| Gold has indelible physical and chemical properties and has occupied a very important position in human history for thousands of years.It is of great significance to predict the price of gold.With the development of statistical science and computer technology,people continue to try to use machine learning methods to predict financial market data,and put forward a variety of data feature mining prediction models.Considering a large number of disturbing factors in the financial market,the gold price data contains a lot of noise,which hinders the construction of the model.Machine learning method has a strong ability to deal with nonlinear data,but it will pay too much attention to noise data in the training process,which will lead to poor fitting effect and reduce its ability to predict gold price.Therefore,this paper first uses wavelet transform to denoise the data and eliminate the noise part in the gold price data,so as to reduce the impact of short-term noise interference on the machine learning model and improve the prediction ability of the model.BP neural network and short-term memory network(LSTM)are two machine learning methods that can be used to fit time series.BP neural network model is a multilayer feedforward network model,which is mainly trained by error reverse transfer algorithm.The model can realize nonlinear fitting of data by flexibly adjusting the number of middle layers of the network and the connection of processing units of each layer.Wavelet neural network is an improvement of BP neural network.The main improvement is to use the nonlinear wavelet basis function to replace the activation function in BP neural network,and skillfully combine wavelet transform with neural network.This construction makes the wavelet neural network model inherit the advantages of wavelet analysis and BP neural network,so that the model has a more flexible and effective function approximation ability.Long and short term memory network(LSTM)overcomes the phenomena of gradient disappearance and gradient explosion in cyclic neural network by introducing the concept of gating into cyclic neural network.The independent storage structure of neurons in LSTM can solve the problem of long order dependence of data.On the basis of LSTM and the idea of wavelet neural network,this paper constructs a new wlstm neural network,which also uses nonlinear wavelet basis to replace the activation function of LSTM.so that LSTM network has a better ability to fit gold price data.By using a variety of machine learning methods and traditional time series models,this paper makes an empirical analysis on gold futures price data,so as to verify the applicability of the improved machine learning model to gold futures data analysis.This paper selects Comex gold futures data of New York Mercantile Exchange,and introduces gold price related data such as silver futures,US dollar index and crude oil price as the sample of this study.In the gold price prediction experiment,from the perspective of univariate input and multivariable input,a variety of models are used to model and predict the gold price,and ARIMA model is used as the comparison model to verify the prediction ability of the improved model based on wavelet analysis.Finally,aiming at the defects of the above machine learning methods and the future development direction,this paper makes an in-depth discussion on the optimization of the model. |
