Abstract
The challenge this study addresses involves understanding and modeling the complex issues related to freight transportation, which is crucial because of its significant economic effect. Traditionally, freight demand modeling starts with analyzing freight production (FP) and freight attraction (FA) values. However, traditional transportation modeling methods often struggle to accurately capture and predict freight demand because of complex user behaviors and the involvement of multiple stakeholders throughout the supply chain. Previous research has predominantly focused on developing quantitative demand models, leading to varied perspectives in this field. To address these issues, a comprehensive literature review was conducted to compare and synthesize findings from various studies into a coherent narrative. The FP and FA models were critically examined, including their techniques, key variables, data requirements, and evaluation methods. This literature review highlighted current challenges and proposed future research directions. The key findings from the literature reveal significant insights, such as the weak correlation between freight demand and traffic because of the diverse goods and variations in shipment sizes. This underscores the need for separate freight and trip generation models. Therefore, future research should integrate traditional surveys with emerging data sources and combine conventional statistical models with advanced deep learning techniques, leveraging the strengths of both approaches. This narrative review offers valuable context and insights that can complement systematic reviews or meta-analyses on freight modeling, providing a broader understanding that quantitative analyses alone might overlook.
Freight transportation is a cornerstone of most global economies. In Canada, for instance, every dollar of gross domestic product (GDP) is associated with approximately 0.52 ton-kilometer freight transportation ( 1 ). Similarly, in the United States, the transportation and warehousing sector contributes around 8%–9% to the overall GDP, encompassing road, rail, air, and sea freight transportation along with warehousing and storage services ( 2 ). Beyond its economic contributions, the freight industry is a crucial component of the global supply chain, enabling the movement of essential goods worldwide.
Capturing user behavior is important in developing demand models and transportation planning ( 3 ). In freight transportation, understanding user behavior in demand models is complex because of the involvement of multiple stakeholders with different preferences and constant fluctuations in factors, such as economic conditions, fuel prices, regulations, and technological advancements, which affect market dynamics ( 4 , 5 ). In freight demand modeling, it is essential to differentiate between demand generation and traffic generation. The former arises from economic activities and is measured in weight, volume, or value, and the latter results from logistical choices and is described in truck trips ( 3 , 6 ). Demand generation models have also been referred to as freight generation (FG) ( 3 ) or commodity-based models (7–9) in the literature, and traffic generation models are expressed as the freight trip generation (FTG) models ( 3 ), encompassing trip-based ( 8 , 10 , 11 ) and delivery-based models ( 12 ). Despite the multitude of units and methodologies used for modeling freight movements, all these units can be categorized within the broader realms of FG and FTG ( 13 ). “Generation” encompasses attraction and production, which leads to the division of FG and FTG into freight attraction (FA) and freight production (FP), as well as freight trip attraction (FTA) and freight trip production (FTP) ( 14 ). Figure 1 shows the distinction between FG and FTG more clearly, highlighting differences in their underlying causes, units of measurement, and commonly used terms.
A comparison between the freight generation (FG) and freight trip generation (FTG) concepts.
Considering the large volume of cargo transported worldwide annually, its increasing trend in the future ( 15 ), and undergoing shifts and changes across different sectors and regions, it is important to organize and enhance the literature on freight demand modeling. The main objective of this study is to conduct a narrative literature review and offer a thorough overview of demand modeling in freight transportation. Therefore, after introducing the review methodology in the following section, FG and FTG are discussed in depth, followed by an extensive review of their different aspects, including modeling techniques, utilized variables and data sets, and evaluation approaches. Finally, future research directions for resolving the challenges of freight demand modeling are presented. This study will provide readers with an understanding of the conducted research, identify key findings, draw connections between different studies, and develop a cohesive narrative about FG modeling.
Methodology
This study offers an extensive review of FG and FTG models in urban areas. Previous research in freight demand modeling has primarily focused on quantitatively synthesizing FG and FTG models, underscoring the complex and multilayered nature of this field.
Present practices in freight demand modeling emphasize that the accuracy of FG and FTG models relies on various factors. These include FG and FTG patterns, the selection of modeling structures, the appropriateness of explanatory variables, and the methods employed to convert establishment-level generations into zonal-level generations. Therefore, comparing the outputs of FG and FTG models across different studies has become challenging.
Given the diverse perspectives within freight demand modeling research, this study involves a narrative review to consolidate these varied viewpoints into a coherent narrative. This methodology enables a comprehensive overview of available evidence on freight demand modeling to be provided, presenting practical implications and recommendations. In addition, this narrative review can potentially complement future systematic reviews or meta-analyses on the same subject. It provides additional insights, contextual information, and interpretations that quantitative analyses might overlook.
