Mask-Aware AI Transformer Outperforms Benchmarks in Forecasting GDP From Incomplete Data

Forecasting gross domestic product has always been one of the hardest problems in economics, and a new study argues that the missing piece may literally be the missing data. A research team led by Asiye Omidian and Maral Mirzamohammadi has built a deep-learning architecture called a mask-aware Multivariate Correlation Transformer, or MCT, that is explicitly designed to forecast GDP from short, heterogeneous, and incomplete annual macroeconomic histories. In a fixed-origin evaluation spanning 2020 to 2024 across 14 countries, the model achieved a root mean squared error of 0.141693 trillion current U.S. dollars, roughly 23.8 percent lower than the strongest original comparator, an encoder–decoder LSTM, and about 15.8 percent lower than iTransformer, the best of the modern forecasting architectures tested. The work, published in Discover Artificial Intelligence, addresses a problem that has long frustrated both statisticians and machine-learning researchers: how to squeeze reliable predictions from economic panels that are riddled with gaps.

The core difficulty is structural. GDP evolves jointly with labor-market conditions, investment, trade, prices, exchange rates, and fiscal variables, and these relationships shift across countries and over time. Recent shocks, including the COVID-19 pandemic, inflationary surges, supply-chain disruptions, and large exchange-rate movements, have only sharpened the instability and nonlinearity of macroeconomic dynamics. Meanwhile, cross-country economic databases are rarely balanced. Historical coverage differs across variables, countries, and sources; some country-year observations are simply absent. The conventional responses, dropping incomplete variables or filling gaps with imputed values, either discard useful information or blur the crucial distinction between what was actually observed and what was manufactured by an algorithm. The MCT takes a different route: it keeps the missingness visible throughout the entire forecasting pipeline.

Technically, the framework rests on four integrated components. First, an explicit binary observation mask accompanies every input, telling the network which of the entries in a 10-year by 14-variable window are genuine measurements and which are unavailable. Second, a reconstruction pathway is trained jointly with the forecaster, but its supervision is applied only where verified ground truth exists, namely originally observed values that are deliberately hidden during training. Naturally unavailable entries are never assigned unverified reconstruction targets. Third, cross-variable interactions are modeled through an asymmetric cosine-normalized query–key similarity attention mechanism rather than the conventional scaled dot-product attention used in standard Transformers. Because the query and key projections are learned independently, the resulting similarity matrix is directional, and the authors are careful to stress that these are internal neural similarities, not statistical correlations or causal relationships. Fourth, an input-adaptive dynamic feedforward module blends three learnable transformation kernels with weights generated from a global summary of the current representation, while a similarity-structure self-distillation loss encourages the forecasting pathway to preserve the cross-variable interaction structure learned by the reconstruction pathway.

The empirical design is deliberately restrictive, which is part of what makes the results notable. The dataset assembles 14 macroeconomic variables, from inflation and unemployment to capital stock and the human-capital index, for 14 countries including the United States, China, India, Brazil, and Nigeria, drawing on the IMF World Economic Outlook, the IMF International Financial Statistics, and the Penn World Table. The target is nominal GDP at current prices in U.S. dollars. Neural models are trained globally on pooled country-specific 10-year input windows, with country-balanced sampling so that economies with longer histories do not dominate optimization. No country identifier or embedding is supplied; the model must learn shared dynamics across very different economies. Training covers 1980 to 2015, validation runs from 2016 to 2019, and testing is reserved for 2020 through 2024, a period chosen, knowingly, to include some of the most turbulent macroeconomic conditions in decades.

Evaluation proceeds from a single forecast origin at the end of 2019. The 2020 forecast is generated directly, while forecasts for 2021 through 2024 are produced recursively, with each model’s own predictions appended to the rolling input window and no realized test-period outcomes fed back into the forecasting path. This fixed-origin recursive protocol yields 70 forecast errors per model, 14 countries times five horizons, and the authors treat those errors as the dependent panel they are. Rather than assuming independence, the primary statistical comparisons use standard errors clustered simultaneously by country and target year, with Holm adjustment across the seven original benchmarks. All pairwise squared-loss improvements favor the MCT, with Holm-adjusted p-values below 0.05, and two complementary sensitivity analyses, target-year-block inference and a 10,000-replicate country-cluster bootstrap, reach the same qualitative conclusion.

