl.exclude(_DATE_)).to_numpy()  # 权重
    logger.info('权重收益准备数据,period={}', period)

    np.seterr(divide='ignore', invalid='ignore')

    rr = np.where(rr == rr, rr, 0.0)
    # 累计收益分资产，资金不共享
    # 由于是每天换仓，所以不存在空头计算不准的问题
    out[:] = np.cumprod(rr * ww + 1, axis=0)

    logger.info('权重收益计算完成,period={}\n{}', period, out.tail(1).to_string())
    return out
```

--------------------------------

### IC Calculation and Metrics

Source: https://context7.com/wukan1986/alphainspect/llms.txt

Calculate Information Coefficient (IC) matrices and derived metrics like IR to evaluate factor predictive power.

```python
import polars as pl
from alphainspect.ic import calc_ic, rank_ic, mutual_info

df = pl.read_parquet('data/data.parquet')

# 定义因子和收益率
factors = ['SMA_005', 'SMA_010', 'SMA_020']
forward_returns = ['RETURN_OO_01', 'RETURN_OO_05']

# 计算RankIC矩阵
df_ic = calc_ic(
    df,
    factors=factors,
    forward_returns=forward_returns,
    method='rank_ic'            # 或 'mutual_info' 计算互信息
)

print(df_ic.head())

# 计算IC均值和IR
from alphainspect.calc import calc_mean, calc_ir

ic_mean = calc_mean(df_ic)  # IC均值
ic_ir = calc_ir(df_ic)      # IR = IC均值 / IC标准差
print(f"IC均值: {ic_mean}")
print(f"IR值: {ic_ir}")
```

--------------------------------

### Numba Parallel Sub-Portfolio Returns Calculation

Source: https://github.com/wukan1986/alphainspect/blob/main/deprecated/_nb.txt

Calculates sub-portfolio returns with Numba's parallel processing. It simulates fund allocation, rebalancing based on frequency, and handles potential losses.

```python
@jit(nopython=True, nogil=True, fastmath=True, cache=True, parallel=True)
def _sub_portfolio_returns(m: int, n: int,
                           weights: np.ndarray, returns: np.ndarray,
                           funds: int = 1,
                           freq: int = -1,
                           init_cash: float = 1.0) -> np.ndarray:
    """资金分成N份。每隔N天取一次权重，这N天内每次都用上次的市值乘以权重得到新市值，乘以收益率得到日终市值

    输出的数据全不为nan
    """
    if freq == -1:
        freq = m
    # 记录每份的收益率
    out = np.zeros(shape=(m, funds), dtype=float)
    # 资金分成funds份
    for i in prange(funds):
        cashflow = np.zeros(shape=(m, n), dtype=float)  # 多头为0，空头为2*weights
        val = np.zeros(shape=(m, n), dtype=float)
        val[:, 0] = init_cash  # 初始资金全放第0列
        last_sum = init_cash
        for j in range(i, m):
            # 调仓节点，如果设成m表示只第一天进行调仓
            if (j % freq == i) and (last_sum > 0):
                val[j] = last_sum * weights[j]
                # 负数变正数*2，正数变成0
                cashflow[j] = np.maximum(-val[j], 0) * 2
            else:
                # 不调仓，直接取昨天的现金和市值做为盘前值
                cashflow[j] = cashflow[j - 1]
                val[j] = val[j - 1]

            val[j] = returns[j] * val[j]
            last = val[j] + cashflow[j]  # 在这里，很有可能做空亏成负数
            last_sum = np.sum(last)
            if last_sum <= 0:
                # 亏损了，将市值移动到现金流中，不玩了
                cashflow[j] += val[j]
                val[j] = 0

        # 如果不水平求和，就能看到每支股票的累计收益
        np_sum(val + cashflow, axis=1, out=out[:, i])

    return out
```

--------------------------------

### Analyze Factor Turnover and Autocorrelation

Source: https://context7.com/wukan1986/alphainspect/llms.txt

Calculates factor autocorrelation and quantile-based turnover to assess trading costs and factor stability.