To identify relevant papers for this review, Google Scholar was used, which focused on papers published after 2017 to capture recent advancements in the field. A comprehensive set of keywords associated with freight modeling was used, including “Freight Generation,”“Freight Trip Generation,”“Freight Demand Modeling,”“Aggregate Freight Modeling,”“Disaggregate Freight Modeling,”“Establishment-level Freight Modeling,” as well as their combinations with “GPS Dataset” and “Trip Estimation.” Recognizing that “generation” encompasses production and attraction, keywords such as “Freight Production,”“Freight Trip Production,”“Freight Attraction,” and “Freight Trip Attraction” were included. In addition, to address the overlap between freight and establishment-based service trips, the search was extended to include “Service Trip Generation” and sector-specific terms (e.g., “Hospital Freight Trips” and “Restaurant Freight Trips”). This keyword expansion strategy ensures broader coverage of the literature and reduces the risk of missing relevant studies.
After compiling an initial list of articles, each paper was evaluated by reviewing its abstract and conclusion sections to assess its relevance to the focus of this study. Papers closely aligned with the objectives were selected for a full review. In addition, a backward and forward reference search was employed, examining frequently cited papers to ensure coverage of important studies and significant contributions in the field. Figure 2 shows the research questions and the details of the methodology employed in each step.
Steps in methodology employed for this review.
Freight (Trip) Generation
Freight (trip) generation is fundamental in modeling freight movements, acting as the initial step in traditional four-step models ( 14 , 16 ). Investigating FG offers insights into establishments’ needs, aids in estimating current freight traffic, and enables predicting future freight movement. This knowledge equips policymakers for more informed decisions, enhancing the efficiency and sustainability of freight transportation ( 6 , 17 ). The concept of generation encompasses production and attraction and can take on distinct interpretations based on the units of analysis ( 18 ). The FG can be computed at macro levels, such as cities or at micro levels, including individual establishments ( 16 ). This determination depends on the available data set and the specific objectives of the study.
When examining freight generation, it is essential to differentiate between FG and FTG. Recent contributions in freight demand modeling encompass various aspects of modeling FG and its influencing factors ( 11 , 14 , 19–25). These works provide FG quantities for FTG calculations in the later stages of demand modeling ( 11 ). In addition, certain studies undertake the dual task of modeling and estimating FG and FTG ( 3 , 6 , 19 , 26 ). A third category focuses exclusively on FTG modeling, investigating diverse techniques and influential factors ( 7 , 13 , 27–31). In their modeling methods, Gonzalez-Feliu and Sánchez-Díaz (13) categorize FG and FTG estimation into four main groups, based on a comprehensive literature review: (1) constant generation rates; (2) linear and nonlinear functional forms; (3) utility functions; and (4) probability-based random generation models. Of note, their study compares constant rates and a selection of linear and nonlinear functional forms for FTG computation.
Modeling Technique
Traditionally, the prevailing method for estimating FG and FTG involves using constant rates per employee or unit of establishment area. As highlighted by Holguín-Veras et al. ( 3 ), this approach may be valid for calculating FG; however, it is often inappropriate for FTG estimation and can lead to calculation errors because there is no direct proportionality between FTG and business size variables in most cases. They also mentioned that when considering various zones with the same number of employees, applying constant rates would yield identical FTG values despite the mix of establishments within a zone. In addition, their research findings demonstrated that only 18% of industry sectors exhibited a consistent FTG rate per employee.
Over the last two decades, researchers have shown growing interest in adopting functional forms for modeling, primarily because of their higher accuracy and ability to simulate establishment behavior more realistically. This approach has proven beneficial for achieving better results and providing more accurate recommendations. Regression models have emerged as the preferred method within functional forms ( 16 , 30 ). In particular, researchers have commonly employed simple linear regression and multiple regression models, utilizing the ordinary least squares (OLS) method to establish causal relationships between dependent and independent variables when estimating FG and FTG ( 3 , 10 , 11 , 13 , 14 , 16 , 19 , 20 , 22 , 25–31).
To explore the effect of various classification systems on FP and FTP patterns, Pani and Sahu ( 32 ) turned to hierarchical linear modeling (HLM), a more complex form of OLS. In addition, Doustmohammadi et al. ( 29 ) utilized stepwise linear regression (SLR) and Bayesian linear regression (BLR) in Birmingham to formulate truck trip equations, although their findings revealed no discernible distinctions between these two methods. To ensure accurate results, adhering to certain fundamental assumptions when using the OLS method is crucial. These assumptions encompass the need for a linear relationship between the dependent and independent variables, randomness in the data set, avoidance of strong correlations between independent variables, adherence to a normal distribution for the error term, consistency in error variances (homoscedasticity), and independence among the error terms ( 14 , 20 ). To tackle nonlinearity between the dependent and independent variables, most studies utilized logarithmic transformations on the dependent variable, independent variable, or both ( 7 , 11 , 13 , 19 , 21 , 22 , 27 , 32 ).
