Are there any Linear Regression Function in SQL Server?

Are the开发者_运维百科re any Linear Regression Function in SQL Server 2005/2008, similar to the the Linear Regression functions in Oracle ?

To the best of my knowledge, there is none. Writing one is pretty straightforward, though. The following gives you the constant alpha and slope beta for y = Alpha + Beta * x + epsilon:

-- test data (GroupIDs 1, 2 normal regressions, 3, 4 = no variance)
WITH some_table(GroupID, x, y) AS
(       SELECT 1,  1,  1    UNION SELECT 1,  2,  2    UNION SELECT 1,  3,  1.3  
  UNION SELECT 1,  4,  3.75 UNION SELECT 1,  5,  2.25 UNION SELECT 2, 95, 85    
  UNION SELECT 2, 85, 95    UNION SELECT 2, 80, 70    UNION SELECT 2, 70, 65    
  UNION SELECT 2, 60, 70    UNION SELECT 3,  1,  2    UNION SELECT 3,  1, 3
  UNION SELECT 4,  1,  2    UNION SELECT 4,  2,  2),
 -- linear regression query
/*WITH*/ mean_estimates AS
(   SELECT GroupID
          ,AVG(x * 1.)                                             AS xmean
          ,AVG(y * 1.)                                             AS ymean
    FROM some_table
    GROUP BY GroupID
),
stdev_estimates AS
(   SELECT pd.GroupID
          -- T-SQL STDEV() implementation is not numerically stable
          ,CASE      SUM(SQUARE(x - xmean)) WHEN 0 THEN 1 
           ELSE SQRT(SUM(SQUARE(x - xmean)) / (COUNT(*) - 1)) END AS xstdev
          ,     SQRT(SUM(SQUARE(y - ymean)) / (COUNT(*) - 1))     AS ystdev
    FROM some_table pd
    INNER JOIN mean_estimates  pm ON pm.GroupID = pd.GroupID
    GROUP BY pd.GroupID, pm.xmean, pm.ymean
),
standardized_data AS                   -- increases numerical stability
(   SELECT pd.GroupID
          ,(x - xmean) / xstdev                                    AS xstd
          ,CASE ystdev WHEN 0 THEN 0 ELSE (y - ymean) / ystdev END AS ystd
    FROM some_table pd
    INNER JOIN stdev_estimates ps ON ps.GroupID = pd.GroupID
    INNER JOIN mean_estimates  pm ON pm.GroupID = pd.GroupID
),
standardized_beta_estimates AS
(   SELECT GroupID
          ,CASE WHEN SUM(xstd * xstd) = 0 THEN 0
                ELSE SUM(xstd * ystd) / (COUNT(*) - 1) END         AS betastd
    FROM standardized_data pd
    GROUP BY GroupID
)
SELECT pb.GroupID
      ,ymean - xmean * betastd * ystdev / xstdev                   AS Alpha
      ,betastd * ystdev / xstdev                                   AS Beta
FROM standardized_beta_estimates pb
INNER JOIN stdev_estimates ps ON ps.GroupID = pb.GroupID
INNER JOIN mean_estimates  pm ON pm.GroupID = pb.GroupID

Here GroupID is used to show how to group by some value in your source data table. If you just want the statistics across all data in the table (not specific sub-groups), you can drop it and the joins. I have used the WITH statement for sake of clarity. As an alternative, you can use sub-queries instead. Please be mindful of the precision of the data type used in your tables as the numerical stability can deteriorate quickly if the precision is not high enough relative to your data.

EDIT: (in answer to Peter's question for additional statistics like R2 in the comments)

You can easily calculate additional statistics using the same technique. Here is a version with R2, correlation, and sample covariance:

