As part of the ongoing research and development of the predictive models we regularly measure the model accuracy on real delivered software projects. 

 

Model/Module	Number of survey questions	Module supports tradeoffs and cost analyses	Relative error when predicting escaped defect density
Basic Historical model	None.  Historical failure data must be collected from your own organization.	NO	If data is recent, complete and similar in application type this is generally the most accurate predictor
Basic SEI CMMi model	One question	NO	If level <2 - 83% if level >= 2 19%
Basic Industry type	One question	NO	154%
Shortcut model	15 questions	YES	See below chart
Full-scale model	Between 60 and 120 questions depending on how you answer initial questions	YES	See below chart
Rome Labs model	From 44 to several hundred depending on how many surveys you answer	YES	Depends on selection of application type factor and number of questions answered
GUESSING	None	NO	Guessing during the early phases of development generally results in a relative error of 400% or more.  Guessing gets better as code becomes more complete.