Software effort models should be assessed via leave-one-out validation

Context: More than half the literature on software effort estimation (SEE) focuses on model comparisons. Each of those requires a sampling method (SM) to generate the train and test sets. Different authors use different SMs such as leave-one-out (LOO), 3Way and 10Way cross-validation. While LOO is a deterministic algorithm, the N-way methods use random selection to build their train and test sets. This introduces the problem of conclusion instability where different authors rank effort estimators in different ways. Objective: To reduce conclusion instability by removing the effects of a sampling method's random test case generation. Method: Calculate bias and variance (B&V) values following the assumption that a learner trained on the whole dataset is taken as the true model; then demonstrate that the MAT and runtime values for LOO are similar to N-way by running 90 different algorithms on 20 different SEE datasets. For each algorithm, collect runtimes, B&V values under LOO, 3Way and 10Way. Results: We observed that: (1) the majority of the algorithms have statistically indistinguishable B&V values under different SMs and (2) different SMs have similar run times. Conclusion: In terms of their generated B&V values and runtimes, there is no reason to prefer N-way over LOO. In terms of reproducibility, LOO removes one cause of conclusion instability (the random selection of train and test sets). Therefore, we depreciate N-way and endorse LOO validation for assessing effort models. (C) 2013 Elsevier Inc. All rights reserved.