Freight transportation is a service representing derived demand, which is derived from the direct demand for the movement of commodities, materials, goods, products or logistics from the production source to the final attraction market (Tolliver, 1994). Associated with the freight service is the level of service (generalised costs, prices, rates, impedance or utilities, benefits, savings) with which the freight transportation is performed (Fair and Williams, 1975). In addition, freight demand has both temporal and spatial dimensions for the wide varieties of commodities (Tolliver, 1994). Hence, freight transportation demand (FTD) is dependent on the following conceptual form:

(1)

where F_D is the derived freight demand, A is the activity generating the direct demand between parties p for the commodity c, S is the level of service of freight transportation provided, t is time dimension and x is spatial geographical dimension.

The study of freight demand forecasting is important for development policy options, society, economy and environment, yet it is relatively less researched, more complicated and differentiated with more variables (markets, attributes, parties, policies) that are more difficult to model and surveyed as well as more integrated into overall management (logistics production process, inventory control and stocks) than passenger transportation demand (Ortuzar and Willusem, 2002).

This paper summarises approaches to freight demand modelling. It begins with a brief historical development of freight demand forecasting, followed by the main factors, difficulties and differences of freight characteristics from passenger characteristics. It then presents modelling requirements, existing methodologies and critically reviews the available models with respect to uses and limitations. A simple case study with actual data is then presented. It concludes with likely state-of-the-art developments.

Research on freight demand modelling lags behind that of passenger travel modelling (Regan and Garrido, 2000). However, the economics of military and business logistics and freight movement actually start from the Industrial Revolution in the 19^th century (Fair and Williams, 1975).

Freight transportation planning, of which freight demand is an important determinant, and supply of freight services have been of major interest since the early 1960s (Shankar and Pendyala, 2000). Published freight studies reflect different modelling approaches. Historically, the modelling is confined to separate non-overlapping well-defined principles and approaches, which are presented in section 4. Separation is based on geography, data nature and methodology principles. Freight demand modelling is separated from shipper behaviour, carrier behaviour and shipper-carrier relationships which would affect freight demand (Regan and Garrido, 2000).

Two types of derived demand for freight transport (Fair and Williams 1975):

1. Production freight demand: volume movement of basic materials to processing and manufacturing plants as well as inter-plant volume movements. Heavily influenced by size of production region, degree of region specialization, high concentrations of origins, destinations and mainline routes (haul-lines) and population density (cities have net freight inflow). Due to scattering of originating sources, freight termination at processing centres is more concentrated. Commodity characteristics (perishability, flammability, fragileness, bulkiness, weight and ease of handling) and market pattern (shipper/carrier cost and service requirements of adequacy, economy and service quality) shape the movement. Bulk freight over long hauls rely heavily on railroads, water carriers and pipelines.
2. Distribution freight demand: distribution of finished materials, products and perishables as value-added merchandise to the ultimate industrial and consumer markets. Importance lies in the commodity value, revenue generation and service costs. Major categories include foods, refined liquid products and manufactured nonperishables. Different conditions from production freight – packaging required, loss/damage is serious, expedition & careful handling needed, reduce deterioration of perishables and customers demand fast, safe and dependable delivery to lower costs and risks of large inventories. Expedited freight service is normally expensive, except for bulk freight for low-cost volume (ore, coal, grains) and volume shipments (unit train operations). Trend is towards trucks for short-haul and rail ("piggyback") for long-haul.

Production freight amount substantially exceeds that of distribution freight as fuels are consumed, raw materials lose considerable volume and weight. The trend is towards zero-inventory with smaller, more frequent shipments & deliveries with increasing usage of motor freight. Smaller shipments lead to increased need for assembly and disassembly operations with more handling and clerical works (Fair and Williams 1975).

Freight transportation is commonly measured and described by either commodity movements or vehicle movements. Freight demand exists across the entire network from suppliers and producers to processing plants to intermediaries (inventory, warehouses, distribution centres) to attraction installers, retails and customers. Since demand is derived from the socio-economic production system with respect to land-use production and the type of commodity determines the suitable transport modes, the primary focus of freight demand modelling should be commodity movements (Regan and Garrido, 2000).

