Network evolution analysis of nickel futures and the spot price linkage effect based on a distributed lag model

Abstract

In many cases, the correlation between time series has a certain lag effect. To study the lag correlation between two time series variables, we select London Metal Exchange (LME) nickel futures and spot prices from 3 January 2008 to 29 December 2017 as sample data to carry out stationarity tests, cointegration tests and Granger causality tests to determine the stationarity and correlation of two time series. Then, we use the method of combining the distributed lag model and sliding window method to construct a network. We select the best sliding window length through a sensitivity test. The time series is reconstructed into a complex network by taking the types of patterns as the nodes and the conduction relationship between the patterns as the edges. The number of transitions between patterns is defined as the weight of the edge. The results show that the spot price changes are caused by the change in nickel futures price and that the optimal sliding window length is 64. Additionally, 12 types of patterns account for a large proportion of the patterns in the network. Six patterns are the main intermediaries of pattern transmission and appear centrally with the change in the market environment. Therefore, the relationship model between these futures and spot prices has remained stable for a long time. Combining the positive and negative news of the market, we identify the timing of the change in the relationship model and can use this approach to improve the accuracy of early warning methods. This study provides a method to construct a complex network using a distributed lag model, which can help analyze two real time series variables with lag correlation.

Incorporation of causality structures to complex network analysis of time-varying behaviour of multivariate time series

“…The transfer of a time-varying multivariate system to a complex network to analyse its evolution has been employed in a number of studies 13 , 36 , 40 , 61 , 62 . Although the exact methodology for this transfer varies, the general approach can be seen as similar.…”

“…The previously mentioned studies employing complex networks for the evolution of multivariate systems by Jiang et al, Yu et al, Qi et al, Dong et al, and Yu et al 13 , 36 , 40 , 61 , 62 also subsequently analyse their constructed networks. This analysis takes a number of forms but primarily it determines properties and metrics of the network that can then be used to describe the evolution of the original system.…”

“…This analysis takes a number of forms but primarily it determines properties and metrics of the network that can then be used to describe the evolution of the original system. This analysis can be characterised as the following types: determination of the degree of the nodes, the distribution of this degree, and the comparison of this to standard network models, such as a scale free network 13 , 36 , 40 , 61 ; the distribution of edge weights (frequency), comparison to power law distribution, and total level of self-loops within the network (i.e. when a causality pattern does not change over a time step) 13 , 36 .…”

“…This analysis can be characterised as the following types: determination of the degree of the nodes, the distribution of this degree, and the comparison of this to standard network models, such as a scale free network 13 , 36 , 40 , 61 ; the distribution of edge weights (frequency), comparison to power law distribution, and total level of self-loops within the network (i.e. when a causality pattern does not change over a time step) 13 , 36 . Common network methods are used, such as the closeness of the network 40 , and the betweenness of nodes 36 , 40 , 61 , 62 .…”

“…These complex network approaches to discovering behaviour have been employed by a number of researchers 9 , 12 , 13 , 15 , 36 – 40 , demonstrating their usage in the field. However, there still exists opportunity for expansion within these techniques 13 , 33 .…”