SAS信用卡评分之变量分段
这一篇的文章来讲变量分段，在我之前的文章中，涉及到变量分段的代码是有的，早开始的等高等宽分啊，后面的基于基尼系数以及基于iv值对于字符变量的分类都有。链接在这,这！这！这！。
sas字符变量基于iv值的最优分类
sas字符变量基于基尼系数的最优分类
sas分类变量（3）之排序后按人数分组
sas分类变量（2）之等宽分组
sas分类变量（1）之按分位数分组
但是我要是没记错的话，我好像还欠你们一个数值变量的最优分段。这篇文章如果后面还可以贴的话，那就附上数据集批量输出数值变量最优分段的代码。为什么说“可以贴的话”，那是我之前写过一篇文章，微信提醒我超过20000字啦，叫我删掉。
进入正题。当评分卡建模的时候用的逻辑回归，最好是将连续变量变成分段变量，即字符变量。把字符变量，观测的种类达到10种以上的时候，建议分下类，最好每个变量（无论数值还是字符）控制在3-7段之间，这是我的建议哈，要是你领导叫你分8段，你就千万就要听领导的。
然后，我来说下，我这里的最优分段怎么就是最优的呢?
01字符变量
先发这张图给粘上来，然后我就用简单粗暴的语言解释最优是怎么最优的。就是先把每种情况都列出来，刚开始每一种情况都是一类，然后你还要输入因变量，所以1中就是找出最优的二元分割方法，把原来的一大群先分两大类，然后第2的套路还是跟1一样的，知道分成5份。你问我二元分割最好的指标是什么，你还记得我之前写的代码之前都有带“基尼系数”，“iv值”吗，就是按照这个指标去分的啦。然后这里还要说一点就是，你要是一个变量，总共就1、2、3、4种情况，然后还要最优分段分五份这不是为难嘛。假设你觉得4种情况的分层没有特点，想分的有特点一点，那就可以试着分成3份，2份。分出来的结果对比一下iv值，要是3份的iv值比4份还高或者一样的话，那就是3份还要好些，因为我们都知道变量分段越多iv值越高。
02
数值变量
这是数值变量最优分段的图，其实套路跟字符变量很像，但是数值变量就多了顺序，所以还是有点跟字符有点不像。首先连续变量被分为大量等距的小分段，譬如区间是100的变量，然后就分成50段，那么就是1-2就是一组。那按照跟刚才的字符变量一样的分法类似，就是先分两份，只是对于字符变量多了顺序。但是这里这里要注意一点就是，你本来1、2、3代表的是类的话，在这里就需要把他转成字符，就不要是数值丢进去分段。同样的，要是你不知道分几段的时候，试几次，看下iv值，取一个你觉得最好的iv值。
我是分割线
好的，两种变量的分类也就这样了啦，好像也没写多少字哦，那就贴代码吧。
options mlogic;
options nomlogic;
%macro gvalue(binds,m_value);
proc sql noprint;
%local i j R N;/*生成局部变量*/
select max(bin)into:R from &binds;/**/
select sum(total) into: N from &binds;/**/
%do i=1 %to &R;
%local N_&i._1 N_&i._2 N_&i._s N_s_1 N_s_2;
Select sum(Ni1) into :N_&i._1 from &BinDS where Bin =&i ;
Select sum(Ni2) into :N_&i._2 from &BinDS where Bin =&i ;
Select sum(Total) into :N_&i._s from &BinDS where Bin =&i ;
Select sum(Ni1) into :N_s_1 from &BinDS ;
Select sum(Ni2) into :N_s_2 from &BinDS ;
%end;
quit;
/* 检查缺失值 */
%do i=1 %to &R;
%do j=1 %to 2;
%local N_&i._&j;
%if (&&N_&i._&j=.) or (&&N_&i._&j=0) %then %do ;
%let &M_Value=.;
%return;
%end;
%end;
%end;
%do i=1 %to &r;
%local E_&i;
%let E_&i=0;
%do j=1 %to 2;
%let E_&i = %sysevalf(&&E_&i - (&&N_&i._&j/&&N_&i._s)*%sysfunc(log(%sysevalf(&&N_&i._&j/&&N_&i._s))) );
%end;
%let E_&i = %sysevalf(&&E_&i/%sysfunc(log(2)));
%end;
%local E;
%let E=0;
%do j=1 %to 2;
%let E=%sysevalf(&E - (&&N_s_&j/&N)*%sysfunc(log(&&N_s_&j/&N)) );
%end;
%let E=%sysevalf(&E / %sysfunc(log(2)));
%local Er;
%let Er=0;
%do i=1 %to &r;
%let Er=%sysevalf(&Er+ &&N_&i._s * &&E_&i / &N);
%end;
%let &M_Value=%sysevalf(1 - &Er/&E);
%return;
%mend;
%macro CalcMerit(BinDS, ix, M_Value);
%local n_11 n_12 n_21 n_22 n_1s n_2s n_s1 n_s2;
proc sql noprint;
select sum(Ni1) into :n_11 from &BinDS where i<=&ix;
select sum(Ni1) into :n_21 from &BinDS where i> &ix;
