This study addresses the critical challenge of financial forecasting for small businesses, which often struggle with fluctuating demand, seasonal sales patterns, and tight profit margins. Accurate forecasting is essential for optimizing resources, improving profitability, and making data-driven decisions in a dynamic market. To enhance the accuracy and efficiency of forecasting models, this paper introduces a novel approach combining machine learning models with metaheuristic optimization algorithms. Specifically, the Dynamic Attention Recurrent (DAR) model optimized with Logarithmic Transformation (LogTrans) is evaluated at various stages. In the baseline evaluation, the DAR + LogTrans model demonstrated outstanding performance with an MSE of 0.00075, RMSE of 0.0274, and R-squared of 0.861, indicating its strong predictive capability. After applying optimization techniques, DAR + LogTrans achieved remarkable improvements, reaching an MSE of 1.88E-07, RMSE of 4.36E-04, and R-squared of 0.968, showcasing substantial gains in accuracy and generalization. The results emphasize the potential of metaheuristic optimization, such as the Whale Optimization Algorithm (WAO), Bat Algorithm (BA), and Particle Swarm Optimization (PSO), in improving model performance. These findings provide valuable insights for small business owners seeking to implement advanced forecasting models that can adapt to market fluctuations. The optimized models, particularly DAR + LogTrans, offer a powerful tool for improving decision-making, managing cash flow, and enhancing operational efficiency, with significant implications for the future of financial forecasting in small businesses.
Read MoreDoi: https://doi.org/10.54216/JSDGT.050201
Vol. 5 Issue. 2 PP. 01-29, (2025)
Accurate prediction of customer behavior remains a core methodological and operational challenge in retail economics and financial decision-making, particularly as institutions increasingly depend on data-driven forecasting systems to improve credit risk assessment, refine customer segmentation, and deliver targeted financial services in competitive and rapidly changing markets. In practice, the economic value of customer behavior prediction lies in its direct connection to profit maximization, loss minimization, and resource allocation efficiency: retailers seek to anticipate spending tendencies and product affinities to reduce marketing waste and optimize inventory, while financial institutions aim to infer creditworthiness and repayment capacity to reduce default exposure and enhance portfolio stability. Despite the demonstrated advantages of data-driven approaches, the predictive performance of advanced learning systems in such contexts is frequently constrained by the dual challenge of high-dimensional, heterogeneous feature spaces and the sensitivity of model outcomes to hyperparameter choices, often resulting in limited generalization, unstable convergence, or performance degradation when applied to unseen customer groups. To address these constraints, this study develops an integrated optimization framework that couples a high-capacity predictive model with a physics-inspired search mechanism, namely the Kirchhoff’s Law Algorithm (KLA), and employs it for automated hyperparameter optimization of an End-to-End Attention Long Short-Term Memory model (EALSTM), thereby reducing reliance on manual tuning and improving model reliability under financially meaningful data complexity. In addition to introducing the KLA-driven optimization pipeline, the study conducts a rigorous comparative evaluation against established state-of-the-art metaheuristic optimizers, including Particle Swarm Optimizer (PSO), Biogeography Based Optimizer (BBO), Whale Optimization Algorithm (WOA), Bat Algorithm (BA), Artificial Protozoa Optimizer (APO), Genetic Algorithm (GA), and Stochastic Fractal Search (SFS), enabling a systematic assessment of how different search dynamics influence predictive quality in customer analytics applications. Experimental evaluation is performed using an enhanced customer dataset that integrates demographic descriptors, behavioral spending indicators, and financially meaningful constructs—thereby better reflecting real-world decision environments where customer profiling depends on both consumption behavior and financial capacity—and the results demonstrate that the KLA + EALSTM configuration consistently achieves the strongest predictive performance across the full suite of regression metrics. Specifically, KLA + EALSTM attains a Mean Squared Error (MSE) of 3.60 × 10−7, a Root Mean Squared Error (RMSE) of 0.00728, a Mean Absolute Error (MAE) of 0.000372, a Mean Bias Error (MBE) of 0.000091, a correlation coefficient (r) of 0.972, a coefficient of determination (R2) of 0.969, a Relative RMSE (RRMSE) of 0.095, a Nash–Sutcliffe Efficiency (NSE) of 0.971, and a Willmott Index (WI) of 0.977, collectively indicating extremely low error magnitude, minimal systematic bias, strong explanatory power, and high agreement between predicted and observed outcomes, and representing a substantial improvement over the unoptimized EALSTM baseline. From an economic and financial viewpoint, these gains are practically consequential because they strengthen the reliability of predictive decision-support systems used for credit scoring, personalized marketing, customer valueassessment, and financially efficient resource allocation, where even small prediction errors can translate into measurable cost, risk, or revenue impacts. Overall, the findings provide strong empirical support for physics-inspired, non-parametric optimization as a robust mechanism for improving predictive accuracy, stability, and generalization in customer analytics, and they position KLA-based optimization as a scalable and methodologically efficient solution for next-generation retail and financial analytics systems operating under high-dimensional behavioral and financial data conditions.
