An Empirical Comparative Analysis of LSTM-Based Time Series Models for Skill Gap Forecasting

Ongoing skills mismatches between available talent pools and changing business needs remain an important obstacle to organisational productivity and country level economic competitiveness. Predicting the locations of emerging competency gaps that may eventually translate into business process failures requires tools that can effectively extract patterns of sequence demand over time. In this paper, we report the results of an empirical investigation into the performance of seven sequence architectures, namely, Vanilla LSTM, Stacked LSTM, Bidirectional LSTM, CNN-LSTM, LSTM with Bahdanau Attention, GRU, and Transformer, on four synthetic demand sequences over ninety six months for the Technology, Healthcare, Finance, and Manufacturing industries. Using identical evaluation protocols for all architectures, we show that the LSTM with Bahdanau Attention architecture yields the minimum forecasting error for all four metrics: RMSE = 0.0643, MAE = 0.0472, MAPE = 5.09%, R2 = 0.9142. Notably, the LSTM with Bahdanau Attention architecture improves the RMSE error over the baseline Vanilla LSTM architecture by 21.9%. Using the Wilcoxon signed-rank test, we validate the statistical significance of the pairwise differences between architectures. Moreover, we report that the standard deviations of the architectures over five separate training runs remain below 0.001 for all architectures. These results offer evidence-based, metric-driven recommendations for the construction of the next generation of workforce intelligence forecasting systems.