-- test data (GroupIDs 1, 2 normal regressions, 3, 4 = no variance)
WITH some_table(GroupID, x, y) AS
(       SELECT 1,  1,  1    UNION SELECT 1,  2,  2    UNION SELECT 1,  3,  1.3  
  UNION SELECT 1,  4,  3.75 UNION SELECT 1,  5,  2.25 UNION SELECT 2, 95, 85    
  UNION SELECT 2, 85, 95    UNION SELECT 2, 80, 70    UNION SELECT 2, 70, 65    
  UNION SELECT 2, 60, 70    UNION SELECT 3,  1,  2    UNION SELECT 3,  1, 3
  UNION SELECT 4,  1,  2    UNION SELECT 4,  2,  2),
 -- linear regression query
/*WITH*/ mean_estimates AS
(   SELECT GroupID
          ,AVG(x * 1.)                                             AS xmean
          ,AVG(y * 1.)                                             AS ymean
    FROM some_table pd
    GROUP BY GroupID
),
stdev_estimates AS
(   SELECT pd.GroupID
          -- T-SQL STDEV() implementation is not numerically stable
          ,CASE      SUM(SQUARE(x - xmean)) WHEN 0 THEN 1 
           ELSE SQRT(SUM(SQUARE(x - xmean)) / (COUNT(*) - 1)) END AS xstdev
          ,     SQRT(SUM(SQUARE(y - ymean)) / (COUNT(*) - 1))     AS ystdev
    FROM some_table pd
    INNER JOIN mean_estimates  pm ON pm.GroupID = pd.GroupID
    GROUP BY pd.GroupID, pm.xmean, pm.ymean
),
standardized_data AS                   -- increases numerical stability
(   SELECT pd.GroupID
          ,(x - xmean) / xstdev                                    AS xstd
          ,CASE ystdev WHEN 0 THEN 0 ELSE (y - ymean) / ystdev END AS ystd
    FROM some_table pd
    INNER JOIN stdev_estimates ps ON ps.GroupID = pd.GroupID
    INNER JOIN mean_estimates  pm ON pm.GroupID = pd.GroupID
),
standardized_beta_estimates AS
(   SELECT GroupID
          ,CASE WHEN SUM(xstd * xstd) = 0 THEN 0
                ELSE SUM(xstd * ystd) / (COUNT(*) - 1) END         AS betastd
    FROM standardized_data
    GROUP BY GroupID
)
SELECT pb.GroupID
      ,ymean - xmean * betastd * ystdev / xstdev                   AS Alpha
      ,betastd * ystdev / xstdev                                   AS Beta
      ,CASE ystdev WHEN 0 THEN 1 ELSE betastd * betastd END        AS R2
      ,betastd                                                     AS Correl
      ,betastd * xstdev * ystdev                                   AS Covar
FROM standardized_beta_estimates pb
INNER JOIN stdev_estimates ps ON ps.GroupID = pb.GroupID
INNER JOIN mean_estimates  pm ON pm.GroupID = pb.GroupID

EDIT 2 improves numerical stability by standardizing data (instead of only centering) and by replacing STDEV because of numerical stability issues. To me, the current implementation seems to be the best trade-off between stability and complexity. I could improve stability by replacing my standard deviation with a numerically stable online algorithm, but this would complicate the implementation substantantially (and slow it down). Similarly, implementations using e.g. Kahan(-Babuška-Neumaier) compensations for the SUM and AVG seem to perform modestly better in limited tests, but make the query much more complex. And as long as I do not know how T-SQL implements SUM and AVG (e.g. it might already be using pairwise summation), I cannot guarantee that such modifications always improve accuracy.

This is an alternate method, based off a blog post on Linear Regression in T-SQL, which uses the following equations:

The SQL suggestion in the blog uses cursors though. Here's a prettified version of a forum answer that I used:

table
-----
X (numeric)
Y (numeric)

/**
 * m = (nSxy - SxSy) / (nSxx - SxSx)
 * b = Ay - (Ax * m)
 * N.B. S = Sum, A = Mean
 */
DECLARE @n INT
SELECT @n = COUNT(*) FROM table
SELECT (@n * SUM(X*Y) - SUM(X) * SUM(Y)) / (@n * SUM(X*X) - SUM(X) * SUM(X)) AS M,
       AVG(Y) - AVG(X) *
       (@n * SUM(X*Y) - SUM(X) * SUM(Y)) / (@n * SUM(X*X) - SUM(X) * SUM(X)) AS B
FROM table

I've actually written an SQL routine using Gram-Schmidt orthoganalization. It, as well as other machine learning and forecasting routines, is available at sqldatamine.blogspot.com

At the suggestion of Brad Larson I've added the code here rather than just direct users to my blog. This produces the same results as the linest function in Excel. My primary source is Elements of Statistical Learning (2008) by Hastie, Tibshirni and Friedman.

