| 摘 要: 在通过LightGBM、Prophet、HotWinters等单一预测算法，以及通过引入搜索指数等消费者行为数据作为预测变量等无法达到汽车品牌销量预测精度的情况下，基于销量数据特征遴选出HotWinters、Prophet、LightGBM三个预测模型，并自主构建了Musgrave方法，以此四个算法构建了组合预测模型，并结合熵值法作为权重动态变化的方法，构建了“动态变权组合预测策略”。本策略使用“单地区多品牌维度、多品牌维度、多地区多品牌维度”三种方式进行六期预测并检验预测效果，结果表明三种方式预测误差中位数分别为7.50%、6.11%、9.61%，因此，本策略能够满足对具有复杂多变特征的数据进行预测的需要。 |
| 关键词: 动态变权；组合预测；Musgrave；熵值法 |
|
中图分类号: TP311
文献标识码: A
|
|
| Research on Dynamic Variable Weight Combination Prediction with Mixed Characteristics of Multi-objective Time Series |
|
DONG Zhixue, GONG Yue, HU Yong, GAN Mengzhuang
|
(Algorithm Department, Beijing Autohome Information Technology Co., Ltd., Beijing 100190, China )
602932093@qq.com; gongyue.best@163.com; hyalone@126.com; 531609777@qq.com
|
| Abstract: Prediction accuracy of automobile brand sales cannot be achieved through single prediction algorithms such as LightGBM, Prophet, HotWinters, and the introduction of consumer behavior data such as search indexes as predictors. Based on the characteristics of sales data, this paper proposes to select three predicting models, LightGBM, HotWinters and Prophet, and independently build Musgrave method. With these four algorithms, a combined forecasting model is constructed, and a "dynamic variable weight combination predicting strategy" is constructed by taking the entropy method as a method of dynamic weight change. This strategy uses three methods of "single-region multi-brand dimension, multi-brand dimension, and multi-regional and multi-brand dimensions" to conduct six-period predictions and test the prediction results. Results show that the median prediction errors of the three methods are 7.50%, 6.11%, 9.61%. Therefore, this strategy can meet the needs of predicting data with complex and changeable characteristics. |
| Keywords: dynamic variable weight; combined prediction; Musgrave; entropy method |
