统计与数据科学学院知识库

黄金是金融市场上最重要的商品之一,它是一种兼具货币和金融属性的特殊贵重金属，同时拥有高流动、高安全和低收益的特性。黄金期货价格的变化受到多种因素的综合影响，例如供需关系、全球经济状况、国家政策、汇率、股市指数与原油价格等。黄金期货价格波动较为复杂，再加上突发事件的干扰，使黄金期货价格预测变得越来越具有挑战性。传统的预测方法难以满足当前环境下对黄金期货价格短期预测的要求。因此本文从特征选择与特征融合两个角度出发分别对黄金期货价格及其影响因素进行分析，旨在探索提高黄金期货价格预测精度的新方法。主要工作如下:
在特征选择方面，首先选用弹性网络对中国黄金期货及其影响因素进行选择和降维，利用相空间重构法与时差相关分析法分别确定黄金期货本身的滞后期及压缩变量的滞后期;预测模型方面，在 BP、ELM、LSSVR 三种模型的基础上选用粒子群优化算法(PSO)和灰狼优化算法GWO)对基础模型参数进行优化并进行滑动窗口预测。预测结果表明，GWO-ELM 模型相较于其他模型有着更快的寻优速度，具有鲁棒性，在此基础上为了提高黄金期货价格预测方向精度，对已选GWO-ELM 模型进行再次优化,设定方向精度与水平精度结合的适应度函数去训练网络模型，结果表明更改适应度函数的预测模型在一步预测与多步预测的方向精度都要好于原有的 GWO-ELM 模型，最后针对一步预测与多步预测的实验结果，设计不同的交易策略进行回测，通过策略收益率等相关指标进一步验证了更改适应度函数的 GWO-ELM 模型有着良好的预测效果。
在特征融合方面，不仅考虑黄金期货价格及市场等因素的影响，同时在影响因素中纳入投资者关注度度量指标即百度指数数据集，通过 Pearson 相关系数MIV 和 KPCA 方法对混合数据进行筛选及特征融合，对比分析信息贡献率来确定MIV 与KPCA 最佳值，选用对比得出稳健性较强的 GWO-ELM 模型进行-步预测，对比特征选择方法在混合数据集中的预测效果，得出特征选择方法在多维度信息的黄金期货价格预测方面更加适用。

Gold is one of the most important commodities in the financial market. It is a special precious metal that combines both currency and financial attributes, while possessing the characteristics of high liquidity, high security, and low yield. The change in gold futures prices is influenced by various factors, such as supply and demand, global economic conditions, national policies, exchange rates, stock market indices, and crude oil prices. The fluctuation of gold futures prices is quite complex, and the interference of unexpected events makes gold futures price prediction increasingly challenging. Traditional prediction methods are difficult to meet the requirements for short-term prediction of gold futures prices in the current environment. Therefore, this thesis analyzes the gold futures prices and their influencing factors from the perspectives of feature selection and feature fusion, aiming to explore new methods to improve the accuracy of gold futures price prediction. The main work is as follows:

Regarding feature selection, the elastic net is first used to select and reduce the dimensionality of the Chinese gold futures and their influencing factors. The phase space reconstruction method and time-lagged correlation analysis method are used to determine the lag period of the gold futures themselves and the lag period of the compressed variables, respectively. In terms of prediction models, based on the BP, ELM, and LSSVR models, the particle swarm optimization algorithm (PSO) and gray wolf optimization algorithm (GWO) are used to optimize the basic model parameters and conduct sliding window prediction. The prediction results show that the GWO-ELM model has a faster optimization speed and robustness compared to other models. To improve the directional accuracy of gold futures price prediction, the already selected GWO-ELM model is further optimized by setting an adaptive function that combines directional and horizontal accuracy to train the neural network model. The results show that the prediction model with the modified fitness function has better directional accuracy in both one-step and multi-step prediction than the original GWO-ELM model. Finally, different trading strategies are designed to backtest the experimental results of one-step and multi-step prediction, and relevant indicators such as strategy yield are used to further verify the good prediction performance of the modified fitness function GWO-ELM model.

Regarding feature fusion, not only the impact of gold futures prices and market factors is considered, but also the investor attention measurement index, namely the Baidu search index dataset, is included in the influencing factors. The Pearson correlation coefficient, MIV, and KPCA methods are used to screen and fuse the mixed data, and the information contribution rate is compared and analyzed to determine the optimal threshold for MIV and KPCA. The GWO-ELM model with strong robustness is selected for one-step prediction, and the prediction effects of feature extraction methods are compared for the mixed dataset. It is found that feature extraction methods are more suitable for predicting gold futures prices in multi-dimensional information.