In cases where OLS assumptions are violated, researchers explored alternative modeling methods. In papers where they face issues, such as heteroskedasticity and non-normality in the error term of the OLS, they used the generalized linear model (GLM) ( 7 , 27 ), generalized linear regression model with a log link (GLML), or the gamma log generalized linear regression model (GLGLM) ( 20 ). Recognizing the presence of outliers and deviations from model assumptions, Sánchez-Díaz employed a robust linear regression method with sandwich estimators when estimating FTA parameters. This approach accounts for these deviations and makes parameter estimates more reliable ( 6 ).
In addition, when modeling FG and FTG across geographic areas, spatial units introduce spatial interdependencies that should not be overlooked. To address this challenge and improve model accuracy, some studies incorporate spatial modeling techniques, such as the spatial lag model (SLM) and the spatial error model (SEM) ( 11 , 26 ). Novak et al. ( 21 ) employed the spatial autoregressive model (SAR) and the spatial moving average model (SMA) to enhance their models’ fit.
Researchers have explored alternative methods for modeling FG and FTG in recent studies. One commonly employed alternative is multiple classification analysis (MCA). Researchers have found that MCA offers several advantages: it is easier to apply than OLS, provides slightly better estimation power because of increased degrees of freedom, and is more straightforward in its implementation ( 7 , 14 , 25 , 31 ). In addition, Kulpa ( 16 ) developed regional truck trip generation equations using trip generation rates, multiple regression, and artificial neural networks (ANN). Their results indicated that multiple regression and ANNs can yield more accurate predictions than trip generation rates, depending on the number of independent variables. The ANN’s strength lies in uncovering latent relationships among dependent and independent variables. However, an ANN lacks functional forms, such as regression equations, and requires a substantial sample size, which are considered disadvantages. Furthermore, Lim et al. ( 22 ) evaluated and compared seven machine learning (ML) algorithms with the OLS method. These algorithms included Lasso, decision tree regression (DTR), random forest regression (RFR), gradient boosting regression (GBR), support vector regression (SVR), Gaussian process regression (GPR), and multilayer perceptron (MLP). The ML algorithms introduce complexity to the modeling process, and they enhance modeling performance. The results of this study demonstrated that alternative ML models can reduce the root mean square error (RMSE) by up to 30%. Balla et al. ( 14 ) also categorized modeling techniques into parametric and nonparametric. They conducted a comparative analysis to evaluate the effectiveness of various methods in modeling FG at different levels, including industrial segments, regions, and states. Parametric methods, such as OLS, weighted least squares (WLS), seemingly unrelated regression (SUR), and robust regression (RR), were contrasted with nonparametric approaches, including MCA and SVR. They noted a distinct advantage of nonparametric methods, highlighting that these approaches do not require information about the distribution of the data set. Their research findings demonstrated that nonparametric methods outperformed parametric methods in predicting FG. Figure 3 shows a summary of the discussion on the methodology of freight demand modeling.
Summary of methodology details in freight demand modeling.
Variables
Selecting suitable variables is critical in modeling freight demand to estimate FG and FTG, especially when specifying functional forms. The choice of variables depends on the nature and scope of available data sets and the study’s aim. Variables typically fall into two categories: (1) dependent variables, representing the outcomes of interest; and (2) independent variables, which act as predictors. In this section, both variables will be explored, examining how previous studies have approached variable selection.
Dependent Variable
The dependent variable, often referred to interchangeably as the response variable ( 11 , 14 ) and explained variable ( 13 ), is the focal point of investigation and measurement in experimental or analytical studies and represents the desired outcome or effect of interest to researchers. This variable is the core distinction between FG and FTG models, with FG models typically featuring weight, volume, or value, as the dependent variable, with units such as tons per year ( 14 , 19 , 20 , 22 , 25 ), tons per day ( 23 ), cubic meters per employee ( 6 ), million dollars per year ( 22 ), and INR 10 million per year ( 11 ). In contrast, the FTG models utilize the number of trips as their dependent variable, employing units such as trips per week ( 6 , 7 , 27 , 28 , 30 ) and trips per day ( 3 , 16 , 26 , 31 ) in previous analyses.
