The performance of Long Short-Term Memory (LSTM) networks in financial time series forecasting is heavily dependent on two interdependent factors: the selection of input features and the configuration of network hyperparameters. Existing methodologies treat these as isolated sequential problems, optimizing one while holding the other fixed, which guarantees a suboptimal convergence. This paper introduces a Multi-Objective Modified Firefly Algorithm (MOMFA) that executes joint feature selection and hyperparameter optimization as a single unified search. Candidate solutions are encoded as 30-dimensional vectors combining a discrete binary feature mask with continuous hyperparameter spaces. The standard Firefly Algorithm is modified via Lévy flight exploration and a sigmoid transfer function to bridge the continuous and discrete search spaces. Optimizing for both forecast accuracy (RMSE) and model complexity, the framework generates a Pareto-optimal set of network configurations.
Udayanga et al. (Tue,) studied this question.