测试IC及ICIR加权策略.zip_ICIR加权啥意思_icicir_icir加权权重_因子ICIR_测试ICIR加权方法

共1个文件

py：1个

版权申诉

32 浏览量 2022-09-22 14:13:55 上传评论 1 收藏 2KB ZIP 举报

在IT领域，尤其是在量化投资和金融数据分析中，IC（Information Coefficient）和ICIR（Information Coefficient Improvement Ratio）是衡量因子预测能力的重要指标。本文将深入探讨IC和ICIR的概念，以及如何使用Python来实现ICIR加权策略。让我们理解什么是IC。IC，即信息系数，是用来评估一个因子对未来收益预测能力的统计量。它通过计算因子得分与未来收益之间的相关性来度量。IC值越接近1，表示因子的预测能力越强；接近0则表示因子预测能力较弱；如果为负，则表示因子与未来收益呈现负相关。 ICIR，即信息系数改善比率，是在原有IC基础上引入的一种优化策略。它旨在通过动态调整因子权重来提升组合的整体预测效果。ICIR不仅考虑了因子之间的相关性和预测能力，还考虑了权重分配对整体IC的影响。目标是最大化组合的ICIR，从而提高投资决策的准确性。 Python在实现ICIR加权策略时，通常涉及以下步骤： 1. 数据准备：收集并清洗因子数据和对应的投资标的的收益率数据。 2. 计算IC：使用皮尔逊相关系数或其他相关性测量方法计算每个因子与未来收益的IC。 3. 构建权重矩阵：根据因子的IC值和其它因素（如稳定性、可解释性等）构建初始权重矩阵。 4. 优化权重：通过迭代算法（如梯度下降、遗传算法等）优化权重，目标函数为最大化ICIR。 5. 验证策略：使用回测框架（如Backtrader、Zipline等）验证优化后的策略，评估其在历史数据上的表现。 6. 持续监控：实时更新因子数据和权重，确保策略适应市场变化。在Python中，可以利用pandas库处理数据，numpy库进行数值计算，scipy库进行统计分析，以及sklearn或optimization库进行权重优化。例如，以下是一个简单的ICIR计算和优化的伪代码： ```python import pandas as pd import numpy as np from scipy.stats import pearsonr # 假设df是包含因子得分和未来收益的数据框 df = pd.read_csv('因子数据.csv') # 计算IC ic = {} for factor in df.columns[:-1]: corr, _ = pearsonr(df[factor], df['收益']) ic[factor] = corr # 初始化权重 weights = np.ones(len(ic)) / len(ic) # 定义优化目标函数 def optimize_weights(weights): # 使用权重计算ICIR，并返回目标值 pass # 使用优化算法（如梯度下降） from scipy.optimize import minimize result = minimize(optimize_weights, weights, method='gradient_descent') # 获取最优权重并应用到策略 optimal_weights = result.x ``` 以上代码只是一个基础示例，实际应用中可能需要根据具体需求进行复杂的数据预处理、因子选择、权重约束等操作。此外，ICIR加权策略并非万能，还需要结合其他投资理论和实践进行综合决策。 IC和ICIR是量化投资中用于评价因子预测能力和优化策略的重要工具。Python提供了一系列强大的库，能够帮助我们有效地实现ICIR加权策略，从而提升投资组合的收益潜力。在实际应用中，理解和熟练运用这些概念和技术对于提高投资决策的科学性和有效性至关重要。

