均值方差投资组合模型案例（数据和MATLAB代码）.rar_matlab投资组合_mean_ori3j_投资组合_组合模型

%Author: Moeti Ncube clear all; close all; format long g; %Indicator to run section 1 ind=1; if ind==1 %% Section 1: Only need to be run once per day %Pull 10 years of daily stock prices, up to yesterday, from the S&P 500 stock list in %s&p500.xlsx file. pull_stocks; %Resaves stock data with addition statistics: % mean_data: 30 day historical average ohlc prices % std_data: 30 day historical std ohlc prices % z-data: z score movement from historical price levels %(currently resaves data from 2008 up to yesterday's %date) resave_stock_data %Reformats files into a format that is allows for cross-sectional %backtesting aggregate_stocks %Pull S&P500 data as baseline indicator sp_price='^GSPC'; getyahoo10(sp_price,pwd); %% end sp_price='^GSPC'; %% Section 2: This section performes the optimization %File directory labels folder_base=pwd; folder_stock_data_base=[pwd,'\stock_data_base']; folder_stock_data=[pwd,'\stock_data']; %Number of days in optimized period hist_lag=90; %Out of sample date begin_date=floor(now)-1; end_date=floor(now)-1; %Risk tolerance parameter for mv criterion mw_alpha=.01; %Transaction costs (Based on tiered pricing structure from IB) trans_cost=0.0035; %Bid alpha: how many alpha standard deviations off the open price bid_alpha=.25; %Strategy overview: %The Algorithm take each stock, within the s&p 500, and submits the following limit orders: % 1) Bid Limit order: mean-alpha std deviations % 2) Off Limit order: mean+alpha std deviations % The historical performance of each bidding strategy over % the past 'hist_lag' period is treated as individual strategy within a % portfolio basket of strategies. %The algorithm assigns a weighting, between 0 and 1, to each individual strategy, %so that the Mean-Variance criteria over the entire portfolio basket of %strategies is optimized. %This code applies a unique approach to this %optimization (see optimization section), using ideas from dynamic programming, %to quickly compute the optimization of a large portfolio matrix %The optimal allocation, determine over the previous hist_lag period, is %then applied the next day 'out of sample'. %This procedure is iteratively backtested from the 'begin_date' to the %'end_date'; the daily % return performance is computed and stored. %Variables: %port_hist: The historical performance of the optimal collection of stocks %port_matrix: The optimized basket of stocks %column1: The stock location %column2: Long=1, Short=-1; %column3: weight of row, sum of column3 equals 1 %column4: return of row out of sample. datevector=datenum(begin_date):1:datenum(end_date); d_agg=[]; n_agg=[]; m_agg=[]; s_agg=[]; z_agg=[]; for iter=1:length(datevector) %Create dataset matrices for daily backtests, if iter==1 [d_temp,n_temp,m_temp,s_temp,z_temp]=stock_timeseriespull_all(datevector(iter)-hist_lag,datevector(iter)-2,folder_stock_data,folder_base); d_hist=[d_agg;d_temp]; n_hist=[n_agg;n_temp]; m_hist=[m_agg;m_temp]; s_hist=[s_agg;s_temp]; z_hist=[z_agg;z_temp]; else [d_temp,n_temp,m_temp,s_temp,z_temp]=stock_timeseriespull_all(datevector(iter)-2,datevector(iter)-2,folder_stock_data,folder_base); f=length(find(d_hist(:,1)==d_hist(1,1)))+1; d_hist=[d_hist(f:end,:);d_temp]; n_hist=[n_hist(f:end,:);n_temp]; m_hist=[m_hist(f:end,:);m_temp]; s_hist=[s_hist(f:end,:);s_temp]; z_hist=[z_hist(f:end,:);z_temp]; end [d_fcst_1,n_fcst_1,m_fcst_1,s_fcst_1,z_fcst_1]=stock_timeseriespull_all(datevector(iter)-1,datevector(iter)-1,folder_stock_data,folder_base); [d_fcst_2,n_fcst_2,m_fcst_2,s_fcst_2,z_fcst_2]=stock_timeseriespull_all(datevector(iter),datevector(iter),folder_stock_data,folder_base); d_agg=[d_hist;d_fcst_1;d_fcst_2]; n_agg=[n_hist;n_fcst_1;n_fcst_2]; m_agg=[m_hist;m_fcst_1;m_fcst_2]; s_agg=[s_hist;s_fcst_1;s_fcst_2]; z_agg=[z_hist;z_fcst_1;z_fcst_2]; d_length=hist_lag+1; [stock_list,d_final,m_final,s_final,z_final,last_z]=get_stock_data(d_length,d_agg,n_agg,m_agg,s_agg,z_agg); %% Optimization