Research On The Dynamic Optimization Of Financial Hedging Method For Bulk Commodities Based On Deep Learning

Bulk commodities are important raw materials for enterprises.Their price fluctuations directly affect the procurement cost,the production and operation decisions of enterprises,the inflation level and the development of the national economy(Wu,2012;Tian and Tan,2014).The sharp fluctuation of commodity prices will bring huge risks to enterprises,which may lead to huge losses.In China,the government attaches great importance to financial hedging in managing price risk.For example,in 2021,the General Office of the State Council pointed out that the futures companies should provide risk management services to small and medium-sized enterprises,and help them use hedging tools to deal with the risk of sharp fluctuations in raw material prices.However,it is not easy to build an effective hedging strategy.The main reason is that the commodity prices often show complex nonlinear characteristics because of a variety of complex factors both at home and abroad.While the mainstream hedging strategies are often based on linear or quasi linear assumptions to model the price fluctuation process and the correlation between futures and spot prices.As a result,these models can not effectively extract the important nonlinear information in the actual price series,resulting in ineffective hedging strategy or even significant losses.We creatively construct a new deep learning driven dynamic hedging method to effectively manage the price risk of bulk commodities.When constructing this new hedging method,several aspects are considered.(1)We select representative bulk commodities,analyze and obtain important international and domestic influencing factors which affect the commodity prices.We then analyze the characteristics of these time series in detail,and preliminarily judge the applicability of the current mainstream hedging strategies.(2)If the current mainstream hedging strategies are not appropriate,how should we design and build a new more effective hedging method?(3)Based on the actual data,this paper empirically tests whether the new hedging method is robustly superior to the mainstream hedging methods.(4)The interpretability of newly constructed hedging method should be further analyzed to reveal more useful information.Based on the above considerations,a series of experiments has been carried out,and the main research results are as follows.(1)The price fluctuation series of copper,aluminum and zinc present complex nonlinear characteristics.The analytical methods used in this paper include descriptive statistical analysis,price and volatility graph,correlation coefficient heat map,principal component analysis and time-varying autocorrelation coefficient analysis.The nonlinear characteristics includes fat tail,sharp peak,non-normality,long memory and time-varying autocorrelation.In addition,other factors also show complex nonlinear characteristics of at tail,sharp peak.Besides,the correlation coefficients between these factors are relatively scattered,which means that high-dimensional nonlinear data should be modelled.(2)The newly constructed GARCH-(B)LSTM-ANN hybrid model can extract complex nonlinear information more effectively than the GARCH-ANN hybrid model widely used in the literature.So the new model can improve the volatility prediction accuracy.We combine the GARCH,ANN,LSTM and BLSTM model to construct new deep learning hybrid model,and analyze whether the constructed model gains better volatility prediction ability based on the actual data of copper,aluminum and zinc in China.Six deep learning models,including ANN、LSTMANN 、 BLSTM-ANN 、 GARCH-ANN 、 GARCH-LSTM-ANN and GARCHBLSTM-ANN,are compared in detail.In special,for the ANN part,different network depths(the number of hidden layers)and network widths(the number of neurons in each hidden layer)are tested.The empirical results show that,the best volatility prediction models for next two weeks,three weeks and four weeks for copper are GARCH-LSTM-ANN and GARCH-BLSTM-ANN.For aluminum,the best volatility prediction models are GARCH-LSTM-ANN and LSTM-ANN(only for a few special cases).And for zinc,the best volatility prediction models are GARCH-LSTM-ANN and GARCH-ANN(only for a few special cases).Therefore,in most cases,adding GARCH forecasts and memory cells can help to improve price volatility prediction accuracy for copper,aluminum and zinc in China.(3)The deep learning hedging method based on risk early warning can obtain significantly better hedging performance than the mainstream hedging methods.Regarding to the mainstream hedging methods,although in theory the dynamic hedging methods should show better hedging performance than the static hedging methods due to the time-varying characteristics of financial time series,such as volatility clustering,time-varying variance and covariance.However,many studies have found that the dynamic hedging methods are not significantly better than or even weaker than the static hedging methods.The reason may be that the traditional dynamic hedging methods can not effectively deal with the complex nonlinear characteristics of financial time series.Therefore,this paper integrates the deep learning technology that can effectively deal with the nonlinear characteristics of data into the hedging method,constructing the deep learning hedging method,which can dynamically adjust and optimize the hedging ratios according to the risk early warning information obtained from the deep learning volatility prediction model.The hedging performance of deep learning hedging method is compared with the mainstream static hedging method OLS and the dynamic hedging method DCC.The empirical results show that compared with the traditional OLS and DCC hedging methods,the deep learning hedging method can obtain a higher risk-adjusted return in out-of-sample.The empirical results also show that the deep learning hedging methods can bring significant economic benefits to enterprises,especially in 2020,when the price fluctuates more sharply.Finally,this paper further verifies the robustness of the empirical results under three different settings.(4)We further provide explanatory analyses for the deep learning prediction method and the deep learning hedging method.Firstly,we give explanatory analyses on the effectiveness of deep learning hedging method.The results show that,this method can significantly improve the return of the hedged portfolio and at the same time significantly control the risk of the portfolio,especially the downside risk.Secondly,we analyze the interpretability of deep learning method,which can provide effective risk early warning information.The results are as follows.First,overall,adding GARCH forecasts or memory cells can improve the prediction ability of deep learning hybrid model.Moreover,the effectiveness of GARCH forecasts tends to be positively correlated with the complexity of data,and the effectiveness of memory cells is positively correlated with the significance of longterm memory phenomenon in the volatility series.Second,based on the risk warning factor importance results,it is found that there is financialization phenomenon in the pricing of copper,aluminum and zinc,besides their price