software prediction models forecasting the costs of software development
TRANSCRIPT
Software Prediction Models
Forecasting the costs of software development
Prediction Study Outcomes Vary
Estimation-by-analogy beats regression Or not Classification and regression trees (CART)
beats regression Or not Artificial neural networks beat regression Or not
Why Are The Results Conflicting?
Poor data or research procedure Complex techniques may require expert
users; hence applications may vary Small sample size Measurement process that is flawed Selective use of differing parameters may
result in different rankings
Key Terms
Accuracy indicator– Some measure of a process– A summary statistic based on that measure
Leave-one-out cross-validation Arbitrary function approximator taxonomy
– Many-data versus sparse-data– Linear versus nonlinear– Supervised versus unsupervised
Reliability versus validity
Indicator 1: MMRE
Mean magnitude of relative error (MMRE) is an average where the MRE=|actual-prediction|/actual
Claimed advantages of MMRE– Compare across data sets*– Independent of units– Compare across differing prediction models*– Scale independence
*An hypothesis challenged by this paper
Indicator 2: MER
Magnitude of the error relative to the estimate (MER) is defined asMER = |actual-prediction|/prediction
Indicator 3: AR
The absolute residual (AR) is defined asAR = |actual-prediction|
Other Measures
Standard deviation (SD) Relative standard deviation (RSD) Log standard deviation (LSD) Balanced relative error (BRE) Inverted balanced relative error (IBRE)
Standard Deviation of Residuals, Denoted SD
zero is ˆ residuals
theofmean thebecause occurstion simplifica The 2.
StDevfor formula general theNOT is formula This 1.
:Notes
)1(
ˆ
SD 1
2
iii
n
iii
yyr
n
yy
Algebraic Simplification
1
ˆ
1
1)(
residuals. theofmean the and
residual therepresent ˆLet
1
2
1
2
1
2
n
yy
n
r
n
rrresidualsStDev
r
iyyr
n
iii
n
ii
n
ii
thiii
Relative Standard Deviation (RSD)
)1(
ˆ
RSD 1
2
n
xyyn
i i
ii
Log Standard Deviation (LSD)
?ˆlnln
and residual theof variance the
where
)1(
2 LSD
2
1
22
iii
i
n
ii
yye
e
n
e
Balanced Relative Error (BRE)
0ˆ,
ˆ)ˆ(
0ˆ,)ˆ(
BREyy
y
yy
yyy
yy
Inverted Balanced Relative Error (IBRE)
0ˆ,
ˆ)ˆ(
0ˆ,)ˆ(
IBREyy
y
yy
yyy
yy