Read MoreDoi: https://doi.org/10.54216/JSDGT.050202
Vol. 5 Issue. 2 PP. 30-58, (2025)
The growing availability of granular customer-level data has intensified the demand for accurate and robust predictive models in retail economics and consumer finance, particularly for forecasting financially relevant indicators such as savings capacity and credit-related measures, where prediction inaccuracies can lead to inefficient pricing strategies, misallocation of financial resources, and distorted risk assessments. Traditional statistical and econometric approaches often struggle to model the nonlinear and high-dimensional relationships inherent in such data, motivating the use of advanced deep learning techniques combined with intelligent optimization strategies. This study proposes an integrated economic and financial analytics framework that couples a sequence-to-sequence deep learning architecture (Sequence-to-Sequence, Seq2Seq) with state-of-the-art metaheuristic optimization algorithms for automated hyperparameter tuning, with particular emphasis on the Puma Optimizer–Seq2Seq (PO + Seq2Seq) configuration. The framework systematically evaluates multiple baseline deep learning models and enhances them through metaheuristic-driven optimization to address challenges related to convergence stability, generalization capability, and model sensitivity in customer-level financial prediction. Empirical analysis shows that the PO + Seq2Seq model consistently outperforms all baseline and alternative optimized configurations across all evaluation stages, achieving a Mean Squared Error of 2.05 × 10−5, Root Mean Squared Error of 4.52 × 10−3, Mean Absolute Error of 2.05 × 10−4, and a very small Mean Bias Error of 5.40 × 10−5, together with strong goodness-of-fit and efficiency indicators, including a correlation coefficient of 0.987, R2 of 0.983, Nash–Sutcliffe Efficiency of 0.986, and Willmott Index of 0.988. From an economic and financial perspective, these findings demonstrate that the proposed PO + Seq2Seq framework provides a reliable and scalable predictive tool for customer analytics, enabling more accurate assessment of financial behavior, improved customer segmentation, and enhanced decision support in retail finance and consumer-oriented financial systems, while highlighting the critical role of metaheuristic optimization in unlocking the full predictive potential of deep learning models for real-world economic applications.
Read MoreDoi: https://doi.org/10.54216/JSDGT.050203
Vol. 5 Issue. 2 PP. 59-86, (2025)
Accurate forecasting of gold prices remains a critical challenge in financial markets due to the nonlinear, nonstationary, and regime-dependent nature of commodity price dynamics, particularly for gold quoted against the US dollar (XAU/USD), which plays a central role as a safe-haven asset, inflation hedge, and portfolio diversifier. Motivated by the growing limitations of traditional econometric and manually tuned machine learning approaches in handling long-horizon, multi-timeframe financial data, this study proposes a robust forecasting framework that integrates deep learning with metaheuristic optimization. The main contribution of this work lies in the systematic combination of a Deep Pyramid Recurrent Neural Network (DPRNN) with advanced metaheuristic algorithms for automated hyperparameter optimization, with particular emphasis on Greylag Goose Optimization (GGO), alongside other state-of-the-art optimizers. Using historical XAU/USD data spanning from 2004 to February 2025 across multiple temporal resolutions, baseline model evaluation demonstrates that DPRNN outperforms other deep learning architectures prior to optimization, achieving a Mean Squared Error (MSE) of 0.0589, Root Mean Squared Error (RMSE) of 0.2426, and coefficient of determination (R2) of 0.79. Following optimization, the proposed GGO-optimized DPRNN framework yields a substantial performance enhancement, reducing the MSE to 2.05 × 10−5 and RMSE to 4.52 × 10−3, while simultaneously increasing the correlation coefficient to 0.987 and R2 to 0.983, with near-perfect agreement metrics reflected by a Nash–Sutcliffe Efficiency of 0.986 and Willmott Index of 0.988. These results confirm the effectiveness of GGO in navigating complex hyperparameter search spaces and significantly improving predictive accuracy and stability. From an economic and financial perspective, the findings underscore the practical value of metaheuristic-optimized deep learning models for enhancing gold price forecasting, supporting more informed investment decisions, improved risk management, and greater market efficiency in volatile and uncertain financial environments.
Read MoreDoi: https://doi.org/10.54216/JSDGT.050204
Vol. 5 Issue. 2 PP. 87-116, (2025)
