Comparative Analysis of LSTM-Based Models for Daily Gold Price Forecasting Using Time Series and Sentiment Features

Abstract

This study examines the performance of three forecasting approaches for predicting daily gold prices: a Pure Long Short-Term Memory (LSTM) model, a Hybrid LSTM + Prophet model, and an LSTM model enhanced with sentiment features. Daily price data from 2013 to 2023 were used, and all evaluation metrics were computed on inverse-transformed values to ensure accurate interpretation. The Pure LSTM model shows strong predictive capability, achieving a testing RMSE of 17.4500, MAE of 12.9924, R² of 0.9431, and a MAPE of 0.72%. The Hybrid LSTM + Prophet model performs considerably worse on unseen data, with a testing RMSE of 71.6272 and an R² close to zero, indicating that Prophet’s decomposition is unsuitable for highly volatile price movements. The LSTM + Sentiment model provides the best overall performance, with a testing RMSE of 17.2769, MAE of 12.5515, R² of 0.9442, and a MAPE of 0.70%, showing modest but consistent improvement over the baseline LSTM. These findings highlight the suitability of LSTM-based models for non-seasonal financial time series and demonstrate that sentiment can contribute additional predictive value.