基于循环神经网络的交易量预测问题的研究

Research on Transaction Volume Prediction Based on LSTM Network

随着人工智能技术的快速发展,以大数据为依托的智能化的金融科技场景受到越来越多的关注,通过对海量金融类数据进行分析,可以为其各个场景提供精准的决策支持。通过预测某金融机构产品交易量,可以提前1天发现异常的交易流量和渐变趋势,为业务部门提供决策支持,同时也可为系统扩缩容提供决策支持,进而减少系统线上压力或者释放不必要的系统资源。线性算法比如AR模型、MA模型、ARMA模型等由于对历史数据有着强假设条件,在对文中研究处理的非稳定数据集中的节假日期间的交易量的预测结果较差。利用深度学习算法,设计并实现了基于LSTM的交易量预测模型LSTM-WP(LSTM-WebPredict),通过发现与学习历史数据的特征,可以在线性算法的基础上,将对节假日交易量预测的准确度提高约8%。

With the rapid development of artificial intelligence technologies,intelligent financial technology sce-narios based on big data have attracted increasing attention.Through analysis of massive financial data,accurate de-cision-making support can be provided for various scenarios.By predicting transaction volumes of a certain financial product of a bank,abnormal transaction flows as well as gradual trends can be predicted one day in advance.This can provide decision support for both business plan formulation and system expansion/contraction strategy,thereby reduc-ing online system pressure or releasing unnecessary system resources accordingly.However,this study's data set shows that due to strong assumptions on historical data,linear algorithms such as Autoregressive(AR)model,Moving Average(MA)model,and Autoregressive Moving Average(ARMA)model yield relatively poor performance on predicting holiday transaction volumes.This article uses deep learning algorithms to design and implement an LSTM-based trading volume prediction model LSTM-WP(LSTM-WebPredict),which improves the accuracy of holiday trade forecast by roughly 8%.