The benchmark sweep is unusually broad. Beyond the original statistical models, exponential smoothing, ARIMA, and vector autoregression, and recurrent neural networks including CNN, LSTM, bidirectional LSTM, and encoder–decoder LSTM, the expanded comparison adds five simple rules such as persistence and random walk with drift, five contemporary architectures including PatchTST, TimesNet, Crossformer, and TiDE, and six conventional machine-learning methods including CatBoost, XGBoost, and Random Forest. None reproduces the MCT’s point-forecast accuracy. Among simple rules, the best performer, an AR(1) model fitted to log GDP growth, records an RMSE of 0.536350, nearly four times the MCT figure. CatBoost leads the machine-learning family at 0.239600. The MCT also posts the lowest Theil’s U, at 0.159992, indicating the largest relative reduction in error compared with a no-change persistence reference.

Ablation experiments probe which components matter. Replacing the complete MCT block with a standard Transformer produces the largest deterioration, a 15.60 percent increase in RMSE, followed by swapping the normalized-similarity attention for conventional dot-product attention, which costs 12.42 percent. Removing the reconstruction pathway raises error by 8.68 percent, while dropping the dynamic feedforward module or the self-distillation objective costs 6.77 and 5.50 percent respectively. The authors caution that some of these comparisons also change parameter counts, so the percentages should not be read as isolated causal effect sizes. Robustness was further tested under additional random corruption at 10, 30, and 50 percent masking rates and under six structured synthetic missingness scenarios, including block, ragged-edge, variable-specific, country-specific, and MAR-like and MNAR-like mechanisms. The MCT retained the lowest RMSE in every condition, with its advantage over a standard Transformer widening to roughly 27 percent under the heaviest random corruption.

The reconstruction audit offers a rare moment of verified honesty in the imputation literature. Under a fixed 25 percent audit mask, the model recovered 1,591 deliberately withheld observations with an RMSE of 0.062474 and a pooled R-squared of 0.958807, results that are consistent across countries. But no accuracy claim is made for the 383 naturally unavailable entries, because no verified ground truth exists for them. Attention diagnostics reinforce the theme of careful interpretation. The learned similarity matrices are substantially asymmetric, with a mean maximum asymmetry of 0.6966, confirming they are not Pearson-type correlation matrices. Descriptive attention profiles are dispersed rather than concentrated: exports, GDP, employment, exchange rates, and capital receive the highest mean incoming attention, with stable rankings across five random seeds, but the authors explicitly reject any causal reading of these weights.

The study is equally candid about its limits. The evidence rests on a single forecast origin, five target years, 14 countries, and annual data, and the evaluation is retrospective rather than a genuine real-time-vintage experiment. The 2025 to 2030 outputs, which project, for example, U.S. nominal GDP rising from 28.720 to 32.270 trillion dollars and China’s from 18.995 to 26.020 trillion, are reported separately as post-evaluation model-based projections with empirically calibrated uncertainty bands, not as validated forecasts. The model is also deterministic, so no formal probabilistic prediction intervals are offered, and its accuracy comes at a computational price: more parameters, more memory, and longer training than simpler benchmarks, even though its estimated operation count per sample is about 60 percent lower than PatchTST’s. Still, the central message stands on its own terms. By jointly modeling what is observed, what is missing, and how fourteen macroeconomic variables move together, the MCT demonstrates that treating missingness as information rather than nuisance can measurably improve GDP forecasting in exactly the settings, short, incomplete, cross-country annual data, where forecasters need help most.

Subject of Research: Deep-learning-based GDP forecasting from incomplete multivariate macroeconomic time series

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

Article References: Omidian, A., Keivan, Z., Omidian, A., & Mirzamohammadi, M. (2026). A mask-aware multivariate correlation transformer for GDP forecasting with incomplete macroeconomic time series. Discover Artificial Intelligence, 6(1), Article 1454. https://doi.org/10.1007/s44163-026-02538-9

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02538-9

Keywords: GDP forecasting, transformer, missing data, macroeconomic time series, attention mechanism, machine learning, multivariate forecasting, reconstruction, self-distillation, IMF data, Penn World Table, econometrics