```python
import polars as pl
import matplotlib.pyplot as plt
from alphainspect.turnover import (
    calc_auto_correlation,
    calc_quantile_turnover,
    create_turnover_sheet
)
from alphainspect.utils import with_factor_quantile
from alphainspect import _DATE_

df = pl.read_parquet('data/data.parquet')
df = with_factor_quantile(df, 'SMA_010', quantiles=5, by=[_DATE_], factor_quantile='_fq')

# 计算因子自相关（间隔1、5、10、20期）
df_auto_corr = calc_auto_correlation(
    df,
    factor='SMA_010',
    periods=[1, 5, 10, 20]
)
print(df_auto_corr.head())
# 输出：
# ┌────────────┬────────┬────────┬────────┬────────┐
# │ date       │ AC01   │ AC05   │ AC10   │ AC20   │
# ├────────────┼────────┼────────┼────────┼────────┤
# │ 2018-01-05 │ 0.9523 │ 0.8756 │ 0.7892 │ 0.6234 │
# └────────────┴────────┴────────┴────────┴────────┘

# 计算分层换手率
df_turnover = calc_quantile_turnover(
    df,
    factor_quantile='_fq',
    periods=[1, 5, 10, 20]
)

# 生成完整换手率分析图表
create_turnover_sheet(
    df,
    factor='SMA_010',
    factor_quantile='_fq',
    periods=[1, 5, 10, 20],
    axvlines=('2020-01-01',)
)

plt.show()
```

--------------------------------

### Dual-Factor Grouping Analysis

Source: https://context7.com/wukan1986/alphainspect/llms.txt

Performs statistical analysis on two factors simultaneously to identify return characteristics across combined quantiles.

```python
import polars as pl
import matplotlib.pyplot as plt
from alphainspect.plotting import create_describe2_sheet
from alphainspect.utils import with_factor_quantile
from alphainspect import _DATE_

df = pl.read_parquet('data/data.parquet')

# 对两个因子分别分层
df = with_factor_quantile(df, 'SMA_010', quantiles=3, by=[_DATE_], factor_quantile='_fq_sma')
df = with_factor_quantile(df, 'STD_010', quantiles=5, by=[_DATE_], factor_quantile='_fq_std')

# 双因子分组分析
create_describe2_sheet(
    df,
    col='RETURN_OO_05',                    # 要统计的列（可以是收益率）
    factor_quantiles=['_fq_sma', '_fq_std']  # 两个因子分层列
)

# 生成3个热力图：Mean、Std、Count
# 可用于分析3x5=15个组合的收益特征

plt.show()
```

--------------------------------

### Numba Optimized NumPy Tile Function

Source: https://github.com/wukan1986/alphainspect/blob/main/deprecated/_nb.txt

Replicates a 2D NumPy array multiple times along the second axis using Numba. This is an optimized version of `np.tile` for specific 2D tiling scenarios.

```python
@jit(nopython=True, nogil=True, fastmath=True, cache=True)
def np_tile(arr, reps):
    m, n = arr.shape
    out = np.empty(shape=(m, n * reps), dtype=arr.dtype)
    for i in range(reps):
        out[:, i * n:(i + 1) * n] = arr
    return out
```

--------------------------------

### Perform Event Study Analysis

Source: https://context7.com/wukan1986/alphainspect/llms.txt

Analyzes factor performance around specific events, including price changes, layered returns, and win rates.

```python
import polars as pl
import matplotlib.pyplot as plt
from alphainspect.events import with_around_price, create_events_sheet
from alphainspect.utils import with_factor_quantile
from alphainspect import _DATE_

df = pl.read_parquet('data/data.parquet')

# 添加事件前后的标准化价格（前5后15期）
df = with_around_price(
    df,
    price='CLOSE',
    periods_before=5,
    periods_after=15
)

# 因子分层
df = with_factor_quantile(df, 'STD_010', quantiles=3, by=[_DATE_], factor_quantile='_fq')

# 事件分析：分析波动率大于0.005时的表现
create_events_sheet(
    df,
    condition=pl.col('STD_010') > 0.005,  # 事件条件
    fwd_ret_1='RETURN_OO_05',
    show_long_short=True,
    factor_quantile='_fq',
    axvlines=('2020-01-01', '2024-01-01')
)

# 生成的6个子图：
# - 事件前后平均累计收益曲线（按分层）
# - 最高层误差条图
# - 事件胜率时序图
# - 分层累计收益曲线
# - 事件发生次数分布图

plt.show()
```

--------------------------------

### Apply 1D Function Along Axis with Numba

Source: https://github.com/wukan1986/alphainspect/blob/main/deprecated/_nb.txt

Applies a 1D function along a specified axis of a 2D NumPy array using Numba for acceleration. Useful for vectorized operations on rows or columns.

