A mask-aware multivariate correlation transformer for GDP forecasting with incomplete macroeconomic time series

Accurate Gross Domestic Product (GDP) forecasting is challenging when macroeconomic histories are multivariate, heterogeneous, short, and incomplete. This study proposes a mask-aware Multivariate Correlation Transformer (MCT) that integrates explicit observation masks, supervised reconstruction where ground truth is available, asymmetric cosine-normalized query–key attention, dynamic feedforward processing, and similarity-structure self-distillation. The analysis uses annual data for 14 countries from the IMF World Economic Outlook, IMF International Financial Statistics, and Penn World Table, with neural models trained globally on pooled 10-year input windows and one-year multivariate targets. Forecast performance is evaluated from a common origin at the end of 2019, with H1 generated directly for 2020 and H2–H5 generated recursively through 2024 without incorporating realized test-period outcomes or re-estimating model parameters. The benchmark set includes statistical and recurrent models, persistence/trend/growth rules, contemporary forecasting architectures, and conventional machine-learning methods. Using equal-country aggregation, the Proposed MCT achieves a root mean squared error (RMSE) of 0.141693, a mean absolute error (MAE) of 0.137203, a symmetric mean absolute percentage error (sMAPE) of 11.6653%, and Theil’s U of 0.159992. The strongest additional point-forecast comparator is iTransformer with RMSE = 0.168200. Dependence-aware inference against the seven original benchmarks, repeated-seed experiments, ablations, and random and structured missingness stress tests provide complementary evidence for the observed performance. Reconstruction accuracy is verified only for deliberately withheld observations with known ground truth, while learned attention scores are interpreted as internal neural similarities rather than statistical correlations or causal relationships. The evidence remains specific to the single-origin 2020–2024 evaluation. The 2025–2030 outputs are reported separately as post-evaluation model-based projections rather than additional validated forecasts.