Alphalens 是非常著名的一个python因子分析库。但是该库由于目前已经不怎么维护，问题非常多。最新的使用建议使用alphalens-reloaded，地址：stefan-jansen/alphalens-reloaded: Performance analysis of predictive (alpha) stock factors (github.com)。
由于该库的demo都是基于国外雅虎财经的接口yfinance。如果使用国内的akshare配合使用会出现一下问题。需要对Alphalens接口非常熟悉。建议阅读其原始接口的注释，特别是get_clean_factor_and_forward_returns方法。
def get_clean_factor_and_forward_returns(factor,
prices,
groupby=None,
binning_by_group=False,
quantiles=5,
bins=None,
periods=(1, 5, 10),
filter_zscore=20,
groupby_labels=None,
max_loss=0.35,
zero_aware=False,
cumulative_returns=True):
"""
Formats the factor data, pricing data, and group mappings into a DataFrame
that contains aligned MultiIndex indices of timestamp and asset. The
returned data will be formatted to be suitable for Alphalens functions.
It is safe to skip a call to this function and still make use of Alphalens
functionalities as long as the factor data conforms to the format returned
from get_clean_factor_and_forward_returns and documented here
Parameters
----------
factor : pd.Series - MultiIndex
A MultiIndex Series indexed by timestamp (level 0) and asset
(level 1), containing the values for a single alpha factor.
::
-----------------------------------
date | asset |
-----------------------------------
| AAPL | 0.5
-----------------------
| BA | -1.1
-----------------------
2014-01-01 | CMG | 1.7
-----------------------
| DAL | -0.1
-----------------------
| LULU | 2.7
-----------------------
prices : pd.DataFrame
A wide form Pandas DataFrame indexed by timestamp with assets
in the columns.
Pricing data must span the factor analysis time period plus an
additional buffer window that is greater than the maximum number
of expected periods in the forward returns calculations.
It is important to pass the correct pricing data in depending on
what time of period your signal was generated so to avoid lookahead
bias, or delayed calculations.
'Prices' must contain at least an entry for each timestamp/asset
combination in 'factor'. This entry should reflect the buy price
for the assets and usually it is the next available price after the
factor is computed but it can also be a later price if the factor is
meant to be traded later (e.g. if the factor is computed at market
open but traded 1 hour after market open the price information should
be 1 hour after market open).
'Prices' must also contain entries for timestamps following each
timestamp/asset combination in 'factor', as many more timestamps
as the maximum value in 'periods'. The asset price after 'period'
timestamps will be considered the sell price for that asset when
computing 'period' forward returns.
::
----------------------------------------------------
| AAPL | BA | CMG | DAL | LULU |
----------------------------------------------------
Date | | | | | |
----------------------------------------------------
2014-01-01 |605.12| 24.58| 11.72| 54.43 | 37.14 |
----------------------------------------------------
2014-01-02 |604.35| 22.23| 12.21| 52.78 | 33.63 |
----------------------------------------------------
2014-01-03 |607.94| 21.68| 14.36| 53.94 | 29.37 |
----------------------------------------------------
groupby : pd.Series - MultiIndex or dict
Either A MultiIndex Series indexed by date and asset,
containing the period wise group codes for each asset, or
a dict of asset to group mappings. If a dict is passed,
it is assumed that group mappings are unchanged for the
entire time period of the passed factor data.
binning_by_group : bool
If True, compute quantile buckets separately for each group.
This is useful when the factor values range vary considerably
across gorups so that it is wise to make the binning group relative.
You should probably enable this if the factor is intended
to be analyzed for a group neutral portfolio
quantiles : int or sequence[float]
Number of equal-sized quantile buckets to use in factor bucketing.
Alternately sequence of quantiles, allowing non-equal-sized buckets
e.g. [0, .10, .5, .90, 1.] or [.05, .5, .95]
Only one of 'quantiles' or 'bins' can be not-None
bins : int or sequence[float]
Number of equal-width (valuewise) bins to use in factor bucketing.
Alternately sequence of bin edges allowing for non-uniform bin width
e.g. [-4, -2, -0.5, 0, 10]
Chooses the buckets to be evenly spaced according to the values
themselves. Useful when the factor contains discrete values.
Only one of 'quantiles' or 'bins' can be not-None
periods : sequence[int]
periods to compute forward returns on.
filter_zscore : int or float, optional
Sets forward returns greater than X standard deviations
from the the mean to nan. Set it to 'None' to avoid filtering.
Caution: this outlier filtering incorporates lookahead bias.
groupby_labels : dict
A dictionary keyed by group code with values corresponding
to the display name for each group.
max_loss : float, optional
Maximum percentage (0.00 to 1.00) of factor data dropping allowed,
computed comparing the number of items in the input factor index and
the number of items in the output DataFrame index.
Factor data can be partially dropped due to being flawed itself
(e.g. NaNs), not having provided enough price data to compute
forward returns for all factor values, or because it is not possible
to perform binning.
Set max_loss=0 to avoid Exceptions suppression.
zero_aware : bool, optional
If True, compute quantile buckets separately for positive and negative
signal values. This is useful if your signal is centered and zero is
the separation between long and short signals, respectively.
cumulative_returns : bool, optional
If True, forward returns columns will contain cumulative returns.