fluctuations are affected by both international and domestic factors.The most important factors for the three commodities exist similarities and differences.Specifically,for copper,the US dollar exchange rate and WTI crude oil futures price show the greatest impact on the future copper price volatility,followed by copper spot price and Shanghai copper futures price.For aluminum,WTI crude oil futures price is also an important factor affecting aluminum price volatility in the future.In addition,the Chinese and American financial markets(Dow Jones Industrial Average and CSI 300 index)also have significant impacts on aluminum price volatility.For zinc,the impact of WTI crude oil futures price is also high.In addition,the impact of metal markets such as copper price in China and zinc spot price in London Metal Exchange are also significantly high.What’s more,the GARCH forecasts are also important risk warning factors for copper,aluminum and zinc,which verifies the effectiveness of adding GARCH forecasts into neural network.Third,the dynamic sensitivity analysis of price volatility to risk warning factors shows that there exist complex nonlinear functional relationships between the volatility and the risk warning factors,and there also exist interactive effects between different risk warning factors.The main contributions of this paper are as follows.(1)This paper creatively constructs a new deep learning hedging method,which can obtain significantly better hedging performance.In literature,the hedging methods have roughly experienced the changes from the static methods such as OLS,Va R and VECM,to the dynamic methods such as the GARCH-type methods,in order to obtain better hedging performance.We find that the futures and spot prices often show complex nonlinear fluctuations,while the current mainstream static and dynamic hedging methods are often modeled based on linear or quasilinear assumptions,which may lead to ineffective hedging results because they can not effectively extract non-linear information.To solve this problem,this paper introduces deep learning technology to extract the relevant nonlinear information in the price fluctuation series,and combines it with the mainstream hedging methods to construct deep learning hedging methods.The empirical results show that the dynamic hedging method driven by deep learning can robustly and significantly improve the hedging performance.(2)This paper creatively constructs a more effective risk warning model for commodity price fluctuation risk.Previous studies in the field of financial volatility prediction often emphasize the use of econometric models,especially the GARCHtype models,to predict price volatility,such as Hammodeh and Yuan(2008),Khalifa et al.(2011),Trück and Liang(2012),Chkili et al.(2014),and so on.Recently,with the development of deep learning technology,researchers began to explore using the hybrid model of ANN and GARCH to improve the volatility prediction accuracy,and gradually they realized the necessity of combining deep learning technology with expertise in the financial field,such as Kristjanpoller and Minutolo(2015),García and Kristjanpoller(2019).However,ANN is a kind of memory free network,which can not effectively deal with the long memory phenomenon in financial time series.This paper further constructs a new deep learning hybrid model,which further adds the memory cells to the neural network model to further improve the volatility prediction ability of the hybrid model.(3)This paper expands the research on the combination of hand-engineered features and deep learning technology,and finds that adding hand-engineered features can significantly improve the volatility prediction ability of the deep learning model.This model can effectively combine the financial expertise and the deep learning technology.In fact,Lecun et al.(2015),a very influential paper,pointed out that the deep learning model composed of multiple processing layers can automatically extract the required features from the data.Therefore,they should hardly need the hand-engineered features obtained by experts.However,as the financial market is affected by many factors,so the price fluctuation is quite complex.How to effectively design a deep learning framework to extract useful features from the financial data is not clear.In this paper,the deep learning hybrid model constructed by integrating GARCH forecasts partially reveals this problem,in which GARCH forecasts can be regarded as hand-engineered features extracted by financial expertise.The results show that the deep learning technology can not replace the professional knowledge in the financial field.On the contrary,they should be combined.Therefore,the findings of this paper complement and expand the research results of Lecun et al.(2015),and provide useful guidance for effectively integrating finance knowledge to build an effective deep learning hybrid model.(4)This paper extends the research on interpretability for deep learning model in the field of finance.Recently,the interpretability for deep learning models are getting more and more attention by both academia and industry.In the financial field,researchers have also made some attempts recently.For example,Gu et al.(2020)explained why the deep learning model can better predict the cross-sectional return of stocks through the reduction of R2,sum of squared partial derivatives,and the interaction between explanatory variables.Gu et al.(2021)found variable importance for stock return in the deep learning prediction model also through the reduction of R2.However,compared with the large number of deep learning prediction research in the financial field,the research on interpretability is far away from meeting the needs.This paper expands the existing research,and creatively explore the interpretabilities for deep learning model by analyzing the effectiveness of memory cells and hand-engineered features,the sensitivity of price volatility to risk warning factors.By doing so,we want to reveal the decision-making mechanisms for deep learning model.(5)This paper innovatively provides a systematic hedging framework.This paper takes hedging theory,nonlinear price fluctuation risk,deep learning risk warning model,deep learning hedging method,the interpretabilities for deep learning method and deep learning hedging method as a whole.The lack of any of these five parts will affect the integrity of this paper.The first part,i.e.the hedging theory,is the basis and starting point of this research.The research about nonlinear characteristics in the second part leads to the doubt on the effectiveness of the mainstream hedging methods in managing the price risk.The third part attempts to build a new deep learning hybrid model to extract the complex nonlinear information in the data,so as to improve the volatility prediction accuracy,and provide effective risk warning information.Based on the risk warning information,the fourth part further dynamically optimizes the hedging ratios,constructing an effective deep learning hedging method.After finding a more effective hedging method,this paper further reveals the source of the effectiveness,which is shown in the fifth part.That is,the interpretabilities for deep learning method and deep learning hedging method.