select sum(Ni2) into : n_12 from &BinDS where i<=&ix ;
select sum(Ni2) into : n_22 from &binDS where i> &ix ;
select sum(total) into :n_1s from &BinDS where i<=&ix ;
select sum(total) into :n_2s from &BinDS where i> &ix ;
select sum(Ni1) into :n_s1 from &BinDS;
select sum(Ni2) into :n_s2 from &BinDS;
quit;
%local N E1 E2 E Er;
%let N=%eval(&n_1s+&n_2s);
%let E1=%sysevalf(-( (&n_11/&n_1s)*%sysfunc(log(%sysevalf(&n_11/&n_1s))) +
(&n_12/&n_1s)*%sysfunc(log(%sysevalf(&n_12/&n_1s)))) / %sysfunc(log(2)) ) ;
%let E2=%sysevalf(-( (&n_21/&n_2s)*%sysfunc(log(%sysevalf(&n_21/&n_2s))) +
(&n_22/&n_2s)*%sysfunc(log(%sysevalf(&n_22/&n_2s)))) / %sysfunc(log(2)) ) ;
%let E =%sysevalf(-( (&n_s1/&n )*%sysfunc(log(%sysevalf(&n_s1/&n ))) +
(&n_s2/&n )*%sysfunc(log(%sysevalf(&n_s2/&n )))) / %sysfunc(log(2)) ) ;
%let Er=%sysevalf(1-(&n_1s*&E1+&n_2s*&E2)/(&N*&E));
%let &M_value=&Er;
%return;
%mend;
%macro BestSplit(BinDs, BinNo);
%local mb i value BestValue BestI;
proc sql noprint;
select count(*) into: mb from &BinDs where Bin=&BinNo;
quit;
%let BestValue=0;
%let BestI=1;
%do i=1 %to %eval(&mb-1);
%let value=;
%CalcMerit(&BinDS, &i, Value);
%if %sysevalf(&BestValue<&value) %then %do;
%let BestValue=&Value;
%let BestI=&i;
%end;
%end;
data &BinDS;
set &BinDS;
if i<=&BestI then Split=1;
else Split=0;
drop i;
run;
proc sort data=&BinDS;
by Split;
run;
data &BinDS;
retain i 0;
set &BinDs;
by Split;
if first.split then i=1;
else i=i+1;
run;
%mend;
%macro CandSplits(BinDS, NewBins);
proc sort data=&BinDS;
by Bin PDV1;
run;
%local Bmax i value;
proc sql noprint;
select max(bin) into: Bmax from &BinDS;
%do i=1 %to &Bmax;
%local m&i;
create table Temp_BinC&i as select * from &BinDS where Bin=&i;
select count(*) into:m&i from Temp_BinC&i;
%end;
create table temp_allVals (BinToSplit num, DatasetName char(80), Value num);
run;quit;
%do i=1 %to &Bmax;
%if (&&m&i>1) %then %do;
%BestSplit(Temp_BinC&i, &i);
data temp_trysplit&i;
set temp_binC&i;
if split=1 then Bin=%eval(&Bmax+1);
run;
Data temp_main&i;
set &BinDS;
if Bin=&i then delete;
run;
Data Temp_main&i;
set temp_main&i temp_trysplit&i;
run;
%let value=;
%GValue(temp_main&i, Value);
proc sql noprint;
insert into temp_AllVals values(&i, "temp_main&i", &Value);
run;quit;
%end;
%end;
proc sort data=temp_allVals;
by descending value;
run;
data _null_;
set temp_AllVals(obs=1);
call symput("bin", compress(BinToSplit));
run;
Data &NewBins;
set Temp_main&Bin;
drop split;
run;
/* Clean the workspace */
/*proc datasets nodetails nolist library=work;*/