Independent Variable
Selecting the independent variables in freight demand modeling is a crucial step that depends on several factors, including the study’s scope, objectives, and data availability ( 13 , 27 ). These variables are often referred to interchangeably as independent, explanatory, predictor, or input variables, and they are the basis for calculating the dependent variable. The unit of analysis in FG and FTG modeling (e.g., zone-level or establishment-level) influences the selection of independent variables. For example, zonal FG or FTG models based on aggregated data often emphasize socioeconomic indicators, such as population density and the number of establishments. In contrast, establishment-level models typically rely on disaggregated data and emphasize attributes specific to individual businesses or facilities ( 13 ).
In FG models, three types of independent variables have been used. The first group includes human resources, such as employment (which can take different forms depending on sector or industry) ( 11 , 14 , 19–22, 25 , 32 ), population ( 11 , 20 ), annual payroll ( 22 ), and GDP ( 11 ). The second category is composed of industrial resources, including petroleum and electricity consumption ( 19 ). The final group encompasses infrastructure and network variables, such as area (including specific sector areas) ( 11 , 14 , 19 , 20 , 25 ), effect of ports ( 11 , 21 ), length of transportation networks (containing variables such as railways and highways length [ 11 , 21 ]), and the number of establishments and factories ( 11 , 21 ).
For FTG modeling, the independent variables utilized in the previous studies were categorized into three groups. The first group is the number of employees, which was mostly utilized in previous studies of different countries for modeling FTG ( 3 , 13 ). This variable is a reliable representation of the establishment size and is more readily available than other establishment size variables ( 27 , 28 ).
The second group includes variables that exhibit inconsistencies in their utilization in previous studies because of data unavailability, ineffectiveness, or study scope. As an alternative to “employment” for the “establishment size” variable, some researchers used the area of the establishment ( 33 ). Some studies found that employment-based models performed better than area-based ones ( 27 , 30 ). Sánchez-Díaz ( 6 ) used employment and area as independent variables and found that area-based models sometimes outperformed employment-based ones. The combination of those variables was not feasible in a model because of their high correlation. Some studies also introduced spatial variables, such as the width of the front street for each establishment ( 26 , 27 ), land value ( 26 ), land use variables, such as industry sectors ( 6 , 7 , 26 , 27 ), and dissimilarity index ( 28 ). In addition, several papers utilized the number of inhabitants as the independent variable to include regional characteristics in FTG modeling ( 16 , 28 ).
The third group contains variables highlighted by previous studies to enhance the explanatory power of future FTG modeling. For instance, Sánchez-Díaz ( 6 ) recommended the inclusion of sales, storage space, location, and commodity type in addition to employment, area, and commercial sector.
Service Trip Generation and Sector-Specific FTG
Although this study focuses on FG and FTG, it is essential to acknowledge related research on service trip generation because many urban establishment-based trip models analyze freight and service trips simultaneously. Holguín-Veras et al. ( 34 ) estimated service trip attraction models using OLS regression, testing linear, logarithmic, exponential, and power specifications. The dependent variable was the number of service trips attracted per establishment per day, which was primarily explained by employment, with sectoral and geographic effects also considered; the power function emerged as the best-performing model. Gonzalez-Calderon et al. ( 35 ) studied Medellín’s urban establishments, modeling service trip production and attraction using OLS linear and nonlinear forms. Their dependent variable was the number of service trips produced or attracted, with predictors including employment, floor area, vehicles, parking availability, and service duration. These works highlight the close methodological parallels between freight and service trip generation.
In addition, several studies have examined sector-specific FTG. Oliveira et al. ( 36 ) estimated freight trips for pubs and restaurants in Belo Horizonte, applying OLS with area and employment as predictors, and found establishment size to be the strongest explanatory factor, with a coefficient of determination of R2 = 0.81. Oliveira ( 37 ) extended this work to nine Brazilian municipalities, using OLS and Tobit models with employment and floor area as predictors; robust log–log regression yielded the best results across city types. Patil et al. ( 38 ) investigated restaurants in Indian cities, modeling FTA and FTP. Daily trip counts were explained by employment (EMP), seating capacity (SEAT), vehicle ownership (VEH), gross floor area, and interaction terms (EMP × VEH, SEAT × VEH), with Poisson regression outperforming linear regression because of the count nature of the data. Mafla-Hernández et al. ( 39 ) analyzed large urban facilities, including shopping centers, hotels, hospitals, universities, and marketplaces, and applied OLS regression with linear, log–linear, and log–log specifications. Daily FTP and FTA were modeled using predictors, such as beds, rooms, students, freight weight, floor area, and employment, with log–log models generally achieving the highest R2 values (often > 0.90). Finally, Sánchez-Díaz ( 6 ) examined freight demand across diverse urban establishments, highlighting sectoral differences in the influence of employment and floor area.
Both service trips and sector-specific FTG models were not the central focus of this study; this section showed that their modeling approaches, dependent variables, and independent variables are consistent with the broader FTG literature, which was presented in previous sections.
