squeezing variation for profit

Copyright 2005 Northrop Grumman Corporation0

SqueezingVariation forProfitCMMI® Technology Conference and User GroupDenver, Colorado

Thursday, November 17, 2005

D. R. CorpronDivision Manager & Six Sigma Master Black BeltNorthrop Grumman Corporation


Background…

� CMMI Level 4,Quantitatively Managedcovers both theorganizational and projectaspects of processstability and capability

� Stability and Capabilityare not just nobleconcepts, they haveeconomic value and areabout managing variation


The problem…

� The economic value ofrendering processesstable and capable isoften incalculable

� And, the return oninvestment of placingmore processes underquantitative managementlikewise isindeterminable

� So, how to quantify thebenefit?


Situations where variation manifests asschedule misses is a problem..

� Projects miss committeddelivery dates due tosystematic underestimatesof the effort to performtasks

� Projects miss committeddelivery dates due to poorexecution and control ofproject tasks

� Missed delivery datesoften have direconsequences

In Cost ofQuality lingo,

these arefailure costs

Schedule often is a majorconcern of Customers


The concept of failure costs…� Defect correction� Budget misses� Processing discrepancy

reports (DR’s)� Retesting� Unscheduled downtime� Inventory shrinkage� Schedule misses� Invoice errors� Payroll errors� Erroneous status reporting� Lost data

Internal Failure Cost arethe costs that result from afailure to…“Do it right the first time.”


Example…

� The Build Schedulingsubprocess was put underQuantitative Management

� Same process usedacross several projects todetermine scheduleperformance


Data were collected for 30 samples…

20Build 2558Build 1519Build 512Build 24337Build 1410Build 415Build 2319Build 13-1Build 311Build 2220Build 121Build 289Build 2123Build 1128Build 1

95Build 3062Build 2011Build 1031Build 2910Build 1987Build 913Build 288Build 182Build 8

4Build 2753Build 1790Build 74Build 262Build 165Build 6

± DaysEarly or

LateBuild

± DaysEarly or

LateBuild

±DaysEarly or

LateBuild

Fairly clear that schedule performance isan issue


Histogram reveals the shape of the data…

Before

Fre

que

n cy

1501209060300

16

14

12

10

8

6

4

2

0

Histogram of Delivery Performance

Thisdistribution

clearly is non-normal (typicalfor time data)

Data that are not normal present analytic challenges


Further analysis shows the data to be alognormal distribution…

Before+5

Perc

ent

1000100101

99

95

90

80

70

60504030

20

10

5

1

Loc 3.172Scale 1.020N 30AD 0.473P-Value 0.226

Probability Plot of Before+5Lognormal - 95% CINote that 5 is

added to eachnumber to

eliminate anynegativevalues

All points fallwithin the 95%confidenceinterval


Charting the data shows the process to bestable…

Build

Natu

ralL

ogo f

Day

sE a

r ly

orL a

te

Bui ld

28

Build

25

Build

22

Build

19

Build

16

Build

13

Build

10

Bui ld

7

Build

4

Build

1

7

6

5

4

3

2

1

0

_X=3.172

UCL=6.224

LCL=0.121

I Chart of Before+5Using Box-Cox Transformation With Lambda = 0.00This process

is stable asno special

cause pointsappear

Number actually ise3.172 or 24 days less 5

or 19 days

Stable processes lend themselves to improvement


The process is not capable…

654321

USL*transformed data

Sample Mean* 3.1725StDev(Overall)* 1.02886

LSL *Target *USL 5Sample Mean 42.6Sample N 30StDev(Overall) 64.4483

LSL* *Target* *USL* 1.60944

After Transformation

Process DataZ.Bench -1.52Z.LSL *Z.USL -1.52Ppk -0.51Cpm *

Overall Capability

% < LSL *% > USL 96.67% Total 96.67

Observed Performance% < LSL* *% > USL* 93.56% Total 93.56

Exp. Overall Performance

Process Capability of Before+5Using Box-Cox Transformation With Lambda = 0

Fitteddistribution

Actualdata

Can expect theprocess to belate 94% of thetime


An improvement team went to work…

� The team conducted athorough CausalAnalysis and Resolution

� They implemented anew Build Schedulingprocess


Data were collected for 30 new samples…

-2Build 5517Build 4524Build 35-2Build 547Build 444Build 3422Build 5311Build 434Build 33

