B样条与LSTM结合的时间序列建模方法研究——以股票收盘价数据为例

Study on Time Series Modeling Method of B-Spline Combined with LSTM—Take the Closing Share Price as an Example

摘要: 时间序列是指将某种统计指标的数值按照时间顺序排列所形成的序列，对时间序列进行分析预测有助于我们提前判断、减少潜在风险，在生产和生活中都有着重大的意义。常见的时间序列数据包括股票数据、商品日销量以及每日的气候等等。其中股票数据是一个高度复杂的非线性系统，对股票长期趋势的预测一直是一个令人感兴趣的话题。本文以股票收盘价数据为例，提出一种新的思路，对时间序列数据进行建模并预测。首先采用B样条回归对股票价格与时间变量之间的非线性关系进行建模。由于B样条曲线只能在一个提前给定的区间内进行拟合，对区间外的预测问题无能为力，因此，为解决这个问题，本文将B样条回归与长短期记忆(Long Short-Term Memory)神经网络模型相结合，构建LSTM-Bspline模型，通过训练神经网络来获取B样条曲线的参数，从而获得预测值。最后通过对某公司股票每日收盘价数据进行分析预测，并将预测结果与经典LSTM神经网络预测结果相比较，证明了LSTM-Spline模型的可行性。

Abstract: Time series refers to the series formed by arranging the values of certain statistical indicators in chronological order. Analysis and prediction of time series can help us to judge in advance and reduce potential risks, which is of great significance in both production and life. Common time series data include stock data, daily sales of goods, daily weather, and so on. Among them, stock data is a highly complex nonlinear system, and the prediction of stock long-term trend is always an interesting topic. Taking stock closing price data as an example, this paper proposes a new way of thinking to model and predict time series data. First the B-spline regression is adopted to model the nonlinear relationship between stock price and time variable, the B-spline curve can only be a in advance within a given interval fitting, outside the range of powerless forecasting problems, therefore, to solve this problem, this paper combined the B spline regression with Long Short-Term Memory neural network model, build LSTM-Bspline model, by training neural network to obtain the parameters of the B-spline curve, the predicted value is achieved. Finally, by analyzing and forecasting the daily closing price data of a company's stock, and comparing the forecasting results with those of classical LSTM neural network, the feasibility of LSTM-Spline model is proved.