集合预报
平均绝对百分比误差
一般化
集成学习
贝叶斯概率
计量经济学
对比度(视觉)
计算机科学
均方误差
统计
时间序列
系列(地层学)
数据挖掘
贝叶斯推理
概率预测
人工神经网络
滑动窗口协议
人工智能
平均绝对误差
数学
均方预测误差
交叉验证
机器学习
自回归积分移动平均
预测建模
作者
Irdayanti,Khairil Anwar Notodiputro,Sachnaz Desta Oktarina
标识
DOI:10.15294/sji.v12i4.32225
摘要
Purpose: There are several forecasting methods, including SARIMA, LSTM, and GRU, which are often claimed to exhibit strong performance in capturing patterns in time series data. However, few studies have conducted direct comparisons among these methods. Therefore, it is necessary to conduct a performance evaluation using empirical data, particularly nickel prices data. This study also aims to improve forecasting performance by combining prediction outputs from deep learning-based models. Methods: This study utilized data on monthly global nickel prices from January 1990 to May 2025. The models developed include SARIMA, LSTM, GRU, and two ensemble approaches: Weighted Averaging and Bayesian Model Averaging (BMA). Model validation was conducted using walk-forward validation with a sliding window approach to evaluate each model’s generalization performance on out-of-sample validation data. The performance was evaluated using MAPE, RMSE, and MAE. Result: The BMA Ensemble approach shows the best performance in forecasting nickel prices, with a MAPE value of 5.39%, RMSE of 1897.84, and MAE of 1133.96. Prediction validation produces MAPE values below 10%, which indicates that the forecasting results are accurate. The ensemble BMA approach is able to produce more accurate and stable predictions compared to other models. Novelty: This study offers a novel approach combining LSTM and GRU through ensemble methods to forecast global nickel prices using monthly historical data from 1990 to 2025. In contrast to previous studies that relied on single models, the proposed method with the ensemble BMA approach demonstrates improved forecasting accuracy and stability.
科研通智能强力驱动
Strongly Powered by AbleSci AI