--Create a table of data
create table #rawdata (id int,area float, rooms float, odd float,  price float)

insert into #rawdata select 1, 2201,3,1,400
insert into #rawdata select 2, 1600,3,0,330
insert into #rawdata select 3, 2400,3,1,369
insert into #rawdata select 4, 1416,2,1,232
insert into #rawdata select 5, 3000,4,0,540

--Insert the data into x & y vectors
select id xid, 0 xn,1 xv into #x from #rawdata
union all
select id, 1,rooms  from #rawdata
union all
select id, 2,area  from #rawdata
union all
select id, 3,odd  from #rawdata

select id yid, 0 yn, price yv  into #y from #rawdata

--create a residuals table and insert the intercept (1)
create table #z (zid int, zn int, zv float)
insert into #z select id , 0 zn,1 zv from #rawdata

--create a table for the orthoganal (#c) & regression(#b) parameters
create table #c(cxn int, czn int, cv float) 
create table #b(bn int, bv float) 


--@p is the number of independent variables including the intercept (@p = 0)
declare @p int
set @p = 1


--Loop through each independent variable and estimate the orthagonal parameter (#c)
-- then estimate the residuals and insert into the residuals table (#z)
while @p <= (select max(xn) from #x)
begin   
        insert into #c
    select  xn cxn,  zn czn, sum(xv*zv)/sum(zv*zv) cv 
        from #x join  #z on  xid = zid where zn = @p-1 and xn>zn group by xn, zn

    insert into #z
    select zid, xn,xv- sum(cv*zv) 
        from #x join #z on xid = zid   join  #c  on  czn = zn and cxn = xn  where xn = @p and zn<xn  group by zid, xn,xv

    set @p = @p +1
end

--Loop through each independent variable and estimate the regression parameter by regressing the orthoganal
-- resiuduals on the dependent variable y
while @p>=0 
begin

    insert into #b
    select zn, sum(yv*zv)/ sum(zv*zv) 
        from #z  join 
            (select yid, yv-isnull(sum(bv*xv),0) yv from #x join #y on xid = yid left join #b on  xn=bn group by yid, yv) y
        on zid = yid where zn = @p  group by zn

    set @p = @p-1
end

--The regression parameters
select * from #b

--Actual vs. fit with error
select yid, yv, fit, yv-fit err from #y join 
    (select xid, sum(xv*bv) fit from #x join #b on xn = bn  group by xid) f
     on yid = xid

--R Squared
select 1-sum(power(err,2))/sum(power(yv,2)) from 
(select yid, yv, fit, yv-fit err from #y join 
    (select xid, sum(xv*bv) fit from #x join #b on xn = bn  group by xid) f
     on yid = xid) d

There are no linear regression functions in SQL Server. But to calculate a Simple Linear Regression (Y' = bX + A) between pairs of data points x,y - including the calculation of the Correlation Coefficient, Coefficient of Determination (R^2) and Standard Estimate of Error (Standard Deviation), do the following:

For a table regression_data with numeric columns x and y:

declare @total_points int 
declare @intercept DECIMAL(38, 10)
declare @slope DECIMAL(38, 10)
declare @r_squared DECIMAL(38, 10)
declare @standard_estimate_error DECIMAL(38, 10)
declare @correlation_coefficient DECIMAL(38, 10)
declare @average_x  DECIMAL(38, 10)
declare @average_y  DECIMAL(38, 10)
declare @sumX DECIMAL(38, 10)
declare @sumY DECIMAL(38, 10)
declare @sumXX DECIMAL(38, 10)
declare @sumYY DECIMAL(38, 10)
declare @sumXY DECIMAL(38, 10)
declare @Sxx DECIMAL(38, 10)
declare @Syy DECIMAL(38, 10)
declare @Sxy DECIMAL(38, 10)

Select 
@total_points = count(*),
@average_x = avg(x),
@average_y = avg(y),
@sumX = sum(x),
@sumY = sum(y),
@sumXX = sum(x*x),
@sumYY = sum(y*y),
@sumXY = sum(x*y)
from regression_data

set @Sxx = @sumXX - (@sumX * @sumX) / @total_points
set @Syy = @sumYY - (@sumY * @sumY) / @total_points
set @Sxy = @sumXY - (@sumX * @sumY) / @total_points

set @correlation_coefficient = @Sxy / SQRT(@Sxx * @Syy) 
set @slope = (@total_points * @sumXY - @sumX * @sumY) / (@total_points * @sumXX - power(@sumX,2))
set @intercept = @average_y - (@total_points * @sumXY - @sumX * @sumY) / (@total_points * @sumXX - power(@sumX,2)) * @average_x
set @r_squared = (@intercept * @sumY + @slope * @sumXY - power(@sumY,2) / @total_points) / (@sumYY - power(@sumY,2) / @total_points)

-- calculate standard_estimate_error (standard deviation)
Select
@standard_estimate_error = sqrt(sum(power(y - (@slope * x + @intercept),2)) / @total_points)
From regression_data

