A Deep Learning-Based Workforce Demand Forecasting Framework with Skill-Gap Prediction for Intelligent Human Resource Planning

Abstract

Organizations increasingly face a mismatch between the pace at which occupational demand shifts and the pace at which internal workforce skills evolve, a gap that traditional, backward-looking human resource (HR) planning practices are poorly equipped to anticipate. Prior research has largely treated workforce demand forecasting and occupational skill analysis as separate problems: time-series and machine-learning studies forecast aggregate employment or job-posting volume without linking predictions to skill content, while skill-taxonomy and natural-language-processing (NLP) studies characterize current skill requirements without a temporal forecasting component. This separation limits the practical value of either line of work for proactive recruitment, reskilling, and upskilling decisions. This paper proposes DL-WSP, an integrated Deep Learning-based Workforce and Skill Prediction framework that couples a temporal deep-learning demand-forecasting module with a skill-gap estimation module, formalized through an explicit mathematical model relating projected role demand, required future skill proficiency, and currently available workforce skill capacity. The framework is evaluated on real, publicly available United States labor-market data: a 1,052-month (1939–2026) Bureau of Labor Statistics Current Employment Statistics series for historical demand, and the Bureau of Labor Statistics Employment Projections skills data product (2025–2035 vintage, built on O*NET) for occupation-level skill importance and projected employment change. A chronologically valid comparison of statistical (ARIMA), machine-learning (Random Forest, XGBoost, linear regression), and deep-learning (LSTM, GRU, BiLSTM) forecasters was executed under identical conditions across four forecast horizons (1, 3, 6, and 12 months). Contrary to prior expectations drawn from adjacent forecasting domains, a simple linear model with engineered lag features outperformed every deep-learning architecture at every horizon on this real series, while tree-based ensembles underperformed even a basic ARIMA baseline — a finding traced to a real, mechanistically identifiable extrapolation limitation of tree ensembles on strongly trending data. A separate, genuine correlation analysis across 831 real occupations found problem-solving, critical/analytical thinking, and science skills most positively associated with real projected 2025–2035 employment growth, independently replicating the direction of BLS’s own published analysis on an earlier data vintage. The full forecast-to-skill-gap computation could not be closed end-to-end on real data within this study due to the unavailability of a multi-industry numeric panel and an organizational skill-availability dataset, a limitation reported explicitly rather than filled with assumed values. The primary novelty of DL-WSP lies in its explicit, auditable linkage between a temporal forecasting output and a skill-gap decision layer, and in its transparent, evidence-based finding that architectural complexity does not automatically improve workforce-forecasting accuracy — a caution directly relevant to HR analytics practitioners considering which methods to deploy.