AG-BiLSTM: A Hybrid Deep Learning Model for Olympic Medal Prediction

Abstract

Time-series forecasting is an important research area in computer science and has been widely applied to complex prediction tasks. Olympic medal prediction is a typical multivariate time-series forecasting problem that supports national sports planning and resource allocation. However, existing methods often struggle to capture nonlinear temporal patterns and long-range dependencies, particularly when historical data are limited. To address these challenges, this study proposes Attention Gated Bidirectional Long Short-Term Memory (AG-BiLSTM), a hybrid deep learning framework that integrates a Bidirectional Long Short-Term Memory (BiLSTM) network, a multi-head self-attention mechanism, and a gated fusion module. BiLSTM is employed to learn bidirectional temporal dependencies from historical Olympic data, while the attention mechanism identifies critical time steps. The gated fusion module further enhances prediction performance by adaptively weighting heterogeneous features. In addition, a country-level temporal splitting strategy is adopted to improve model generalization across countries with different data distributions. Experiments on a self-constructed dataset covering 19 countries demonstrate that AG-BiLSTM achieves an average Root Mean Square Error (RMSE) of 3.74 and consistently outperforms ARIMA, XGBoost, LSTM, BiLSTM, Transformer, and iTransformer models. The proposed model shows superior predictive accuracy across gold, silver, and bronze medal forecasting tasks, with particularly strong performance in gold medal prediction. These results indicate that AG-BiLSTM is an effective and robust approach for Olympic medal prediction under small-sample conditions.