/* delete temp_AllVals %do i=1 %to &Bmax; Temp_BinC&i temp_TrySplit&i temp_Main&i %end; ; */
/*run;*/
/*quit;*/
%mend;
%macro BinContVar(DSin, IVVar, DVVar, MMax, Acc, DSVarMap);
%local VarMax VarMin;
proc sql noprint;
select min(&IVVar), max(&IVVar) into :VarMin, :VarMax from &DSin;
quit;
%local Mbins i MinBinSize;
%let Mbins=%sysfunc(int(%sysevalf(1.0/&Acc)));
%let MinBinSize=%sysevalf((&VarMax-&VarMin)/&Mbins);
%do i=1 %to %eval(&Mbins);
%local Lower_&i Upper_&i;
%let Upper_&i = %sysevalf(&VarMin + &i * &MinBinSize);
%let Lower_&i = %sysevalf(&VarMin + (&i-1)*&MinBinSize);
%end;
%let Lower_1 = %sysevalf(&VarMin-0.0001);
%let Upper_&Mbins=%sysevalf(&VarMax+0.0001);
data Temp_DS;
set &DSin;
%do i=1 %to %eval(&Mbins-1);
if &IVVar>=&&Lower_&i and &IVVar < &&Upper_&i Then Bin=&i;
%end;
if &IVVar>=&&Lower_&Mbins and &IVVar <= &&Upper_&MBins Then Bin=&MBins;
keep &IVVar &DVVar Bin;
run;
data temp_blimits;
%do i=1 %to %Eval(&Mbins-1);
Bin_LowerLimit=&&Lower_&i;
Bin_UpperLimit=&&Upper_&i;
Bin=&i;
output;
%end;
Bin_LowerLimit=&&Lower_&Mbins;
Bin_UpperLimit=&&Upper_&Mbins;
Bin=&Mbins;
output;
run;
proc sort data=temp_blimits;
by Bin;
run;
proc freq data=Temp_DS noprint;
table Bin*&DVvar /out=Temp_cross;
table Bin /out=Temp_binTot;
run;
proc sort data=temp_cross;
by Bin;
run;
proc sort data= temp_BinTot;
by Bin;
run;
data temp_cont;
merge Temp_cross(rename=count=Ni2 ) temp_BinTot(rename=Count=total) temp_BLimits ;
by Bin;
Ni1=total-Ni2;
PDV1=bin;
label Ni2= total=;
if Ni1=0 then output;
else if &DVVar=1 then output;
drop percent &DVVar;
run;
data temp_contold;
set temp_cont;
run;
proc sql noprint;
%local mx;
%do i=1 %to &Mbins;
select count(*) into : mx from Temp_cont where Bin=&i;
%if (&mx>0) %then %do;
select Ni1, Ni2, total, bin_lowerlimit, bin_upperlimit into :Ni1,:Ni2,:total, :bin_lower, :bin_upper
from temp_cont where Bin=&i;
%if (&i=&Mbins) %then %do;
select max(bin) into :i1 from temp_cont where Bin<&Mbins;
%end;
%else %do;
select min(bin) into :i1 from temp_cont where Bin>&i;
%end;
%if (&Ni1=0) or (&Ni2=0) or (&total=0) %then %do;
update temp_cont set Ni1=Ni1+&Ni1 ,
Ni2=Ni2+&Ni2 ,
total=total+&Total
where bin=&i1;
%if (&i<&Mbins) %then %do;
update temp_cont set Bin_lowerlimit = &Bin_lower
where bin=&i1;
%end;
%else %do;
update temp_cont set Bin_upperlimit = &Bin_upper
where bin=&i1;
%end;
delete from temp_cont where bin=&i;
%end;
%end;
%end;
quit;
proc sort data=temp_cont;
by pdv1;
run;
%local m;
data temp_cont;
set temp_cont;
i=_N_;
Var=bin;
Bin=1;
call symput("m", compress(_N_));