Data Set for Freight (Trip) Generation Modeling
Acquiring a relevant and comprehensive data set is a prerequisite for effectively modeling and exploring FG and FTG. Freight demand modeling has gained increased attention recently ( 6 ). Numerous studies have identified the lack of appropriate data sources as a primary obstacle to progress in this area ( 3 , 6 , 14 , 16 , 20 , 25 , 28 , 33 ). The shortage of suitable data can be attributed to various factors, including the proprietary nature of freight-related information ( 13 , 14 , 25 ), the absence of a well-defined framework for systematic data collection ( 14 ), and the inherent challenges and expenses associated with data gathering efforts ( 3 , 17 ). Therefore, innovative approaches and strategies are required to overcome these limitations.
Several developed countries have recognized the significance of periodic commodity flow surveys (CFSs) in advancing their understanding of freight movements and estimating freight demand. In the United States and South Korea, CFS initiatives are conducted at 5-year intervals, providing a wealth of data at the national, state, and metropolitan levels that detail the origins and destinations of commodities (21–23). Meanwhile, certain European nations, such as Belgium and the Netherlands, utilize an annual collection of commodity flow data, where cities are treated as zones for analysis ( 20 ). In Japan, the Tokyo Metropolitan Area Freight Survey is conducted every decade, focusing on the Tokyo region ( 10 , 40 ). In contrast, developing countries, such as India, face significant challenges in freight demand modeling, primarily because of the need for a comprehensive CFS database ( 11 , 14 , 25 ).
For the spatial scale of FTG models, Gonzalez-Feliu and Sánchez-Díaz ( 13 ) have categorized them into three distinct classes based on their scope, as identified via their review of existing studies. These categories encompass the macroscopic level, which deals with broader areas or entire cities; the mesoscopic level, which focuses on neighborhoods or specific streets; and the microscopic level, which zooms in on individual establishments. The spatial scale chosen for the model dictates the need for aggregated or disaggregated data. Aggregated data is sufficient when modeling freight demand at the city or zonal level. However, for establishment-level demand modeling, the availability of disaggregated data becomes crucial. Of note, aggregated data at finer scales may introduce aggregation biases; however, employing disaggregated data provides a more comprehensive and error-reducing approach ( 13 , 22 , 41 ). Striking an appropriate balance between these considerations is pivotal to ensuring the accuracy and reliability of FTG models across different spatial scales.
Recent research on FG and FTG has employed a range of data sources that can be classified into distinct categories. Some studies have utilized establishment-level data acquired from establishment surveys, offering fine-grained, disaggregated insights ( 6 , 7 , 13 , 14 , 20 , 25–28, 31 , 41 ). Others have relied on aggregated data at different spatial scales, encompassing zones, regions, or states ( 10 , 16 , 21 , 24 , 29 ), and some have focused on aggregated data organized by industry sectors ( 22 , 23 ). When primary data sets, such as CFSs or establishment surveys, were unavailable, certain studies turned to secondary data sets as an alternative to estimate freight demand ( 11 , 19 ). These versatile approaches underscore the remarkable adaptability of researchers in overcoming data limitations.
Sánchez-Díaz ( 6 ) categorized the data sets from previous studies into four groups: (1) traffic counts; (2) secondary sources; (3) transport operator data; and (4) establishment-based surveys. He discussed the advantages and disadvantages of each data set for modeling freight demand. According to this study, establishment-based surveys lack route information and are not derived from traffic-related data analysis. In addition, a common drawback among the other three data set types is the absence of a direct connection between freight trips and the originating establishments ( 6 ). Unlike the limitations associated with transport operators’ data, recent research has seen the development of heuristic-based models capable of accurately identifying the origins and destinations of trucks using GPS data sets ( 4 , 5 ).
Freight demand modeling, an important aspect of transportation and logistics, necessitates the integration of data from various sources to enhance estimation accuracy. However, access to emerging and innovative data sets is impeded by a multitude of challenges, including organizational, privacy, technical expertise, and legal issues. Future research should shed light on these challenges and propose potential solutions for future research, advocating for using these data in freight demand modeling ( 42 ).
Stratifying Models
Improving estimation accuracy in modeling is a crucial concern, and various strategies exist to enhance results. One strategy involves stratifying the data and the model, an essential step for modeling freight demand because it helps to distinguish the influential factors of each category on FG and FTG ( 19 ). In addition, categorizing can reduce estimation errors because the members of each category usually have similar characteristics ( 32 ). Several studies emphasize the significance of economic activity-based classification because it simplifies result interpretation. Moreover, economic activities can reflect FG and FTG behavior ( 30 ) and are the source of freight movements ( 32 ). The choice of classification method depends on the model’s objectives, often linked to the data collection purpose and the available classification systems across different countries ( 13 ). This section explores existing categorization methods within three groups: (1) economic activities and commodity type; (2) vehicle type and loading unit; and (3) time of day.