Here it is as a function that takes a table type of type: table (Y float, X double) which is called XYDoubleType and assumes our linear function is of the form AX + B. It returns A and B a Table column just in case you want to have it in a join or something

CREATE FUNCTION FN_GetABForData(
 @XYData as XYDoubleType READONLY
 ) RETURNS  @ABData TABLE(
            A  FLOAT,
            B FLOAT, 
            Rsquare FLOAT )
 AS
 BEGIN
    DECLARE @sx FLOAT, @sy FLOAT
    DECLARE @sxx FLOAT,@syy FLOAT, @sxy FLOAT,@sxsy FLOAT, @sxsx FLOAT, @sysy FLOAT
    DECLARE @n FLOAT, @A FLOAT, @B FLOAT, @Rsq FLOAT

    SELECT @sx =SUM(D.X) ,@sy =SUM(D.Y), @sxx=SUM(D.X*D.X),@syy=SUM(D.Y*D.Y),
        @sxy =SUM(D.X*D.Y),@n =COUNT(*)
    From @XYData D
    SET @sxsx =@sx*@sx
    SET @sxsy =@sx*@sy
    SET @sysy = @sy*@sy

    SET @A = (@n*@sxy -@sxsy)/(@n*@sxx -@sxsx)
    SET @B = @sy/@n  - @A*@sx/@n
    SET @Rsq = POWER((@n*@sxy -@sxsy),2)/((@n*@sxx-@sxsx)*(@n*@syy -@sysy))

    INSERT INTO @ABData (A,B,Rsquare) VALUES(@A,@B,@Rsq)

    RETURN 
 END

To add to @icc97 answer, I have included the weighted versions for the slope and the intercept. If the values are all constant the slope will be NULL (with the appropriate settings SET ARITHABORT OFF; SET ANSI_WARNINGS OFF;) and will need to be substituted for 0 via coalesce().

Here is a solution written in SQL:

with d as (select segment,w,x,y from somedatasource)
select segment,

avg(y) - avg(x) *
((count(*) * sum(x*y)) - (sum(x)*sum(y)))/
((count(*) * sum(x*x)) - (Sum(x)*Sum(x)))   as intercept,

((count(*) * sum(x*y)) - (sum(x)*sum(y)))/
((count(*) * sum(x*x)) - (sum(x)*sum(x))) AS slope,

avg(y) - ((avg(x*y) - avg(x)*avg(y))/var_samp(X)) * avg(x) as interceptUnstable,
(avg(x*y) - avg(x)*avg(y))/var_samp(X) as slopeUnstable,
(Avg(x * y) - Avg(x) * Avg(y)) / (stddev_pop(x) * stddev_pop(y)) as correlationUnstable,

(sum(y*w)/sum(w)) - (sum(w*x)/sum(w)) *
((sum(w)*sum(x*y*w)) - (sum(x*w)*sum(y*w)))/
  ((sum(w)*sum(x*x*w)) - (sum(x*w)*sum(x*w)))   as wIntercept,

((sum(w)*sum(x*y*w)) - (sum(x*w)*sum(y*w)))/
  ((sum(w)*sum(x*x*w)) - (sum(x*w)*sum(x*w))) as wSlope,

(count(*) * sum(x * y) - sum(x) * sum(y)) / (sqrt(count(*) * sum(x * x) - sum(x) * sum(x))
* sqrt(count(*) * sum(y * y) - sum(y) * sum(y))) as correlation,

(sum(w) * sum(x*y*w) - sum(x*w) * sum(y*w)) /
(sqrt(sum(w) * sum(x*x*w) - sum(x*w) * sum(x*w)) * sqrt(sum(w) * sum(y*y*w)
- sum(y*w) * sum(y*w))) as wCorrelation,

count(*) as n

from d where x is not null and y is not null group by segment

Where w is the weight. I double checked this against R to confirm the results. One may need to cast the data from somedatasource to floating point. I included the unstable versions to warn you against those. (Special thanks goes to Stephan in another answer.)