Freight movements result from decisions made in various sectors of the economy concerning production, consumption, and sales that have little to do with transportation per se. However, freight demand research lags behind that of passenger due to several factors including public policy concerns, the lack of efficient methods and tools to model and solve large-scale problems, the difficulty of identifying the decision-makers involved in the process and the primary role of private sector agents who are necessarily concerned about confidentiality and reliability. Data collection problems due to proprietary freight data that is aggregate in geographic scale, difficult to collect from private organisations (Regan and Garrido, 2000) and limited opportunities for extensive roadside interviews (Ortuzar and Willumsen, 2002).

The pricing of freight services influences the level of S in (1) affected by factors including length of contracts, extent of volume discounts, availability of terminal facilities, use of own-account operations (transport, image, reliability, integration), fitness of modes to transport certain commodities and suitability of hierarchical transport systems (Ortuzar and Willusem, 2002).

A freight demand model should be responsive to various influences including macroeconomic factors, demographic trends, socio-economic production dynamics, land-use attraction, government policy, freight logistics practices, transportation infrastructure characteristics and technology (Shankar and Pendyala, 2000). Modelling process proceeds from fundamental principles, parameter identification, model development, calibration, verification, validation (technical, operational and dynamic) and application.

Effective freight demand modelling should incorporate the factors affecting commodity movements including locations, the range of commodities, physical nature (bulk, packaged, secure and refrigerated), operational nature (firm, shipper, carrier), geographical nature (distance, impedance, density), dynamic nature (variations of needs and preferences), pricing (rates are proprietary, flexible and subject to bargaining and negotiations) and data issues (Ortuzar and Willusem, 2002).

Demand modelling should fulfil the following requirements (Boyce, 1998) for effective assessment and usage with respect to freight:

1. Choice variables: complete modelling of all variables in (1)
2. Exogenous variables: Shipper and carrier choices dependent on the above choices with regard to travel times and costs on the networks, demand for inventory and access to destinations
3. Endogenous variables: model solutions appropriate for inputs to land-use and transportation systems as well as interfacing with passenger models for multi-modal activities as in temporal, spatial and structural transferability (Shankar and Pendyala, 2000)
4. Parameter estimation: model capable of being calibrated with available data & validated independently
5. Computing platform: solution using available computing platforms to professionals within acquisition costs and time requirements

Methodological issues for freight behaviour theory (Mahmassani, 2001) are:

1. Identifying the relevant principal choice dimensions in (1) and associated locus of decision-making of a specific, but uncertain freight context (Zijpp and Heydecker, 1996)
2. Determining the appropriate level of disaggregation in freight movement analysis because aggregation over larger scales can cause serious aggregation biases in parametric effects and disaggregate demand models can be suitably aggregated (Shankar and Pendyala, 2000)
3. Importance of optimisation tools and models that reflect shipper/carriers’ behavioural considerations and integrity of freight supply-demand network because of private, competitive operations
4. Integration of freight demand considerations within broader logistic chain

Data issues for freight behaviour theory (Mahmassani, 2001) are:

1. Firm level data: flows, decision frameworks, firm characteristics and activities
2. Collection of data associated with various types of transactions
3. Advanced technologies for passive & unobtrusive data collection: GPS, AVI, EDI
4. Serious data confidentiality, cost and reliability issues
5. Framework for data collection that satisfies firms’ natural reticence simultaneously with the public agencies’ need for data to support planning and policy decisions
6. Flexible & robust basis for data collection, interpretation and aggregation
7. Establish and enforce consistency in combining data from different sources
8. Freight transport behavioural analysis using stated preference, revealed preference and transaction logs with calibration against actual behavioural needs

Freight demand modelling of (1) is inherently difficult, uncertain and incomplete, as explained in previous sections, which can be summarised (Shankar and Pendyala, 2000) as follows:

1. Endogeneity and observational uniqueness of equation systems
2. Unobserved heterogeneity and state dependence
3. Specification and measurement errors and instrumentation
4. Heteroskedasticity and serial correlation

In spite of these, practitioners and researchers have proposed and implemented a wide variety of methodologies under different principles for different applications of different kinds of freight demand, as reviewed by Regan and Garrido (2000) as well as Shankar and Pendyala (2000). A general overview is provided below, followed by specific descriptions of some popular models.

