Date of Publication : September 2026

Abstract: Every major IT program delivery organization collects sprint velocity, commitment reliability, constraint removal rate, and story spillover data through its Agile execution toolchain. Yet this granular behavioral record of how work is actually being done has never been systematically integrated with portfolio-level schedule variance data to train a predictive model for program delay risk. This paper introduces LEXI (Lean Execution Intelligence), the first validated framework for fusing lean Agile execution metrics with structural investment performance features in a supervised machine learning pipeline to forecast schedule delay risk in multiproject IT programs. LEXI trains an XGBoost classifier on a hybrid feature set combining ten lean execution indicators derived from sprint and Program Increment records with fourteen traditional structural and schedule performance variables drawn from the federal IT investment portfolio. A formal comparative experiment demonstrates that LEXI achieves a ROC AUC of 0.938 and a recall of 0.913 on held-out test data, representing improvements of 5.6 and 6.5 percentage points, respectively, over an identical XGBoost model trained without lean execution features. The SHAP global attribution analysis reveals that commitment reliability ratio, constraint removal velocity, and unresolved dependency age are the three lean execution features that contribute most to prediction accuracy, collectively accounting for 31.4 percent of mean absolute feature importance across the full hybrid feature set. Partial Dependence Plot analysis uncovers a nonlinear threshold effect: commitment reliability ratios below 0.72 generate disproportionately large delay risk probability increases that are invisible to traditional schedule variance monitoring until two to four reporting cycles later. These findings establish lean execution metrics as a genuinely informative and previously untapped signal class for IT program delay risk forecasting and provide program delivery leaders with empirically grounded evidence that the behavioral data their Agile toolchains already generate contains specific, actionable predictors of program-level schedule failure.

Reference :