基于多时间尺度组合策略的我国猪肉价格预测研究

A Multi-Time Scales Combination Strategy for Pork Price Forecasting

为提高猪肉价格多步预测精度,基于组合预测思想,提出一种包含多时间尺度的组合预测模型.以日度价格为基准,将周度及月度预测值转化为日度频率数据,形成多时间尺度预测结果,反映价格时序的多维数据生成过程.针对短,中,长期预测需求,构造"日度+周度"以及"日度+周度+月度"组合方案,设计6种权重计算方法,探讨预测步长与时间尺度的匹配关系.以我国集贸市场猪肉平均批发价格为研究对象,实验结果表明:就权重计算方法而言,基于最小二乘的权重优化方法可取得优于单模型的组合预测结果;多时间尺度的组合策略不能改进短期预测性能,但能有效提高中,长期预测精度;组合策略中时间尺度的选择应与预测步长所包含时间跨度相匹配,以充分涵盖预测期限内的价格波动信息.

Aiming to improve the multi-steps forecast accuracy of pork price,a forecast combination model which involves multi-time scales is proposed.Based on daily price data,weekly and monthly forecasts are converted into daily frequency.The forecasts with multi-time scales reflect the data generation process of the time series multi-dimensionally.Regarding the short-,middle-and long-term forecast tasks,"day+week"and"day+week+month"combination schemes are constructed with 6 weighting methods,to investigate the matching relationship between forecast horizon and time-scale.Using the average pork price from wholesale market as an example,experimental results show that,weighting method based on least square is the optimal one which yields superior forecast than the single model.Although multi-time scales combination strategy cannot improve short-term forecast performance,it can effectively enhance the accuracy of middle-and long-term forecast.The selection of time-scales should correspond to the time span within forecast horizon,to fully cover the price volatility during that specific period.