In general, freight modelling can be classified as follows:

1. Trend and time series analysis: using correlation, autoregression and simple growth factor models (ADB, 1988) under constraints suitable for the analysis of time series data and easy to implement, but crude for micro-analysis.
2. Elasticity methods: uses the price elasticity of demand to measure the sensitivity of demand of logistics costs associated with a mode. Useful for performing quick sensitivity analysis w.r.t. to price (ceteris paribus) especially when data is lacking, but crude for micro-analysis
3. Network logistics (network equilibrium) models: Attempting to model decisions of various parties (producers, shippers, carriers and consumers) through a market clearing process for network equilibrium. Assumes conservation of flows throughout the network, that carriers and shippers minimise costs and set prices according to certain rate functions. Generally more complex to implement with great data demands.
4. Spatial price equilibrium models: similar to network equilibrium, but based on market price equilibrium of demand and supply. Complex with big data demands.
5. Spatial interaction models: uses physical interpretation for modelling – production and attraction sources as physical attractions. Hence, freight demand is proportional to land-use and socio-economic activities of the zones and inversely proportional to the impedance. Governed by three fundamental laws – laws of attraction, flow and interaction. Approaches include gravity and network models. Effectiveness depends on level of aggregation, model comprehensiveness and data issues.
6. Economic input-output (econometric) methods: Using economic input and output indicators to determine the levels of economic activity by sector, geographic location and time frame. Inputs include capital, labour and while outputs include industrial production and demand for goods and services. Again, difficulty of model dimension identification and data issues arise.
7. Freight mode split models: made of econometric models (attempts to identify and analyse cause-and-effect and correlative relationships between freight demand and various factors) and network based models (apply optimisation rule to an objective function to forecast freight traffic distribution). Inherits limitations of both models.
8. Direct demand models: one-step forecasting of commodity flow by mode between all O-D pairs using only aggregate data. Useful in the absence of detailed disaggregate data, but cannot capture individual behaviour and mask differences across different units through the type of data aggregation used.
9. Aggregate models: basic unit of observation is an aggregate share of a freight mode at a certain geographical level. Can be derived from optimal behaviour for aggregated market. Two general approaches – total flow approach (estimate total freight volumes by mode) and relative flow approach (share of total freight volume captured by mode). The classical four-step aggregate model with feedback is the most commonly practiced and a general framework for freight has been proposed (Shankar and Pendyala, 2000). Models tend to be empirical and inherit aggregation biases.
10. Disaggregate models: basic unit of observation is the individual firm maker’s choice of a freight mode for a given shipment/haul. Focuses more on behavioural aspects that have not been extensively researched. Similar to aggregate models, uses optimal behaviour for disaggregate firm using utility maximisation. The difference lies only in the type of data used. Common approaches include multinomial logit model approach and inventory theoretic approach. Provides very accurate forecasting grounded in microeconomic theories, conducive to rich empirical specifications and responsive to numerous influencing factors, provided that detailed accurate data can be produced. Limited by extensive data requirement, complexity and computational requirements.
11. Geographical framework: classifies various models according to three broad categories: international, intercity and urban (Regan and Garrido, 2000). Useful due to the uniqueness of the decision makers, forces generating freight movements and technologies involved in each category. However, this is limited in future due to increasing globalisation and considerations across geographical jurisdiction.

Major limitations (Regan and Garrido, 2000) include:

1. Data complexity, incompleteness and uncertainty: collection and preparation of data required by the models
2. Lack of integration between models specific to only one context
3. Rigid assumptions: including decisions independent of other considerations, transport services in each market is fixed and different shipment sizes as competitive with each other
4. No embedment in transportation framework: needed as endogenous solution variables, transferability issues
5. No unified structure for interactions of freight and passenger travels
6. Lack of modelling of new advances: brought about by state-of-the-art developments

Three common freight demand models are described in more details below:

1. Growth factor model (ADB, 1998) for sketch planning:

(2)

where is the forecasted trips or units of shipment for commodity k from origin i to destination j, is the base period trips (assumed, given, measured) with growth factors at origin i and at destination j. Balancing is carried out to meet one of more of the following constraints:

(3)

where a_i^k, b_j^k and D^k are the total trips from origin i, the total trips to destination j and the total trips within the freight network respectively.

 

2. Gravity model (Ortuzar and Willusem, 2002) for aggregate modelling:

(4)

where is the forecasted trips or units of shipment for commodity k from origin i to destination j; O_i^k and D_j^k are supply and demand at zone i (or j); , and are balancing and calibration parameters; the generalised freight transport costs per unit of commodity shipment is:

(5)

where the parameters are as explained in (Ortuzar and Willusem, 2002, 13.2) and the the cost function can be defined as in (Regan and Garrido, 2000, 10).