Setting this to False is useful if you want to analyze how predictive
a factor is for a single forward day.
Returns
-------
merged_data : pd.DataFrame - MultiIndex
A MultiIndex Series indexed by date (level 0) and asset (level 1),
containing the values for a single alpha factor, forward returns for
each period, the factor quantile/bin that factor value belongs to, and
(optionally) the group the asset belongs to.
- forward returns column names follow the format accepted by
pd.Timedelta (e.g. '1D', '30m', '3h15m', '1D1h', etc)
- 'date' index freq property (merged_data.index.levels[0].freq) will be
set to a trading calendar (pandas DateOffset) inferred from the input
data (see infer_trading_calendar for more details). This is currently
used only in cumulative returns computation
::
-------------------------------------------------------------------
| | 1D | 5D | 10D |factor|group|factor_quantile
-------------------------------------------------------------------
date | asset | | | | | |
-------------------------------------------------------------------
| AAPL | 0.09|-0.01|-0.079| 0.5 | G1 | 3
--------------------------------------------------------
| BA | 0.02| 0.06| 0.020| -1.1 | G2 | 5
--------------------------------------------------------
2014-01-01 | CMG | 0.03| 0.09| 0.036| 1.7 | G2 | 1
--------------------------------------------------------
| DAL |-0.02|-0.06|-0.029| -0.1 | G3 | 5
--------------------------------------------------------
| LULU |-0.03| 0.05|-0.009| 2.7 | G1 | 2
--------------------------------------------------------
See Also
--------
utils.get_clean_factor
For use when forward returns are already available.
"""
forward_returns = compute_forward_returns(
factor,
prices,
periods,
filter_zscore,
cumulative_returns,
)
factor_data = get_clean_factor(factor, forward_returns, groupby=groupby,
groupby_labels=groupby_labels,
quantiles=quantiles, bins=bins,
binning_by_group=binning_by_group,
max_loss=max_loss, zero_aware=zero_aware)
return factor_data
源码
使用Akshare获取a股600519数据，然后使用alphalens-reloaded进行最基本的因子分析，因子使用5日均线与10日均线的交叉，代码如下：
import warnings
warnings.filterwarnings('ignore')
import pandas as pd
import alphalens
import seaborn as sns
import akshare as ak
from pytz import timezone
# %matplotlib inline
sns.set_style('white')
# pd.set_option('display.max_columns', None)
# pd.set_option('display.max_rows', None)
# 使用 akshare 的 stock_zh_a_hist 函数
df = ak.stock_zh_a_hist(symbol='600519', period="daily", start_date='20200101', end_date='20201231', adjust="qfq")
# 调整 DataFrame 列名
df.rename(columns={
'日期': 'date',
'开盘': 'open',
'收盘': 'close',
'最高': 'high',
'最低': 'low',
'成交量': 'volume'
}, inplace=True)
df['asset'] = '600519'
# 计算开盘价和收盘价之差
# df['factor'] = df['close']
df['ma5'] = df['close'].rolling(window=5).mean().fillna(0)
df['ma10'] = df['close'].rolling(window=10).mean().fillna(0)
df['factor'] = df['ma5']-df['ma10']
df = df.iloc[20:]
df.head(30)
# 使用dff，不影响原来的df
dff = df
dff['date'] = pd.to_datetime(dff['date'])
dff = dff.set_index(['date', 'asset'])
dff.index = dff.index.set_levels([dff.index.levels[0].tz_localize('UTC'), dff.index.levels[1]])
factor = dff['factor']
# factor.head()
# print(factor)
df['date'] = pd.to_datetime(df['date']).dt.tz_localize('UTC') # convert date column to datetime format with UTC timezone
df.set_index(['date', 'asset'], inplace=True)
# select 'close' column to create the prices dataframe
prices = df['close'].unstack('asset')
prices.head()
print(prices.index.tz)
print(factor.index.levels[0].tz)
# print(prices)
# 现在对factor和prices进行对齐
# factor, prices = factor.align(prices, join='inner', axis=0)
factor_data = alphalens.utils.get_clean_factor_and_forward_returns( factor,
prices,
groupby=None,
binning_by_group=False,
quantiles=2,
bins=None,
periods=(1, 5, 10),
filter_zscore=20,
groupby_labels=None,
max_loss=0.35,
zero_aware=True,
cumulative_returns=True,
)
# factor_data.head()
alphalens.tears.create_full_tear_sheet(factor_data,long_short=False)
结果如图：
常见错误
-
AttributeError: ‘Index’ object has no attribute ‘tz’
时区问题，国外的数据默认都带了时区，国内的tushare、akshare需要自己把时区加上，可以参考上述源码的处理。 -
MaxLossExceededError: max_loss (35.0%) exceeded 100.0%, consider increasing it.
get_clean_factor_and_forward_returns函数默认的max_loss为35.0%，自己也可以配置，最开始使用默认的quantiles=5会出现这个问题，可以把入参quantiles改为2。该因子可分为正数和负数两类。 -
Inferred frequency None from passed values does not conform to passed frequency C
频率问题，解决频率问题可以将数据同步一下，可能是由于部分NaN值或者将factor与prices值对齐。
如有问题欢迎评论区留言或者私信。
扫一扫
7054
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