资源推荐

资源详情

资源评论

收起资源包目录

测试IC及ICIR加权策略.zip （1个子文件）

测试IC及ICIR加权策略.py 9KB

# -*- coding: utf-8 -*- """ Created on Tue Jan 23 09:56:58 2018 @author: Administrator """ import numpy as np import pandas as pd import alphalens import matplotlib.pyplot as plt import os import datetime import time import math df = pd.read_csv('NewGuzhiYinziInitialData.csv') df.drop(df.columns[[0]],axis=1,inplace = True) fg = pd.read_csv('NewGuzhiYinziDirection.csv') fg.drop(fg.columns[[0]],axis=1,inplace = True) fg.iloc[:] = 1 need_test_factor_name = df.columns[6:-2] need_test_factor_direction = fg.iloc[:,0].values needtesttime = pd.unique(df.loc[:,'tdate']) all_single_df_fenshu = [] for itest in needtesttime : #itest = needtesttime[386] print(itest) single_df_fenshu = df.loc[df.tdate == itest,['tdate','security_code','stockname','securityWeight','week_return','stock_end_hangye']] # 把因子值转化为排序值 for jyest in list(range(len(need_test_factor_name))) : # print(jyest) # jyest = 0 single_df = df.loc[df.tdate == itest,:] single_df.insert(0,'judge_trade',single_df['judge_zhangdie'] + single_df['judge_tingpai']) # single_df['judge_trade'] = single_df[:,'judge_zhangdie'] + single_df[:,'judge_tingpai'] xx = single_df.sort_values(['stock_end_hangye','judge_trade',need_test_factor_name[jyest]],ascending = [1,1,need_test_factor_direction[jyest]]) xx = xx[(xx['stock_end_hangye']>0) & (xx['stock_end_hangye']<12)] def ListGivingFenshu(ssff,ssgg): ssfff = ssff[ssgg==0] aaff = list(range(1,np.sum(ssfff == ssfff)+1,1)) + list(10000*np.ones(np.sum(ssfff != ssfff)+np.sum(ssgg>=1),dtype = 'int')) return aaff yy1 = xx.groupby('stock_end_hangye').apply(lambda subf:ListGivingFenshu(subf[need_test_factor_name[jyest]],subf['judge_trade'])) zz1 = [] for i in list(range(1,12)): zz1.extend(yy1[i]) zz2 = pd.Series(zz1,index = xx.index) single_df_fenshu[need_test_factor_name[jyest]] = zz2 all_single_df_fenshu.append(single_df_fenshu) all_end_df_fenshu = pd.concat(all_single_df_fenshu,ignore_index = True) all_all_zz = [] for jjtest in list(range(len(need_test_factor_name))): print(jjtest) all_zz = [] for iitest in needtesttime[:-1]: # print(iitest) single_df_fenshu = all_end_df_fenshu.loc[all_end_df_fenshu.tdate == iitest,:] # 剔除没有数据的股票，并计算所有因子的 IC 值：当前排序分数与下期收益的相关系数 single_factor_fenshu = single_df_fenshu.loc[:,['tdate', 'security_code','week_return','stock_end_hangye',need_test_factor_name[jjtest]]] single_factor_fenshu = single_factor_fenshu.sort_values(['stock_end_hangye',single_factor_fenshu.columns[-1]]) single_factor_fenshu = single_factor_fenshu[(single_factor_fenshu['stock_end_hangye']>0) & (single_factor_fenshu['stock_end_hangye']<12)] def ListGivingCorr(ddgg,ddhh): endddgg = ddgg[ddhh.values<10000] endddhh = ddhh[ddhh.values<10000] # return endddgg.corr(endddhh) #拆分为 10 档，每档内部取平均收益率为该档的收益 if len(endddgg) < 10: return endddgg.corr(endddhh) else: fendangjiange = math.floor(len(endddgg)/10) fendangleave = len(endddgg) - 10 * fendangjiange fendangend = fendangjiange*np.ones(10) + np.append(np.zeros(10-fendangleave),np.ones(fendangleave)) fendang_shouyi = [] fendang_local = [] for kkii in list(range(len(fendangend))): fendang_shouyi.append(np.mean(endddgg.iloc[int(sum(fendangend[:kkii])):int(sum(fendangend[:kkii+1]))])) fendang_local.append(kkii + 1) fendang_shouyi_series = pd.Series(fendang_shouyi) fendang_local_series = pd.Series(fendang_local) return fendang_shouyi_series.corr(fendang_local_series) zz = single_factor_fenshu.groupby('stock_end_hangye').apply(lambda subf:ListGivingCorr(subf['week_return'],subf[single_factor_fenshu.columns[-1]])) zz.fillna(0,inplace = True) if len(zz)>11: break all_zz.append(zz.values) all_all_zz.append(np.column_stack((all_zz))) aaa_all_all_zz = np.vstack(all_all_zz) #plt.plot(all_all_zz[0,:]) # 先测试第 1 个行业，根据IC—IR计算的权重，来合成全新的权重 TestZhidinghangye = 11 aaa_aaa_single_zz = aaa_all_all_zz[list(range(TestZhidinghangye - 1,aaa_all_all_zz.shape[0],11)),:] SlideSize = 200 #计算IC——IR所需要的窗口期 aaa_all_direction = [] aaa_all_quanzhong = [] for TTDay in list(range(200,aaa_aaa_single_zz.shape[1]+1)): # 首次计算 ss_compute_data = np.mean(aaa_aaa_single_zz[:,(TTDay - SlideSize):TTDay],1)/np.std(aaa_aaa_single_zz[:,(TTDay - SlideSize):TTDay],1) ss_compute_data_quanzhong = abs(ss_compute_data)/sum(abs(ss_compute_data)) ss_compute_data_drect = np.zeros(len(ss_compute_data)) ss_compute_data_drect[ss_compute_data>0] = 1 aaa_all_direction.append(ss_compute_data_drect) aaa_all_quanzhong.append(ss_compute_data_quanzhong) aaa_all_direction = np.vstack((aaa_all_direction)) aaa_all_quanzhong = np.vstack((aaa_all_quanzhong)) aaa_all_recordtime = needtesttime[200:] AAAA_Record_result = [] for TTiday in list(range(len(aaa_all_recordtime)-1)): print(aaa_all_recordtime[TTiday]) TTtesttime = aaa_all_recordtime[TTiday] # TTtestdirect = aaa_all_direction[TTiday,:] # TTtestquanzhong = aaa_all_quanzhong[TTiday,:] TTtestdirect = aaa_all_direction[0,:] TTtestquanzhong = aaa_all_quanzhong[0,:] # TTtestquanzhong = [1, 0, 0, 0, 0, 0, 0, 0, 0] single_hangye_df_fenshu = df.loc[(df.tdate == TTtesttime) & (df.stock_end_hangye == TestZhidinghangye),['tdate','security_code','stockname','securityWeight','week_return','stock_end_hangye']] # 把因子值转化为排序值 for jyest in list(range(len(need_test_factor_name))) : # print(jyest) # jyest = 0 single_hangye_df = df.loc[(df.tdate == TTtesttime) & (df.stock_end_hangye == TestZhidinghangye),:] single_hangye_df.insert(0,'judge_trade',single_hangye_df['judge_zhangdie'] + single_hangye_df['judge_tingpai']) # single_df['judge_trade'] = single_df[:,'judge_zhangdie'] + single_df[:,'judge_tingpai'] xx = single_hangye_df.sort_values(['stock_end_hangye','judge_trade',need_test_factor_name[jyest]],ascending = [1,1,int(TTtestdirect[jyest])]) xx = xx[(xx['stock_end_hangye']>0) & (xx['stock_end_hangye']<12)] def ListGivingFenshuHH(ssff,ssgg): ssfff = ssff[ssgg==0] sssfff = ssfff[ssfff == ssfff] if len(sssfff) < 10 : sssfff_fenshu = list(range(len(sssfff),0,-1)) else: fendangjiange = math.floor(len(sssfff)/10) fendangleave = len(sssfff) - 10 * fendangjiange fendangend = fendangjiange*np.ones(10) + np.append(np.zeros(10-fendangleave),np.ones(fendangleave)) sssfff_fenshu = np.zeros(len(sssfff)) for kkii in list(range(len(fendangend))): sssfff_fenshu[int(sum(fendangend[:kkii])):int(sum(fendangend[:kkii+1]))] = 10 - kkii aaff = list(sssfff_fenshu) + list(np.zeros(np.sum(ssfff != ssfff))) + list((-10000)*np.ones(np.sum(ssgg>=1))) return aaff yy1 = xx.groupby('stock_end_hangye').apply(lambda s

评论收藏

内容反馈

版权申诉