The research, conducted by Sana Hamiane and Youssef Ghanou of Moulay Ismail University of Meknes in Morocco together with Gabriella Casalino of the University of Bari Aldo Moro in Italy, combines a convolutional neural network with a long short-term memory network in a single forecasting pipeline. Each component plays a distinct role. The convolutional layers act as feature extractors, scanning windows of multivariable economic data to detect local patterns, such as simultaneous movements in exports, industrial production and inflation, that might signal an upcoming change in output. The LSTM layers then take over, using their gated memory cells to track how these patterns evolve across consecutive quarters and to decide which information from the past should be retained or forgotten when producing a prediction.

This division of labor matters because GDP time series carry two kinds of structure at once. Spatially, the many macroeconomic indicators feeding into the model form intricate cross-correlations that convolutional filters are well suited to capture. Temporally, the economy exhibits long-range dependencies, momentum and regime shifts that recurrent architectures like LSTM were explicitly designed to remember. Models that handle only one of these dimensions tend to leave predictive signal on the table. By stacking the two architectures, the hybrid model can, in principle, exploit both, and the new results suggest that this is exactly what happens in practice.

Before any training began, the researchers confronted a problem that plagues every multivariable forecasting effort: which indicators should be included? Feeding a neural network dozens of weakly related variables invites overfitting and noise. The team turned to the Pearson correlation coefficient, a classical statistical measure of linear association, to rank candidate macroeconomic variables by their correlation with GDP. Only the most relevant features survived this filter. The resulting indicator sets differed by country, reflecting the structure of each economy. For the United States, France, Italy, Spain and Germany the model drew on measures such as private and government consumption, imports and exports, gross fixed capital formation, manufacturing output, consumer price indices, exchange rates and long-term interest rates. For Morocco, the study used a detailed breakdown of sectoral value added, spanning agriculture, mining, construction, transport, finance and other industries.

Architecture alone, however, does not guarantee good forecasts. Deep learning models are notoriously sensitive to hyperparameters: the number of layers, the size of filters and hidden states, learning rates, batch sizes and dropout rates all shape how well a network generalizes. Manual tuning is slow and subjective, and grid searches become computationally explosive as the number of settings grows. The authors addressed this with Bayesian optimization, a sequential model-based search strategy that builds a probabilistic surrogate of the model’s performance and uses it to select the most promising hyperparameter configurations to test next. Instead of blindly exploring the space, the optimizer balances exploration of uncertain regions with exploitation of configurations already known to perform well, a technique that has become a standard tool for squeezing maximum accuracy out of neural networks.

The evaluation was unusually broad for a single study. Quarterly time series from seven countries were used, drawn from the Federal Reserve Bank of St. Louis database for the five advanced economies and from the Haut Commissariat au Plan for Morocco and the Central Bank of Nigeria for the African cases. This mix of large, mature economies, a middle-income economy and a major emerging market gave the model a demanding test: could one framework, adapted country by country, deliver reliable forecasts across radically different data environments? The authors benchmarked their hybrid against standalone LSTM, standalone CNN, a temporal convolutional network known as TCN, and the vector autoregressive model that remains a workhorse of macroeconomic forecasting.

The results were striking. Across all seven countries, the hybrid CNN–LSTM model achieved coefficients of determination between 0.81 and 0.97, meaning it explained the vast majority of the variance in quarterly GDP movements. Prediction errors were correspondingly low, with mean squared error ranging from 0.00017 to 0.014077, mean absolute error from 0.0117 to 0.0960, and mean absolute percentage error between 1.18 and 7.34 percent. Crucially, the model outperformed every benchmark it was compared against, including the individual deep learning models that shared its building blocks. The advantage was not merely a statistical fluke: the authors applied the Diebold–Mariano test, the standard procedure for comparing predictive accuracy between competing forecasts, and found that the hybrid’s improvements were statistically significant in several cases, particularly against the vector autoregressive and temporal convolutional baselines.

Why should the combination beat its parts? The answer likely lies in complementary inductive biases. A standalone CNN sees the input as a grid of values and excels at local pattern detection, but it has limited capacity to model long-term temporal dynamics. A standalone LSTM remembers sequences well but processes multivariable inputs without an explicit mechanism for detecting local interactions between indicators. The temporal convolutional network, though efficient, similarly emphasizes local receptive fields. The hybrid model, by contrast, routes raw multivariable history through convolutional feature extraction before the recurrent layers ever see it, handing the LSTM a cleaner, more informative representation of the recent past. Bayesian optimization then ensures that the architecture is neither too small to capture the dynamics nor so large that it memorizes noise.

The practical implications reach beyond academic leaderboards. Accurate quarterly GDP forecasts inform central bank decisions on interest rates, government budget planning, and private sector investment strategies, and the lag between the end of a quarter and the publication of official GDP estimates creates persistent demand for reliable predictions. A model with mean absolute percentage errors as low as roughly 1 percent in some countries offers a meaningful margin over conventional approaches, and the study’s multivariable design means it can incorporate the kind of high-frequency economic indicators that policymakers already monitor. The authors note that their findings highlight the potential of hybrid deep learning architectures to enhance the accuracy of quarterly GDP forecasting in particular, suggesting a template that could be extended to other macroeconomic aggregates such as inflation, unemployment or trade balances.

The study also illustrates a broader trend in computational economics. Over the past decade, researchers have applied ARIMA models, artificial neural networks, support vector machines and various deep learning algorithms to GDP prediction for countries including China, India, Indonesia, Bangladesh, Egypt and New Zealand, with mixed results. What distinguishes the new work is the systematic combination of feature selection, architectural hybridization and principled hyperparameter optimization, together with a rigorous multi-country evaluation and formal statistical testing of forecast differences. As national statistical agencies and financial institutions accumulate ever larger repositories of economic indicators, the message from this research is that the path to better macroeconomic forecasts may lie not in any single model, but in intelligently combining the strengths of several, and in letting an optimizer do the fine-tuning that human analysts cannot.

Subject of Research: Hybrid deep learning models for multivariable macroeconomic time series forecasting

Article Title: Hybrid CNN–LSTM model optimized by Bayesian method for forecasting gross domestic product based on multivariable data

Article References: Hamiane, S., Ghanou, Y., & Casalino, G. (2026). Hybrid CNN–LSTM model optimized by Bayesian method for forecasting gross domestic product based on multivariable data. International Journal of Data Science and Analytics, 22(1), Article 308. https://doi.org/10.1007/s41060-026-01303-6

Image Credits: AI Generated

DOI: 10.1007/s41060-026-01303-6

Keywords: GDP forecasting, CNN, LSTM, Bayesian optimization, deep learning, macroeconomic indicators, time series, econometrics, Pearson correlation, Diebold-Mariano test, multivariable data, temporal convolutional network