sql - 递归 SQL 语句(PostgreSQL 9.1.4)

Question

PostgreSQL 9.1

业务情况

每个月，都会有一批新的帐户分配给特定的流程。每批可以按月、账户数和账户总余额进行描述。该过程的目标是从客户那里收回部分余额。 然后每月单独跟踪每个批处理(自批处理转移到流程后每月回收的金额)。

目标

我的目标是预测 future 会收回多少金额。

数据定义

create table vintage_data (
    granularity date,       /* Month when account entered process*/
    distance_in_months integer, /* Distance in months from date when accounts entered process*/
    entry_accounts integer,     /* Number of accounts that entered process in a given month*/
    entry_amount numeric,       /* Total amount for account that entered process in a given month*/
    recovery_amount numeric     /* Amount recovered in Nth month on accounts that entered process in a given month */
);

示例数据

insert into vintage_data values('2012-01-31',1,200,100000,1000);
insert into vintage_data values('2012-01-31',2,200,100000,2000);
insert into vintage_data values('2012-01-31',3,200,100000,3000);
insert into vintage_data values('2012-01-31',4,200,100000,3500);
insert into vintage_data values('2012-01-31',5,200,100000,3400);
insert into vintage_data values('2012-01-31',6,200,100000,3300);
insert into vintage_data values('2012-02-28',1,250,150000,1200);
insert into vintage_data values('2012-02-28',2,250,150000,1600);
insert into vintage_data values('2012-02-28',3,250,150000,1800);
insert into vintage_data values('2012-02-28',4,250,150000,1200);
insert into vintage_data values('2012-02-28',5,250,150000,1600);
insert into vintage_data values('2012-03-31',1,200,90000,1300);
insert into vintage_data values('2012-03-31',2,200,90000,1200);
insert into vintage_data values('2012-03-31',3,200,90000,1400);
insert into vintage_data values('2012-03-31',4,200,90000,1000);
insert into vintage_data values('2012-04-30',1,300,180000,1600);
insert into vintage_data values('2012-04-30',2,300,180000,1500);
insert into vintage_data values('2012-04-30',3,300,180000,4000);
insert into vintage_data values('2012-05-31',1,400,225000,2200);
insert into vintage_data values('2012-05-31',2,400,225000,6000);
insert into vintage_data values('2012-06-30',1,100,60000,1000);

计算过程

你可以把数据想象成一个三角矩阵(要预测X值):

distance_in_months                       1      2     3       4      5      6
granularity entry_accounts  entry_amount
2012-01-31  200             100000       1000   2000   3000   3500   3400   3300
2012-02-28  250             150000       1200   1600   1800   1200   1600   (X-1)
2012-03-31  200              90000       1300   1200   1400   1000   (X0)   (X4)
2012-04-30  300             180000       1600   1500   4000   (X1)   (X5)   (X8)
2012-05-31  400             225000       2200   6000   (X2)   (X6)   (X9)   (X11)
2012-06-30  100              60000       1000   (X3)   (X7)   (X10)  (X12   (X13)

算法

我的目标是预测所有缺失点( future )。为了说明这个过程，这是对点X1的计算

1) 使用最多 4 的距离获取前三个月的行总计:

2012-01-31  1000+2000+3000+3500=9500 (d4m3)
2012-02-28  1200+1600+1800+1200=5800 (d4m2)
2012-03-31  1300+1200+1400+1000=4900 (d4m1)

2) 使用最多 3 的距离获取前三个月的行总计:

2012-01-31  1000+2000+3000=6000 (d3m3)
2012-02-28  1200+1600+1800=4600 (d3m2)
2012-03-31  1300+1200+1400=3800 (d3m1)

3) 计算距离 3 和距离 4 的加权平均运行率(由 entry_amount 加权):

(d4m3+d4m2+d4m1)/(100000+150000+90000) = (9500+5800+4900)/(100000+150000+90000) = 20200/340000 = 0.0594
(d3m3+d3m2+d3m1)/(100000+150000+90000) = (6000+4600+3800)/(100000+150000+90000) = 14400/340000 = 0.0424

4)计算距离3和距离4的变化

((d4m3+d4m2+d4m1)/(100000+150000+90000))/((d3m3+d3m2+d3m1)/(100000+150000+90000)) =
= (20200/340000)/(14400/340000) =
= 0.0594/0.0424 = 1.403 (PredictionRateForX1)

5) 使用最大为 3 的距离计算预测月份的行总数:

2012-04-30  1600+1500+4000=7100

6) 使用预测月份的 entry_amount 计算费率

7100/180000 = 0.0394

7) 计算 X1 的预测速率

0.0394 * PredictionRateForX1 = 0.05534

8) 计算X1的金额

(0.05534-0.0394)*180000 = 2869.2

问题

问题是如何使用 SQL 语句计算矩阵的其余部分(从 x-1 到 x13)。很明显，这将需要某种递归算法。

Answer 1

这是一项艰巨的任务，将其拆分以使其更易于管理。我会将其放入带有 RETURN TABLE 的 plpgsql 函数中:

1. 使用交叉表查询为您的“计算过程”矩阵创建一个临时表你需要 tablefunc为此安装的模块。运行(每个数据库一次):

   CREATE EXTENSION tablefunc;
   

2. 按字段更新临时表。

3. 返回表格。

以下演示功能齐全，并使用 PostgreSQL 9.1.4 进行了测试。基于问题中提供的表定义:

-- DROP FUNCTION f_forcast();

CREATE OR REPLACE FUNCTION f_forcast()
  RETURNS TABLE (
  granularity date
 ,entry_accounts numeric
 ,entry_amount numeric
 ,d1 numeric
 ,d2 numeric
 ,d3 numeric
 ,d4 numeric
 ,d5 numeric
 ,d6 numeric) AS
$BODY$
BEGIN

--== Create temp table with result of crosstab() ==--

CREATE TEMP TABLE matrix ON COMMIT DROP AS
SELECT *
FROM   crosstab (
        'SELECT granularity, entry_accounts, entry_amount
               ,distance_in_months, recovery_amount
         FROM   vintage_data
         ORDER  BY 1, 2',

        'SELECT DISTINCT distance_in_months
         FROM   vintage_data
         ORDER  BY 1')
AS tbl (
  granularity date
 ,entry_accounts numeric
 ,entry_amount numeric
 ,d1 numeric
 ,d2 numeric
 ,d3 numeric
 ,d4 numeric
 ,d5 numeric
 ,d6 numeric
 );

ANALYZE matrix; -- update statistics to help calculations


--== Calculations ==--

-- I implemented the first calculation for X1 and leave the rest to you.
-- Can probably be generalized in a loop or even a single statement.

UPDATE matrix m
SET    d4 = (
    SELECT (sum(x.d1) + sum(x.d2) + sum(x.d3) + sum(x.d4))
            /(sum(x.d1) + sum(x.d2) + sum(x.d3)) - 1
            -- removed redundant sum(entry_amount) from equation
    FROM  (
        SELECT *
        FROM   matrix a
        WHERE  a.granularity < m.granularity
        ORDER  BY a.granularity DESC
        LIMIT  3
        ) x
    ) * (m.d1 + m.d2 + m.d3)
WHERE m.granularity = '2012-04-30';

--- Next update X2 ..


--== Return results ==--

RETURN QUERY
TABLE  matrix
ORDER  BY 1;

END;
$BODY$ LANGUAGE plpgsql;

调用:

SELECT * FROM f_forcast();

我已经简化了很多，删除了计算中的一些冗余步骤。
该解决方案采用了多种先进技术。您需要熟悉 PostgreSQL 才能使用它。