%0 Journal Article
%T 基于LSTM、Transformer和LightGBM的机构备付金预测方法<br>A Prediction Method for Institutional Reserve Based on LSTM, Transformer, and LightGBM
%A 冀乃庚
%A 周昕博
%J Computer Science and Application
%P 249-259
%@ 2161-881X
%D 2024
%I Hans Publishing
%R 10.12677/CSA.2024.142025
%X <span style=\"font-size:14px;\">机构备付金是金融机构的重要指标之一，对于评估其稳定性和偿付能力具有重要意义。在第三方支付机构备付金集中存管的背景下，准确预测支付机构备付金的变动对于监管机构风险管理等方面具有重要价值。笔者提出了一种基于LSTM、Transformer和LightGBM的机构备付金预测模型。利用树模型针对表格数据的快速性和准确性，选取交易日志的关键特征；利用Transformer的全局上下文建模能力捕捉财务文件的局部特征；最后采用LSTM算法获取结合后的数据的长期依赖关系。实验结果表明：该模型在机构备付金方面的预测准确性优于ARMA算法、LSTM算法和时序预测Transformer模型。</span><br />
<span style=\"font-size:14px;\">Institutional reserve is one of the important indicators for evaluating the stability and solvency of financial institutions. Accurate prediction of changes in payment institution reserves is of significant value for risk management and regulation by regulatory authorities in the context of centralized custody of reserves for third-party payment institutions. This paper proposes a prediction model for institutional reserve based on LSTM, Transformer, and LightGBM. The LightGBM model is utilized to extract key features of transaction logs, because tree-based model is fast and accurate in tabular data. The Transformer model is utilized to capture the local features of financial documents with its ability to model global context. Lastly, the LSTM algorithm is employed to capture the long-term dependencies of the combined data. Experimental results demonstrate that the proposed algorithm outperforms ARMA, LSTM, and Transformer models in predicting institutional reserves.</span>
%K 深度学习，时间序列预测，长短期记忆网络，Transformer，LightGBM<br>Deep Learning
%K Time Series Forecasting
%K LSTM
%K Transformer
%K LightGBM
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=81101