Artificial Intelligence for Financial Budget Predictions: Neural Networks Lead the Way

Abstract

Predicting financial budgets remains an open challenge. Most studies focus on high-frequency markets, such as stock price prediction, leaving the budget domain aside. This work addresses that gap. We evaluate four artificial intelligence techniques for budget prediction: decision trees, random forests, linear regression, and multilayer perceptron. Additionally, we implement a hybrid version, MLP_GA, where the network hyperparameters are optimized using a genetic algorithm. For training and validation, we use two original datasets not publicly available: quarterly financial statements from Mexican entities (2010-2024) and monthly financial statements from Ecuadorian organizations (2019-2022). Both present characteristics typical of emerging economies, including episodes of high inflation, currency devaluations, fiscal policy changes, and the COVID-19 crisis.  The results show that MLP_GA achieves superior predictive performance compared to the standard multilayer perceptron and the other techniques. Genetic algorithm optimization enables exploring multimodal and non-convex error surfaces, finding hyperparameter configurations (number of hidden neurons, learning rate, L2 regularization coefficient, maximum iterations, and early stopping tolerance) that significantly reduce overfitting, a critical problem when dealing with short time series of only forty to sixty quarterly observations. Statistical tests confirm that the differences are significant.  The contribution is threefold. First, regarding the application domain, we present the first systematic and comparative evaluation of hybrid models combining genetic algorithms with multilayer perceptrons specifically applied to financial budget prediction, distinguishing ourselves from existing studies that focus almost exclusively on high-frequency stock market prediction. Second, regarding the geographical context, we validate our models in two Latin American economies that the literature has systematically ignored: Mexico and Ecuador, a region representing less than one percent of datasets used in the literature according to recent meta-analyses. Third, regarding the methodological justification, we demonstrate that genetic   optimization is particularly effective for moderate-sized data, where more complex architectures like LSTMs or transformers suffer from overfitting due to their high parameter count. The combination of monthly and quarterly analysis improves model robustness, allowing accurate performance in both short-term scenarios and broader projections. These findings suggest that the MLP_GA approach can improve financial budgeting in data-constrained contexts. The paper concludes with a discussion of limitations and future research lines.