Update: added weighted correlation

I have translated the Linear Regression Function used in the funcion Forecast in Excel, and created an SQL function that returns a,b, and the Forecast. You can see the complete teorical explanation in the excel help for FORECAST fuction. Firs of all you will need to create the table data type XYFloatType:

 CREATE TYPE [dbo].[XYFloatType] 
AS TABLE(
[X] FLOAT,
[Y] FLOAT)

Then write the follow function:

    /*
-- =============================================
-- Author:      Me      :)
-- Create date: Today   :)
-- Description: (Copied Excel help): 
--Calculates, or predicts, a future value by using existing values. 
The predicted value is a y-value for a given x-value. 
The known values are existing x-values and y-values, and the new value is predicted by using linear regression. 
You can use this function to predict future sales, inventory requirements, or consumer trends.
-- =============================================
*/

CREATE FUNCTION dbo.FN_GetLinearRegressionForcast

(@PtXYData as XYFloatType READONLY ,@PnFuturePointint)
RETURNS @ABDData TABLE( a FLOAT, b FLOAT, Forecast FLOAT)
AS

BEGIN 
    DECLARE  @LnAvX Float
            ,@LnAvY Float
            ,@LnB Float
            ,@LnA Float
            ,@LnForeCast Float
    Select   @LnAvX = AVG([X])
            ,@LnAvY = AVG([Y])
    FROM @PtXYData;
    SELECT @LnB =  SUM ( ([X]-@LnAvX)*([Y]-@LnAvY) )  /  SUM (POWER([X]-@LnAvX,2))
    FROM @PtXYData;
    SET @LnA = @LnAvY - @LnB * @LnAvX;
    SET @LnForeCast = @LnA + @LnB * @PnFuturePoint;
    INSERT INTO @ABDData ([A],[B],[Forecast]) VALUES (@LnA,@LnB,@LnForeCast)
    RETURN 
END

/*
your tests: 

 (I used the same values that are in the excel help)
DECLARE @t XYFloatType 
INSERT @t VALUES(20,6),(28,7),(31,9),(38,15),(40,21)        -- x and y values
SELECT *, A+B*30 [Prueba]FROM dbo.FN_GetLinearRegressionForcast@t,30);
*/

I hope the following answer helps one understand where some of the solutions come from. I am going to illustrate it with a simple example, but the generalization to many variables is theoretically straightforward as long as you know how to use index notation or matrices. For implementing the solution for anything beyond 3 variables you'll Gram-Schmidt (See Colin Campbell's answer above) or another matrix inversion algorithm.

Since all the functions we need are variance, covariance, average, sum etc. are aggregation functions in SQL, one can easily implement the solution. I've done so in HIVE to do linear calibration of the scores of a Logistic model - amongst many advantages, one is that you can function entirely within HIVE without going out and back in from some scripting language.

The model for your data (x_1, x_2, y) where your data points are indexed by i, is

y(x_1, x_2) = m_1*x_1 + m_2*x_2 + c

The model appears "linear", but needn't be, For example x_2 can be any non-linear function of x_1, as long as it has no free parameters in it, e.g. x_2 = Sinh(3*(x_1)^2 + 42). Even if x_2 is "just" x_2 and the model is linear, the regression problem isn't. Only when you decide that the problem is to find the parameters m_1, m_2, c such that they minimize the L2 error do you have a Linear Regression problem.

The L2 error is sum_i( (y[i] - f(x_1[i], x_2[i]))^2 ). Minimizing this w.r.t. the 3 parameters (set the partial derivatives w.r.t. each parameter = 0) yields 3 linear equations for 3 unknowns. These equations are LINEAR in the parameters (this is what makes it Linear Regression) and can be solved analytically. Doing this for a simple model (1 variable, linear model, hence two parameters) is straightforward and instructive. The generalization to a non-Euclidean metric norm on the error vector space is straightforward, the diagonal special case amounts to using "weights".

Back to our model in two variables:

y = m_1*x_1 + m_2*x_2 + c

Take the expectation value =>

= m_1* + m_2* + c (0)

Now take the covariance w.r.t. x_1 and x_2, and use cov(x,x) = var(x):

cov(y, x_1) = m_1*var(x_1) + m_2*covar(x_2, x_1) (1)

cov(y, x_2) = m_1*covar(x_1, x_2) + m_2*var(x_2) (2)

These are two equations in two unknowns, which you can solve by inverting the 2X2 matrix.

In matrix form: ... which can be inverted to yield ... where

det = var(x_1)*var(x_2) - covar(x_1, x_2)^2

(oh barf, what the heck are "reputation points? Gimme some if you want to see the equations.)

In any case, now that you have m1 and m2 in closed form, you can solve (0) for c.

I checked the analytical solution above to Excel's Solver for a quadratic with Gaussian noise and the residual errors agree to 6 significant digits.

Contact me if you want to do Discrete Fourier Transform in SQL in about 20 lines.