section port_hist=zeros(1,hist_lag); wght_node=[0:.01:1]'; weight_vector=[]; return_vector=[]; row_iter=1; clear return_long; clear long_values; clear return_short; clear short_values; clear z_vector; clear ls_vector; clear stock_vector; for i=1:length(stock_list) %OHLC history, hist mean and std. dev histories. open_hist=d_final(:,2,i); high_hist=d_final(:,3,i); low_hist=d_final(:,4,i); close_hist=d_final(:,5,i); mu_hist=m_final(:,2,i); std_hist=s_final(:,2,i); % Long bidding: bid 'alpha std. deviations' below the open price % If Low of day is less than submitted bid, we assume our order would have % cleared in market. Exit position at the close of the day. bid=mu_hist-bid_alpha*std_hist; return_long=(low_hist<bid).*((close_hist'-bid'-2*trans_cost)./bid')'; return_long(isnan(return_long))=0; off=mu_hist+bid_alpha*std_hist; return_short=(high_hist>off).*((off'-close_hist'-2*trans_cost)./off')'; return_short(isnan(return_short))=0; stock_hist_long=return_long(1:end-1)'; clear mv_temp; for j=1:length(wght_node) port_hist_temp=wght_node(j)*port_hist+(1-wght_node(j))*stock_hist_long; mv_temp(j)=var(port_hist_temp)-mw_alpha*mean(port_hist_temp); end [temp_1,temp_2]=min(mv_temp); port_hist=wght_node(temp_2)*port_hist+(1-wght_node(temp_2))*stock_hist_long; weight_vector=wght_node(temp_2)*weight_vector; weight_vector(row_iter)=(1-wght_node(temp_2)); return_vector(row_iter)=return_long(end); ls_vector(row_iter)=1; stock_vector(row_iter)=i; row_iter=row_iter+1; stock_hist_short=return_short(1:end-1)'; clear mv_temp; for j=1:length(wght_node) port_hist_temp=wght_node(j)*port_hist+(1-wght_node(j))*stock_hist_short; mv_temp(j)=var(port_hist_temp)-mw_alpha*mean(port_hist_temp); end [temp_1,temp_2]=min(mv_temp); port_hist=wght_node(temp_2)*port_hist+(1-wght_node(temp_2))*stock_hist_short; weight_vector=wght_node(temp_2)*weight_vector; weight_vector(row_iter)=(1-wght_node(temp_2)); return_vector(row_iter)=return_short(end); ls_vector(row_iter)=-1; stock_vector(row_iter)=i; row_iter=row_iter+1; end port_matrix_temp=[stock_vector',ls_vector',weight_vector',return_vector']; port_matrix=port_matrix_temp(port_matrix_temp(:,3)>0,:); end date_one=datevector(iter)-hist_lag; date_two=datevector(iter)-1; hist_dtes=date_one:1:date_two; %Pull S&P500 data as baseline indicator, compute cumulative return over the %historical lag period. stock_file=[sp_price,'.csv']; [s1,s2,s3]=xlsread(stock_file); s500_data=[datenum(s2(2:end,1)),s1]; f=find(s500_data(:,1)>=date_one & s500_data(:,1)<=date_two); s500_data_2=s500_data(f,:); close_cumret=(s500_data_2(:,end-1)-s500_data_2(end,end-1))./s500_data_2(end,end-1); clear adj_cumret; for i=1:hist_lag [m1,m2]=min(abs(s500_data_2(:,1)-hist_dtes(i))); adj_cumret(i)=close_cumret(m2); end %Plots the cumulative return performance hold on title('Cumulative Portfolio Return over historical period compared to S&P 500 performance') plot(hist_dtes,cumsum(port_hist),'b') plot(hist_dtes,adj_cumret,'r') legend('Mean-Variance Optimized Cumulative Return','S&P 500 Cumulative Return') datetick('x','yyyymm','keeplimits') xlabel('Dates') ylabel('Cumulative Percentages') a=[cellstr(num2str(get(gca,'ytick')'*100))]; pct = char(ones(size(a,1),1)*'%'); new_yticks = [char(a),pct]; set(gca,'yticklabel',new_yticks); %% %One day out of sample portfolio return oos_date=year(datevector(iter))*10000+month(datevector(iter))*100+day(datevector(iter)); oos_ret=sum(port_matrix(:,end-1).*port_matrix(:,end)); oos_return=[oos_date,oos_ret]