run;
%local Nbins ;
%let Nbins=1;
%DO %WHILE (&Nbins <&MMax);
%CandSplits(temp_cont, Temp_Splits);
Data Temp_Cont;
set Temp_Splits;
run;
%let NBins=%eval(&NBins+1);
%end;
data temp_Map1 ;
set temp_cont(Rename=Var=OldBin);
drop Ni2 PDV1 Ni1 i ;
run;
proc sort data=temp_Map1;
by Bin OldBin ;
run;
data temp_Map2;
retain LL 0 UL 0 BinTotal 0;
set temp_Map1;
by Bin OldBin;
Bintotal=BinTotal+Total;
if first.bin then do;
LL=Bin_LowerLimit;
BinTotal=Total;
End;
if last.bin then do;
UL=Bin_UpperLimit;
output;
end;
drop Bin_lowerLimit Bin_upperLimit Bin OldBin total;
run;
proc sort data=temp_map2;
by LL;
run;
data &DSVarMap;
set temp_map2;
Bin=_N_;
run;
/* Clean the workspace */
/*proc datasets nodetails library=work nolist;*/
/* delete temp_bintot temp_blimits temp_cont temp_contold temp_cross temp_ds temp_map1*/
/* temp_map2 temp_splits;*/
/*run; quit;*/
%mend;
%macro ApplyMap2(DSin, VarX, NewVarX, DSVarMap, DSout);
%local m i;
proc sql noprint;
select count(Bin) into:m from &DSVarMap;
quit;
%do i=1 %to &m;
%local Upper_&i Lower_&i Bin_&i;
%end;
data _null_;
set &DSVarMap;
call symput ("Upper_"||left(_N_), UL);
call symput ("Lower_"||left(_N_), LL);
call symput ("Bin_"||left(_N_), Bin);
run;
Data &DSout;
set &DSin;
IF &VarX < &Upper_1 Then &NewVarX=&Bin_1;
%do i=2 %to %eval(&m-1);
if &VarX >= &&Lower_&i and &VarX < &&Upper_&i Then &NewVarX=&&Bin_&i;
%end;
if &VarX >= &&Lower_&i Then &NewVarX=&&Bin_&i;
DROP &VarX.;
Run;
%mend;
%macro var_namelist(data=,coltype=,tarvar=,dsor=);
%let lib=%upcase(%scan(&data.,1,'.'));
%let dname=%upcase(%scan(&data.,2,'.'));
%global var_list var_num;
proc sql ;
create table &dsor. as
select name
from sashelp.VCOLUMN
where left(libname)="&lib." and left(memname)="&dname." and type="&coltype." and lowcase(name)^=lowcase("&tarvar.") and lowcase(name)^="appl_id";
quit;
%mend;
%macro pub_best(data=,tarvar=,MMax=,ACC=,DSout=);
proc datasets lib=work;
delete _all_;
run;
%var_namelist(data=&data.,coltype=num,tarvar=&tarvar.,dsor=aa);
data _null_;
set aa;
call symput (compress("var"||left(_n_)),compress(name));
call symput(compress("n"),compress(_n_));
run;
%do i=1 %to &n.;
%put &&Var&i.;
%BinContVar(DSin=&data., IVVar=&&Var&i., DVVar=&tarvar.,MMax=&MMax., ACC=&Acc., DSVarMap=AA_1);
%ApplyMap2(DSin=&data., VarX=&&Var&i., NewVarX=N_&&Var&i., DSVarMap=AA_1, DSout=&DSout.);
%END;
%MEND;
这代码有点长，你就直接复制到sas里面看吧。
data=填入原始的数据集
tarvar=因变量;
MMax=分几组;
Acc=刚才是分几组，譬如你是1-100，那么你设定的是0.01，那就是分成100组，建议acc设定在0.01-0.05之间;
DSout=输出数据集。
代码是我调试好的，可以直接用。
数据分析咨询请扫描二维码
