Table 2

The DNN architecture used in this study

HyperparameterAppliedJustification
Normalisation methodStandardScalerEnsures all features have comparable scales, improving convergence and numerical stability
Hidden layers3 fully connected layers (128, 64, 32 neurons)Provides sufficient capacity to model nonlinear temporal and regional price dependencies without excessive complexity
Loss functionMSEAppropriate for continuous-valued regression and forecasting tasks
OptimiserAdam (learning rate = 0.001)Adaptive optimiser, robust for non-stationary and noisy time series
Batch size4Small batch sizes improve generalisation for small-to-medium datasets
Epochs100Allows convergence while monitoring for overfitting, especially with small datasets
Input activationLinearPreserves scaled and PCA-transformed features without distortion
Hidden activationReLUEfficient, avoids vanishing gradients, captures nonlinear dependencies
Output activationLinearSuitable for continuous-valued forecasting
Dropout rate0.2Prevents overfitting by randomly deactivating 20% of neurons during training
Batch NormalisationEnabledStabilises learning, accelerates convergence, reduces internal covariate shift
Early stoppingEnabledPrevents overfitting and avoids unnecessary computation once validation loss plateaus
Data partitioning methodHoldoutPartition the dataset into 70% for training, 10% for validation, and 20% for testing
Source(s): Authors

or Create an Account

Close subscription notice
Close access options