3Build 52177Build 42-4Build 3288Build 5135Build 41-3Build 31

3Build 603Build 5038Build 40-2Build 595Build 49-3Build 39-2Build 5815Build 4828Build 3811Build 579Build 470Build 37

4Build 56-2Build 4623Build 36

± DaysEarly or

LateBuild

± DaysEarly or

LateBuild

± DaysEarly or

LateBuild

The performance looks better. But, byhow much, and what dollar benefit?


The “After” process is still stable…

Build

Natu

ralL

ogof

Day

sE a

r ly

orLa

te

Build

58

Build

55

Build

52

Build

49

Build

46

Build

43

Build

40

Build

37

Build

34

Build

31

6

5

4

3

2

1

0

-1

_X=2.369

UCL=5.701

LCL=-0.964

I Chart of After+5Using Box-Cox Transformation With Lambda = 0.00

Number actually ise2.369 or 11 days less 5

days or 6 days


Comparing the “Before” to the “After”shows a change in the data distribution…

Freq

uenc

y

350300250200150100500

100

80

60

40

20

0

350300250200150100500

Before+5 After+5Loc 3.172Scale 1.020N 30

Before+5

Loc 2.369Scale 1.203N 30

After+5

Histogram of Before+5, After+5Lognormal

The newapproach

clearly reinedin the variation

Fitteddistribution

Actualdata

These parameterswill be used later


And, a hypothesis test shows thatdifference is indeed real…

P-Value > .05;not significant

95% confidenceinterval does notcontain zero

A Mann Whitneytest is similar toa t-test but canbe applied to

non-normal data


And more the process is more capable…

4.83.62.41.20.0

USL*transformed data

Sample Mean* 2.36853StDev(Overall)* 1.21385

LSL *Target *USL 5Sample Mean 22.0333Sample N 30StDev(Overall) 35.7601

LSL* *Target* *USL* 1.60944

After Transformation

Process DataZ.Bench -0.63Z.LSL *Z.USL -0.63Ppk -0.21Cpm *

Overall Capability

% < LSL *% > USL 70.00% Total 70.00

Observed Performance% < LSL* *% > USL* 73.41% Total 73.41

Exp. Overall Performance

Process Capability of After+5Using Box-Cox Transformation With Lambda = 0

Was 94%,now 73%


$15,5231.8%$5040 hrs.140

:::::

$8161.7%$5040 hrs.2

$1241.0%$5040 hrs.1

FailureCost

(Before)probability

LaborCostper

Hour

Resourcesper Day

DaysLate

� Compute the probability of each possible day late using theparameters from the fitted distributions

� Compute the daily failure cost: resource hours × labor rate� Weight the daily failure costs by the probability� Sum all the daily failure costs

Next, compute the failure costs…


The revised process cuts failure costalmost in half…

:::::::

$11,5811.4%$15,5231.8%$5040 hrs.20

0.0%

:

6.3%

4.8%

p(After)

$1,347$3,4730.1%$5040 hrs.140

$728K$1,226KCumulative Failure Cost

::::::

$5,027$8161.7%$5040 hrs.2

$1,908$1241.0%$5040 hrs.1

Overrun(After)

Overrun(Before)p(Before)

LaborCostper

Hour

Resourcesper Day

DaysLate

A net benefitof $500K


Summary…

� The specific benefit fromsqueezing variation out ofa process can becalculated using Cost ofQuality principles and SixSigma techniques

� Knowing the payoffmakes further quantitativemanagement compelling

squeezing variation for profit

Documents