China commodity futures market network is constructed. Commodity is the node of the network, and the network link is defined by the price correlation matrix. We analyze the relationship between the centrality of each commodity in the commodity futures market network and the optimal weight of the commodity portfolio, empirically examine the market system factors and commodity personalized factors that affect the centrality of commodity, and evaluate the effect of network structure on the optimization of commodity portfolio selection under the mean-variance framework. It is found that the commodities with high network centrality are often related to industrial products and have high volatility. Commodities with higher centrality have lower portfolio weights. We put forward a kind of commodity futures investment strategy based on network, according to the network centricity grouping the commodities, the network centricity lower edge of the commodity structure of the portfolio, cumulative yield is better than that of centricity higher core product portfolio, the whole market portfolio yield, but due to large maximum retracement, lead to the stability and ability to resist risk compared with the other two groups of goods combination. The main contribution of this paper is to optimize portfolio selection by establishing the relationship between portfolio weight and commodity centrality by using commodity futures market network as a tool.

Although China has the largest trading volume of commodity futures, limited studies have been devoted to the term structure of Chinese commodity futures. This paper takes the tools in functional data analysis to understand the term structure of commodity futures and forecast its dynamic changes at both short and long horizons. Functional ANOVA has been applied to examine the calendar e_ect of term structure in level and _nd the seasonality in the commodity futures of coking coal and polypropylene. We use an h-step functional autoregressive (1) model to forecast the dynamic change of term structure. Comparing with native predictor, in-sample and out-of-sample forecasting performance indicate that additional forecasting power is gained by using the functional autoregressive structure. Although the dynamic change at short horizons is not predictable, the forecasts appear much accurate at long horizons due to the stronger temporal dependence. The predictive factor method has a better in-sample _tting, but it cannot outperform the estimated kernel method for out-of-sample testing, except for 1-quarter-ahead forecasting.