 

3. Multinomial logit model (Jiang et al, 1999) for disaggregate freight mode choice:

(6)

where P_k is probability for freight mode choice alternative of commodity k under K alternatives in N nested MNL levels, is the calibration parameter associated with n nested level and V_nk is the observed utility associated with n nested level and k alternative.

Using aggregate total international container freight data (ADB 1988, Chapter 4, Appendix I & K), a simple case study for the freight demand forecasting of Asian container freight is carried out with the use of the following gravity model functional form:

(7)

where is the forecasted trips or units of shipment for container freight k (in KTEU/yr) from origin country i to destination country j; O_i^k and D_j^k are supply and demand at country i (or j); and are balancing and calibration parameters; is the per unit generalised container freight transport impedance costs of shipment.

Linearise (7):

(8)

Base year data (O-D matrix and costs) based on 1984 is used for calibration of parameters and . Verification is performed against 1987 data and forecasting carried out for 1990. Results are shown in the attached MS Excel file ADBdata.xls of the relevant years. Calibrated parameters using linear regression on the origin side are and . The forecast results show that the aggregate total agrees well, while the flows do not. This is due to the lack of comprehensiveness of (7) and discrepancy of the data in including all factors involved in container freight.

Various exciting and ambitious research are emerging (Regan and Garrido, 2000). Further research into attenuation of present modelling limitations needs to be conducted. Freight demand models would likely explicitly include carrier or shipper behaviour in order to be more comprehensive with better integrity.

The spatial dimension of (1) would be altered in the new economy where e-commerce and the information explosion would amplify the impacts of intermediaries or third party services (freight forwarders, brokers, facilitators and agents). On one hand, disintermediation driven by ease of information gathering and sharing would cut down the middle-men and lead to failures of third parties. On the other hand, disintegration fuelled by business-to-business e-commerce, emergence of niche markets and direct business approach would benefit third parties.

The emergence of virtual logistics service providers who broker information to shippers, carriers, warehousers and intermediaries needs investigation into the resulting impacts on freight demand. Address how shippers and carriers might co-operate to reduce urban congestion and the development of off-peak operations on freight demand.

Nine likely broad themes (Mahmassani, 2001) are as follows:

1. Linkages between increasing reliance on electronic transactions, logistics & freight demand
2. Impacts of consolidation, cross-border acquisitions and alliances, driven by strategic agreements
3. Increasing intermodalism in freight movement (rail-truck, air and marine container) requiring accurate tracking, records and analysis
4. Examine freight demand of the underlying firm location decisions & land-use pattern
5. Evaluating feasibility, desirability and relevance of infrastructure to policy makers
6. Predict responses of freight parties to various policies for continuing deregulation
7. Regulation of varying vehicle sizes & weights for infrastructure performance
8. Operator responses to regulations for public safety
9. Environmental concerns amidst different and conflicting freight transport environmental performance characteristics

FTD is a derived demand similar to passenger demand, but it is more complex, uncertain with more data collection difficulties involving more parties over more varying time periods and geographical regions. Hence, research and FTD modelling lags behind that of passenger due to the limitations highlighted. A critical review of FTD modelling has been produced with a simple case study using container data under aggregate gravity model. Crude aggregate modelling is used in practice, while choice variables and data reliability issues hamper research development. State-of-the-art developments seek to reduce the present modelling limitations and address integration as well as freight traffic and information management.

Boyce, David. A Guide to Urban Travel Forecasting Models. University of Illinois. Chicago. USA. 1998.

Fair, Marvin L. and Williams Ernest W. Economics of transportation and logistics. Business Publications Inc. 1975.

Jiang, Fei, Johnson, Paul and Calzada, Christian. Freight demand characteristics and mode choice: an analysis of the results of modeling with disaggregate revealed preference data. Journal of Transportation and Statistics. Vol. 2, No. 2. pp. 149-158. 1999.

Regan A.C. and R. Garrido, Freight Demand and Shipper Behavior Modeling: State of the Art, Directions for the Future, Hensher, David (ed) Travel Behavior Research, the Leading Edge, 2001, Pergamon Press, Oxford, pp. 185-216.

Wilson A.G. (1998) Land-use/Transport Interaction Models. Journal of Transport Economics and Policy. Vol. 32. Part I. Pp.3-26.

Zijpp, Nanne J. van der and Heydecker, Benjamin. How many parameters should a traffic model have? University College London. Centre for Transport Studies. England. 1996.