Modeling And Prediction Of The Volatility Of Highway Freight Rate With O2O-based Freight Transportation Platform

Before the advent of the era "Intelligent Logistics + Big Data",China's road freight market has been in a state of asymmetric information and non-circulation for a long time,which is determined by its tripartite freight bargaining system with freight forwarders as intermediaries.Under the traditional tripartite freight bargaining system,freight forwarders,as intermediaries,specifically find the carrier for the shipper and charge a certain fee from it.Due to the opaque price,the carriers lose the normal bargaining power and it hinders the freight rate on the road freight market and the improvement of the operating model.In addition,due to the lack of integrated truck and cargo matching management and optimization,as well as the appropriate freight pricing standards,truckers spend most of their time on finding,distributing and negotiating goods,resulting in a large amount of resource consumption.Nowadays,with the rapid growth of China's road freight market,to ensure the reasonableness of the shipper and carrier's revenue,it is urgent to standardize the freight rate of the traditional road freight market.With the development of intelligent logistics,large amounts of the realtime vehicle and cargo matching information(e.g.waybill information,transaction price,driver information,and vehicle information,etc.),has provided new ideas for building a smart road freight pricing system.Based on the Online To Offline(O2O)road freight platform data,this paper studies the processing method of road freight price anomaly data to ensure the quality of data used for the following volatility analysis and freight price prediction,establishes a series of volatility models and volatility prediction models,analyses the time and regional correlations of freight rates on different transport routes,and proposes a time-delay matrix and short-term freight rate prediction model.Specifically,the research work of this article mainly starts from the following aspects:(1)According to the O2O road freight platform data,a method for analyzing and identifying abnormal characteristics of O2O road freight data is proposed.From the perspectives of time series and probability,the basic characteristics of highway freight rate anomaly data are analyzed in detail,focusing on the scatter diagrams of freight rate-tonnage,freight rate-vehicle length,time-freight rate,etc.And then the combination method of model and clustering idea is proposed to eliminate the abnormal freight rate data – a combination of statistical model,quartile method,and k-means clustering.This combination method proposes the idea of secondary clustering,which solves the problem that it is difficult to determine the number of abnormal data clusters of different types of transport goods and different transport tonnages to a certain extent.This method is versatile.The proposed method is validated using the southwest highway freight data with typical road freight characteristics.The analysis through multiple quantitative index tables such as modeling error curves,showing that the abnormal data combination elimination method was effective for abnormal transportation of different cargo types and different transport tonnages.It has strong versatility and practicality and can reduce the modeling error of freight rate curve,and then provide a data basis for the analysis and prediction of highway freight rate fluctuation and freight rate forecast.(2)Focusing on the characteristics of highway freight rate volatility(FRV),this study summarizes the fluctuation characteristics of highway freight rates and measures the time characteristics,relevant statistical characteristics and industrial transmission characteristics of the fluctuation of freight rates of various transportation routes in the region.Firstly,the definition of highway FRV is given,and then the basic characteristics and fluctuation features of highway freight rate are systematically described from basic statistics,Autoregressive Conditional Heteroskedasticity(ARCH)effect,and stationarity,etc.,verified its nonlinear heteroscedasticity and time-varying characteristics,and further discussed the nonlinear correlation and substitutability between volatilities series of different cities based on the Copula function and nonlinear correlation coefficients.The results of the case study using southwest highway freight data show that the FRV series of different routes show different changing trends and statistical characteristics,but they do not obey the normal distribution and have sharp peaks and thick tails,similar to the financial time series.Besides,market volatility has a certain period.The time-variance variance shown in the series indicates that the classical theory of homoscedasticity assumptions does not conform to the fluctuation characteristics of the time series of highway freight rates.The Generalized ARCH(GARCH)model with skewed t distribution can better capture the time-varying characteristics of the FRV series.When external shocks aggravate FRV,each transportation route presents different market performance,such as the freight rate of the route Chengdu to Kunming is the most sensitive to market fluctuations,and the duration of the shock effect is longer as well,while the freight rate of the route from Chongqing to Kunming is the least sensitive to the market,and the duration of the shock effect is shorter,which indicates the freight market itself is relatively stable.In addition,in terms of the correlations between FRV of different cities,results show that the correlations are nonlinear.Nonlinear correlation coefficients calculated based on Gumbel Copula Function indicates the directions of changes of FRV in cities is basically same,but the substitution effect between FRVs in each city is small.