```python
import numpy as np
from numba import jit, prange


@jit(nopython=True, nogil=True, fastmath=True)
def np_apply_along_axis_1d(func1d, axis, arr, out):
    if axis == 0:
        for i in range(len(out)):
            out[i] = func1d(arr[:, i])
    else:
        for i in range(len(out)):
            out[i] = func1d(arr[i, :])
    return out
```

--------------------------------

### Apply 2D Function Along Axis with Numba

Source: https://github.com/wukan1986/alphainspect/blob/main/deprecated/_nb.txt

Applies a function along a specified axis of a 2D NumPy array using Numba. This is distinct from the 1D version and handles operations differently based on the axis.

```python
@jit(nopython=True, nogil=True, fastmath=True)
def np_apply_along_axis_2d(func1d, axis, arr, out):
    if axis == 0:
        for i in range(out.shape[1]):
            out[:, i] = func1d(arr[:, i])
    else:
        for i in range(out.shape[0]):
            out[i, :] = func1d(arr[i, :])
    return out
```

--------------------------------

### Numba Optimized NumPy Cumulative Product

Source: https://github.com/wukan1986/alphainspect/blob/main/deprecated/_nb.txt

Calculates the cumulative product of a NumPy array along a specified axis using Numba. This function utilizes `np_apply_along_axis_2d` for its implementation.

```python
@jit(nopython=True, nogil=True, fastmath=True, cache=True)
def np_cumprod(arr, axis, out):
    return np_apply_along_axis_2d(np.cumprod, axis, arr, out)
```

--------------------------------

### Numba Optimized NumPy Sum Calculation

Source: https://github.com/wukan1986/alphainspect/blob/main/deprecated/_nb.txt

Computes the sum of a NumPy array along a specified axis with Numba acceleration. Supports the `axis` parameter and uses `np_apply_along_axis_1d`.

```python
@jit(nopython=True, nogil=True, fastmath=True, cache=True)
def np_sum(arr, axis, out):
    # sum支持axis
    return np_apply_along_axis_1d(np.sum, axis, arr, out)
```

--------------------------------

### High-Correlation Factor Filtering

Source: https://context7.com/wukan1986/alphainspect/llms.txt

Identifies and filters out highly correlated factors to prevent multicollinearity in multi-factor models.

```python
import polars as pl
from alphainspect.selection import drop_above_corr_thresh
from alphainspect.ic import calc_ic

df = pl.read_parquet('data/data.parquet')

# 计算多因子IC
factors = ['SMA_005', 'SMA_010', 'SMA_020', 'STD_005', 'STD_010']
df_ic = calc_ic(df, factors, ['RETURN_OO_05'])

# 筛选高相关因子
cols_to_drop, above_thresh_pairs = drop_above_corr_thresh(
    df_ic.select(pl.exclude('date')),
    thresh=0.85                           # 相关性阈值
)

print(f"需要删除的因子: {cols_to_drop}")
# ['SMA_005__RETURN_OO_05']

print(f"高相关因子对: {above_thresh_pairs}")
# [('SMA_005__RETURN_OO_05', 'SMA_010__RETURN_OO_05', 0.92)]
```

--------------------------------

### Numba Optimized NumPy Mean Calculation

Source: https://github.com/wukan1986/alphainspect/blob/main/deprecated/_nb.txt

Calculates the mean of a NumPy array along a specified axis using Numba for performance. Leverages the `np_apply_along_axis_1d` function.

```python
@jit(nopython=True, nogil=True, fastmath=True, cache=True)
def np_mean(arr, axis, out):
    return np_apply_along_axis_1d(np.mean, axis, arr, out)
```
Tokens:
Raw