Smarter, faster, lighter: Deep RVFL networks for predicting Sri Lankan stock market trends

Abstract

Stock market price prediction is important in financial analysis and investment planning. In this study, a Deep Random Vector Functional Link (Deep RVFL) neural network was used to predict the next-day closing prices of the S&P Sri Lanka 20 Index (S&P20) with daily data from January 2010 to July 2025. Recent research has shown growing interest in randomised neural networks for financial forecasting. RVFL models are recognised for their fast training, simplicity, and ability to capture complex patterns without iterative optimisation. Past studies have examined RVFL models for stock price prediction and, in many cases, found them to perform better than elaborated deep learning frameworks. In this study, a Deep RVFL was implemented with three hidden layers, each containing 64 nodes and ReLU activation functions. The architecture includes direct input-output connections to strengthen representational power. In total, 8,769 trainable parameters were tuned using the Adam optimiser over 30 epochs. Before training, the dataset was standardised to support stable convergence and enhance predictive performance. On the test dataset, the model achieved a Mean Squared Error (MSE) of 0.0047 and a Mean Absolute Error (MAE) of 0.0442. These values suggest that the Deep RVFL can deliver precise next-day forecasts with relatively low error. The training history suggested steady convergence and no noticeable overfitting, which reflects the model’s stability. All together, these results indicate that Deep RVFL can serve as a practical tool for short-term stock price prediction in emerging markets. Due to its straightforward implementation, efficiency, and competitive performance, it could be a practical choice compared to more complex deep learning methods.