Robust Time Series Forecasting in Cryptocurrency Markets- An Artificial Bee Colony Optimized Cauchy-Huber Dendritic Neural Network for Bitcoin

This study proposes a novel, robust artificial neural network architecture designed to mitigate the adverse effects of extreme market shocks and outliers in cryptocurrency time series forecasting. The proposed hybrid model 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. Unlike conventional models that suffer from training data poisoning when trained with the Mean Squared Error (MSE) loss, the proposed Cauchy-Huber Dendritic Neuron Model (CH-DNM) down-weights the influence of extreme residuals during training. The model's synaptic weights and thresholds are optimized using the Artificial Bee Colony (ABC) algorithm, which evaluates the fitness of candidate solutions by minimizing the Huber loss rather than the standard MSE. This swarm-intelligence metaheuristic exploits employed-bee, onlooker-bee, and scout-bee search mechanisms to avoid premature convergence to local minima while robustly filtering out extreme residuals. The forecasting performance of the proposed architecture is evaluated on the highly volatile Bitcoin (BTC-USD) index under synthetically injected extreme shock scenarios. Randomly extracted 250-day segments of the daily log-return series are used, 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 a fixed 20-day validation-test horizon. To account for the stochastic nature of ABC-based training, all neural network models are executed over 30 independent runs. Experimental results show that, while all six models perform comparably on the uncontaminated series, the proposed CH-DNM achieves the lowest mean and most stable (lowest standard deviation) test RMSE among all six architectures once outliers are injected into the training data, confirming that the combination of a heavy-tailed activation function and a Huber loss provides an effective defence against training data poisoning.