Stratifying data sets and models based on industry types or economic activities is a widely used approach. The International Standard Industrial Classification (ISIC), presented by the United Nations, is one of the most popular classifications. In this system, industries are classified according to a four-digit hierarchical framework. The ISIC framework encompasses a range of classification levels, starting with 21 sections at the broadest level and extending to 432 classes at the most detailed level. Some reviewed studies used 13 two-digit ISIC classifications to categorize their data set ( 14 , 25 , 32 ). In addition, each country or region has a classification framework, mainly aligned with the ISIC classification system. For example, some of the studies have used the National Industrial Classification system ( 11 ), which is used in India to compile data on employment ( 32 ). Park and Hahn (23) used the Korean Standard Industrial Classification. They first divided their data set into 30 categories and then merged some similar categories to have enough observations for each category for FTG modeling.
Category classification systems vary throughout European countries. For instance, Sánchez-Díaz ( 6 ) used the Swedish Standard Industrial Classification to divide similar economic activities into five categories for modeling urban FTG. In Belgium, Mommens et al. ( 20 ) utilized the Belgian Standard Goods Classification for Transport Statistics and defined 10 commodity types for modeling FG generation. In France, Gonzalez-Feliu and Sánchez-Díaz ( 13 ) utilized Nomenclature of Activities Françaises codes to create hierarchical industrial categories at three different levels, consisting of 8, 27, and 43 categories. They aimed to explore how the aggregation level and category construction affect the quality of FTG estimation. Although these classification codes differ across European nations, most are rooted in the European Community’s Classification of Economic Activities (NACE [ 13 ]). In addition, the NACE code seamlessly corresponds to ISIC at the four-digit level ( 32 ).
The standard economic activity classification code in the United States, Canada, and Mexico is the North American Industry Classification System (NAICS). This hierarchical system spans 2–6 digits, offering five levels of detail. Researchers have employed varying levels of detail within the NAICS codes, depending on the objective of their studies, to categorize their data sets by economic activities ( 21 , 22 , 26 , 29 ).
Despite the existence of the previously mentioned codes for economic activity classification, some studies used outdated codes, local codes, or defined their own categories for classification. Holguín-Veras et al. ( 3 ) used Standard Industrial Classification codes used in the US before defining NAICS codes to divide New York City’s establishments into eight categories. In 2014, Rwakarehe et al. ( 24 ) utilized 22 commodity types based on the Standard Classification of Transported Goods, which is a classification code for Canada. Alho and Silva ( 7 ) also utilized 10 categories based on the City Council Establishment Industry classifications in Lisbon, Portugal, to estimate urban freight parking demand based on the FTG. Puente-Mejia et al. ( 30 ) developed a classification system consisting of 10 categories based on the primary economic activity of shops in Quito, Ecuador. They mentioned that this classification method helps to overcome the complexity of using national standard classification systems, such as NAICS and NACE. Middela et al. ( 19 ) categorized their data set into five groups using specific criteria: (1) joint products; (2) substitutes and complements; (3) demand elasticity; (4) transport mode; and (5) processing degree. They emphasized the need for region-specific commodity grouping because of variations in commodity types and transportation methods. However, using global industry classification codes, such as the ISIC and NAICS, will enhance comparability among worldwide freight demand studies ( 32 ).
Another categorization approach involves classifying based on the type of vehicle. In cases where relevant data is accessible, some studies have incorporated vehicle types alongside industry categories. In a study by Kupla ( 16 ), regional-level modeling of FTG involved stratified models based on commodity type and two vehicle categories (e.g., light and heavy). In addition, De Bakhshi et al. ( 28 ) employed stratified models by categorizing vehicle types into four groups: (1) motorized freight trips; (2) nonmotorized freight trips; (3) motorized two-wheeler freight trips; and (4) motorized freight non-two-wheeler trips. To estimate generated and attracted freight volumes in greater detail, Mommens et al. ( 20 ) categorized their models considering nine loading units, including solid bulk, liquid bulk, containers, other containers, pallets, slings, mobile units, other mobile units, and other, instead of vehicle type.
Although most studies have used the previously mentioned methods, other factors exist for stratifying the model. For instance, Demissie and Kattan ( 5 ) utilized five time periods to develop stratified destination choice models in Calgary, Alberta. Because GPS points within GPS trajectory data include time stamps, this allows us to explore stratification based on the time of day to model the FTG.
