Decile Mean-Based Artificial Neural Network

Multilayer Perceptron (MLP) models have long served as a cornerstone in time series forecasting due to their ability to model complex non-linear relationships. However, their forecasting performance degrades significantly in the presence of anomalies, as standard architectures rely heavily on mean-based aggregation mechanisms that are inherently sensitive to outliers. While various robust neural architectures have been developed to mitigate this vulnerability—often utilizing median-based structures—they frequently suffer from a loss of statistical efficiency in clean data environments or remain susceptible to specific, asymmetric contamination patterns. To bridge this gap between robustness and statistical efficiency, this study introduces a novel neural network architecture: the Decile Mean Artificial Neural Network (DM-ANN). Distinct from conventional networks, DM-ANN implements a specialized decile mean-based aggregation function within its hidden nodes. By trimming extreme values and focusing on central deciles, this mechanism successfully strikes a balance, offering the robust error-insensitivity typically associated with the median alongside the mathematical efficiency of the mean. Due to the non-differentiable nature of the decile-based aggregation function, traditional gradient descent algorithms (such as backpropagation) are sub-optimal for training this architecture. Consequently, the DM-ANN is optimized using the Artificial Bee Colony (ABC) algorithm, a powerful metaheuristic population-based optimization technique capable of navigating complex, non-smooth search spaces to find global optima. The proposed model's forecasting performance and resilience are rigorously evaluated using real-world financial time series datasets under both pristine (clean) and artificially contaminated scenarios with varying levels of outlier intensity. The experimental results demonstrate that DM-ANN consistently outperforms traditional MLPs and existing robust alternatives. It achieves superior forecasting accuracy in stable markets while exhibiting remarkable stability and robustness against heavy-tailed noise and sudden structural breaks, establishing itself as a highly reliable tool for volatile financial forecasting