When Does Combining Software-Effort Estimates Help? Aggregate Benchmark Results from Four Public Project Datasets

Aggregate benchmark results accompanying the research manuscript "When Does Combining Software-Effort Estimates Help? Evidence from Four Public Project Datasets" by Azzeddine Ihsine and Sara Ihsine. The study investigates when combining software-effort estimates improves or degrades predictive accuracy, particularly in the presence of correlated estimation errors and unequal estimator reliability. Five forecast-combination methods are evaluated: equal weighting, inverse-MSE weighting, nonnegative simplex regression, Bates-Granger weighting, and Bates-Granger weighting with Ledoit-Wolf shrinkage. These methods are compared with a calibration-selected individual estimator across four public software-effort datasets: Kitchenham, Desharnais, China, and Maxwell. The experimental protocol includes 100 repeated 70/30 training and test partitions per dataset, together with a completion-year-based sensitivity analysis on Desharnais. All preprocessing is learned within training folds to prevent information leakage. The public archive contains: * Mean and standard deviation of held-out log-MAE and log-RMSE by dataset and method. * Calibration-error correlation summaries. * Win, tie, and loss counts relative to the individual baseline. * Negative-weight counts for covariance-aware combination methods. * Nadeau-Bengio corrected resampled t-test results with Holm adjustment. * Results from the Desharnais time-ordered evaluation. * The scientific tables and figures accompanying the manuscript. * Data provenance, reproducibility limitations, and file-integrity checksums. The results show that forecast combination is not universally beneficial. In Kitchenham, an accurate historical human estimate has a lower mean error than every combination, and equal weighting produces the only statistically significant degradation. In Desharnais, China, and Maxwell, combinations improve average accuracy, but these gains are not statistically significant under the corrected analysis. This public release contains aggregate experimental results only. It does not include raw project records, individual prediction errors, complete split-level results, experimental source code, or proprietary D-POAF platform implementations. Detailed verification materials, including experimental scripts, split assignments, and split-level results, are maintained separately. Access may be requested by qualified reviewers through the corresponding author, subject to applicable data rights and access conditions. The original public datasets were retrieved from the DASE repository (commit 37945fb), whose referenced dataset files are distributed under CC0 1.0. This dataset supports empirical study within the broader D-POAF research programme.