Model Evaluation
In the literature, several metrics are available for evaluating different aspects of models. This section classifies these metrics into two groups and reviews existing studies for each group. The first group focuses on evaluating a model’s goodness of fit by assessing how effectively independent variables explain variance in the dependent variable. Many studies employing regression methods for modeling have employed the R-squared (R2) metric for this ( 3 , 7 , 10 , 11 , 14 , 16 , 19–22, 24 , 25 , 27 , 29 , 30 , 43 ). The R2 values range from zero to one, where zero signifies that the model explains none of the variability, and one indicates a perfect fit where the model accounts for all variability. A low R2 is a common challenge in freight modeling ( 20 ); some studies have achieved satisfactory R2 values ( 3 , 19 , 25 ).
It is important to note that increasing the number of independent variables can artificially inflate R2 without necessarily enhancing the model’s explanatory power because it reduces degrees of freedom. To address this concern, some researchers have turned to the adjusted R2, a modification of the standard R2 that considers the number of independent variables ( 10 , 23 , 26 , 28 , 31 ). In situations involving nonlinear relationships or binary outcomes, such as those described by Sánchez-Díaz ( 6 ), Pseudo R2 is the preferred metric for assessing model goodness of fit. In addition, Gonzalez-Feliu and Sánchez-Díaz ( 13 ) utilized the Pearson coefficient to assess model fit and noted that it represents the R2 metric in a bivariate linear regression.
The second group of metrics relates to validating results, a critical step involving comparing estimated values with observed values using absolute measures ( 24 ). In previous studies, the mean square error (MSE) and its more easily interpretable counterpart, the RMSE, which is in the same unit as the dependent variable, have been widely employed ( 3 , 6 , 7 , 10 , 11 , 14 , 22 , 28 , 31 ). Some investigations have incorporated the percentage root mean square error (PRMSE), expressing RMSE as a percentage of the average actual value ( 24 , 29 ). In addition, the mean absolute error (MAE) and its variation, the mean absolute percentage error (MAPE), have been utilized for validation purposes in select studies ( 7 , 13 , 22 , 26 ). In some instances, the sum of squared errors (SSE) has been applied to select the most appropriate model among several generated using various techniques. A lower SSE signifies a superior fit for the model ( 11 , 21 ). Furthermore, the Akaike information criterion (AIC) has served as another valuable tool for comparing different models and assessing the relative quality of one model compared with others ( 11 , 26 , 32 ).
To consolidate the information from reviewed studies, a comprehensive overview is given in Table 1. This table details key aspects of each paper, including the type of demand model (FG/FTG), the study area, and the method of analysis. In addition, it categorizes independent and dependent variables utilized in each study and specifies the employed validation methods. The last two columns of Table 1 provide additional insights into establishment categories in the supply chain, survey types, and the type and granularity of data used. Freight (trip) generation modeling primarily relied on establishment-based surveys focused on shippers and receivers, with limited attention given to other supply chain actors. Although most studies used data collected primarily for their research, secondary data have been used to supplement primary data and for validation and trip generation model expansion. These data were often aggregated at traffic analysis zones, regional, or state levels to develop the FTG models. This summary aims to furnish a clear and organized comparison between the approaches across the reviewed studies, facilitating an understanding of the similarities and differences in their frameworks.
Summary of Methodological Approaches and Frameworks in Selected Studies
Note: FG = freight generation; FTG = freight trip; MCA = multiple classification analysis; OLS = ordinary least squares; HLM = hierarchical linear modeling; SLR = stepwise linear regression; SLM = spatial lag model; SEM = spatial error model; ANN = artificial neural networks; SVR = support vector regression; GBR = gradient boosting regression; GLM = generalized linear model; GLGLM = gamma log generalized linear regression model; GLML = generalized linear regression model with a log link; DTR = decision tree regression; RFR = random forest regression; GPR = Gaussian process regression; MLP = multilayer perceptron; SAR = spatial autoregressive model; SMA = spatial moving average model; WLS = weighted least squares; RR = robust regression; SUR = seemingly unrelated regression; ANCOVA = analysis of covariance; BLR = Bayesian linear regression; FA = freight attraction; FP = freight production; SSE = sum of squared errors; AIC = Akaike information criterion; RMSE = root mean square error; MAE = mean absolute error; MSE = mean square error; MAPE = mean absolute percentage error; PRMSE = percentage root mean square error; R2 = coefficient of determination; BIC = bayesian information criterion; APE = absolute percentage error; AAPE = average absolute percentage error.
How to Improve Freight (Trip) Generation Models?