(3)To deal with the leverage effect of road FRV and its prediction,the leverage effect of FRV and its influencing factors are studied,and a rolling iterative hybrid heteroscedasticity prediction model with neural networks is proposed.Based on the analysis of the characteristics of FRV,the leverage effects of FRV series are modeled and analyzed.Besides,the reasons for the leverage effect of freight rates are summarized from the aspects of supply and demand,road freight spot market transactions,and types of goods transported.At the same time,the impact and transmission characteristics of external shocks(changes in policies,oil prices,etc.)on the fluctuation of highway freight rates are analyzed.On this basis,a hybrid heteroscedasticity prediction model combining neural network and GARCH-type model is proposed,and one-step-ahead rolling forecasting with re-estimation of the model at each step is used to conduct forecasts.The results of a case study using Southwest highway freight data show that the Southwest freight market has obvious leverage effects,and the Exponential ARCH(EGARCH)model with a skewed t distribution fits well the leverage effect characteristics of each freight rate volatility sequences:negative shocks have a higher impact on volatility than positive shock.That is,under the same shock amplitude,shippers and carriers are more sensitive to bad news,bad news brings more freight rate fluctuations.Although there are leverage effects in the southwest freight market,the reasons for the leverage effects vary from the transportation routes,which mainly due to the different combinations of transportation freight types for each route.In terms of volatility prediction,the case results show that considering the prediction accuracy and prediction significance,the hybrid heteroscedastic model and Autoregressive Integrated Moving Average(ARIMA)model have the smallest rolling iteration prediction error in one-step-ahead forecasts.Reestimating the prediction model at each step can reflect the changes of the prediction environment in real-time,improve the adaptability of the model,and ensure the robustness of the prediction.Results show that re-estimating the prediction model at each step has little improvements on the prediction accuracy of ARIMA,but for the hybrid model,the prediction accuracy is significantly improved.(4)For the issue of the volatility transmission relationship between different FRV,this paper studies the spillover effect between freight rate volatility of different truck lengths in the region and the transmission path of freight rate volatility based on the multivariate GARCH model.The results of the study found that the transmission of different volatility is asymmetric.There is a strong dynamic or constant linkage relationship between the freight rate volatility of different length-specific trucks.The transmission of volatility between freight rates of different large trucks is two-way,and the impact of large trucks on the volatility of freight rates of small trucks is one-way.In the current freight market,large trucks are the main interprovincial capacity supply,and small trucks are mainly short-distance transportation.Therefore,in some provincial capitals,the share of small trucks is occupied by large trucks(5)For the short-term forecast of highway freight rates,this paper proposes a weighted regression model based on the temporal correlation and route-between correlation of routelevel freight rates,named Lagged weight matrix-based Multiple Regression Model(LagWMR)model.This method not only considers the dependence of the traditional time series on the temporal dimension,but also quantifies the transaction behavior of shippers and carriers in the spot freight market,and converts it into the prediction model.On the one hand,compared with the traditional regression model,the Lag-WMR model uses machine learning to select the critical feature combinations that affect the route-level freight rates,on the other hand,this paper proposes a practical time-lag weighting matrix based on the freight rate characteristics.The forecasting method is verified by the southwest highway freight data.The results show that the Lag-WMR model considering the temporal and between-route correlations performs best when compared with benchmarks.Sensitive analysis demonstrates that the improvement of prediction accuracy by only considering the between-route correlation is far greater than the model only considering temporal correlation.To make up the traditional inability to scientifically determine the set of explanatory variables,the method proposed in this paper is to first establish an initial feature set and then screen key features through machine learning,which is a good complement to the optimization of multiple regression models.Furthermore,results indicate that more explanatory variables do not necessarily improve the prediction accuracy of the model,because the overfitting is usually caused by too many features.The"feature importance"function in machine learning can optimize and highlight the most powerful information to the freight rate,avoid information redundancy and overfitting,and thus effectively improve the prediction accuracy.Using the time-lag matrix as the weights in between-route correlations can not only reflect the regional correlation of freight rates but also effectively convey the temporal dependence of freight rates.The LagWMR model provides a new idea for spatial-temporal prediction issues.