提出一种新颖的非线性时间序列预测模型,即滑动窗口二次自回归(MWDAR)模型.MWDAR模型使用部分的历史数据及其二次项构造自回归模型.模型参数用线性最小二乘法估计.应用模型进行预测时,预先选定窗口大小以及模型一次项和二次项的阶次.在每个当前时刻,先根据窗口内的数据估计模型参数,然后根据输入向量及模型参数做出预测.这种预测方法不仅适合小数据集的时间序列预测,而且对大数据集具有极高的计算效率.分别用Henon混沌时间序列数据和真实的股票交易数据作了MWDAR方法与局域线性化方法的1步和多步预测对比实验.结果显示MWDAR方法无论在预测精度上,还是在计算效率上都优于局域线性化方法. A novel nonlinear time series prediction model is proposed, which is sliding window quadratic autoregressive (MWDAR) model.MWDAR model uses part of the historical data and its quadratic autocorrelation model.The model parameters are estimated by linear least square method. When applying the model for prediction, the size of the window and the order of the first and second terms of the model are preselected, and at each current moment, the model parameters are estimated from the data in the window and then predicted based on the input vector and the model parameters. This prediction method is not only suitable for the prediction of time series of small datasets but also has a high computational efficiency for large datasets.The MWDAR method and local linearization method are respectively based on Henon chaotic time series data and real stock transaction data One-step and multi-step prediction experiments show that the MWDAR method is better than the local linearization method both in prediction accuracy and computational efficiency.