AI Summary & Key Takeaways
AI InsightThese results indicate that, once the CH-DNM architecture and its Huber-loss fitness function are held fixed, the choice of metaheuristic optimizer has only a secondary, though non-negligible, effect on forecasting accuracy and robustness, with SCA and GA offering a modest but consistent advantage over ABC and PSO.
📌 3 Key Takeaways
Abstract
This study presents a comparative performance analysis of four population-based metaheuristic optimization algorithms - the Artificial Bee Colony (ABC), the Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and the Sine Cosine Algorithm (SCA) - used to train a single, fixed robust neural architecture for cryptocurrency time series forecasting. The architecture under study is a Cauchy-Huber Dendritic Neuron Model (CH-DNM), which integrates a heavy-tailed Cauchy cumulative distribution function (CDF) as the somatic activation of a Dendritic Neuron Model (DNM) with a Huber robust loss function, so that extreme residuals arising from market shocks are down-weighted rather than allowed to dominate the training gradient. Rather than comparing different network architectures, this study isolates the effect of the training algorithm itself: the CH-DNM structure, its synaptic weights, thresholds, and the Huber-loss fitness function are held identical across all four optimizers, so that any difference in forecasting performance can be attributed strictly to the search dynamics of the optimization algorithm. Each optimizer is evaluated on the highly volatile Bitcoin (BTC-USD) index under synthetically injected extreme shock scenarios. Three independently and randomly extracted BTC-USD log-return series are used, at two observation-window lengths (100 and 250 days) and two validation-test horizons (10 and 20 days), with artificial shocks of ten times the local maximum magnitude injected exclusively into the training portion of each series. Forecast accuracy is assessed using the Root Mean Square Error (RMSE) over the corresponding test horizon, and each optimizer is executed over 30 independent runs per configuration to account for its stochastic nature. Averaged across all series, window lengths, and test horizons, the Sine Cosine Algorithm (SCA) achieves the lowest mean and most stable (lowest standard deviation) test RMSE on the original series, closely followed by the Genetic Algorithm (GA), while ABC and PSO show marginally higher mean error and run-to-run variability. Once outliers are injected into the training data, the gap between optimizers narrows considerably, with all four algorithms achieving very close mean test RMSE, and SCA retaining a small but consistent edge in stability. These results indicate that, once the CH-DNM architecture and its Huber-loss fitness function are held fixed, the choice of metaheuristic optimizer has only a secondary, though non-negligible, effect on forecasting accuracy and robustness, with SCA and GA offering a modest but consistent advantage over ABC and PSO.