The FG modeling is important for understanding and predicting the movement of goods, which supports infrastructure planning, economic development, and environmental management. However, the field faces significant challenges, especially the shortage of detailed data and the inadequacy of existing models to predict freight demand accurately. To address these issues, three main future research directions are proposed: (1) incorporating emerging and innovative data sets; (2) applying contemporary machine (deep) learning models; and (3) the development of FTG models with robust temporal and spatial transferability.
Emerging and Innovative Data Sets
The freight industry represents one of the most promising sectors, driven by advancements in the Internet of Things, with sensors installed on cargo, vehicles, and infrastructure. These sensors can provide insights into freight movements and environmental effects. This growing wealth of data includes various collection methods, sources, and entities, categorized into traditional agency data, opportunistic data, and crowd-sensing data. Table 2 summarizes various emerging and innovative data sets that can be utilized to improve freight demand modeling.
Summary of Emerging and Innovative Data Sets for Freight Demand Modeling
Note: GPS = global positioning system; CCTV = closed circuit television.
Traditional agency data, such as road-mounted and aerial sensors, are typically owned and operated by local or provincial authorities to monitor the movements of freight vehicles. In contrast, telematics data, such as GPS trajectory information and data from freight marketplaces, are predominantly collected and maintained by the private sector. Often proprietary, these data sets are not openly shared with researchers and planners ( 51 ). Opportunistic data sets, which are information collected for one purpose but can be repurposed for another, offer valuable insights in this context. For instance, communication media data (e.g., Call Detail Records) serves as a proxy for understanding the movements of people and goods ( 51 ). Crowd-sensing data analyzes social media posts and trends, revealing shifts in consumer behavior and demand spikes that influence freight movements. A promising future research direction involves integrating these diverse data sources to advance our comprehension and management of freight logistics, enabling more efficient and responsive transportation systems.
Among the emerging and innovative data sources, vehicle telematics data has been the most explored for freight and passenger demand modeling ( 50 ). Vehicle telematics data refers to the information collected from vehicles using telecommunication and informatics technology. This data encompasses various aspects of the vehicle’s operation, location, condition, and usage, which can be analyzed to improve efficiency, safety, and overall fleet management. The growing proportion of passenger and freight vehicles generating telematics data has increased the potential to improve the planning and operation of different transportation services ( 48 , 52 ). Vehicle telematics data relevant to freight demand modeling can be classified into five categories, as listed in Table 3, which is based on information from the literature ( 50 , 53 ). Each category describes one or more vehicle characteristics associated with the telematics device.
Summary of Telematics Data for Freight Demand Modeling
Note: GPS = global positioning system.
Freight demand fluctuates throughout the year, leading to variations in freight demand production and attraction across seasons. To comprehend and study the dynamics of freight demand, it is important to acquire frequent and accurate data. However, this is a challenge when using traditional survey data, necessitating the exploration of alternative data collection methods. Several studies have explored the potential of using vehicle telematics data, especially GPS tracking data, to enhance passenger and freight movement planning and operations ( 5 , 52 , 54 ). Table 4 outlines the primary benefits of leveraging large-scale GPS trajectory data, highlighting opportunities while cautioning researchers about challenges associated with integrating GPS data into freight demand modeling. Researchers have utilized GPS trajectory data in various studies to identify loading and unloading locations of freight vehicles and to analyze trip start and end times, travel durations, distances traveled, and specific routes taken ( 4 , 5 , 55 , 56 ). However, raw GPS trajectory data alone cannot capture the essential characteristics of freight movement; instead, these insights often necessitate the application of mathematical models and the development of algorithms for inference. In addition, GPS trajectory data often lacks secondary information necessary to validate critical trip details, such as vehicle stops and trip purposes ( 4 ). This absence of contextual data complicates the interpretation and reliability of insights derived solely from GPS data. Furthermore, several studies have explored repurposing GPS trajectory data initially collected for other purposes (e.g., geofence alerts for parking locations or travel near fuel stations) to accurately extract crucial details, such as freight vehicle trip purposes, origins, destinations, and other pertinent data. Typically, supplementary sources, such as trip diaries and points of interest, are required to validate and enrich the insights derived from GPS data ( 4 , 57 ). Future studies should integrate GPS trajectory data with other telematics data, as listed in Table 3, and incorporate emerging data sets from Table 2 to enhance the robustness of the FA and production models. Roorda ( 58 ) presented five alternative data collection frameworks to study urban goods movements. This study suggested that conducting a national shipper-based survey and periodically purchasing vehicle tracking data from third-party providers represent a preferred combination because of their complementary strengths for nationwide performance benchmarking of urban goods movements.
Advantages and Disadvantages of Global Positioning System (GPS) Trajectory Data Compared with Survey Data
Contemporary Machine (Deep) Learning Models
Many studies have utilized the OLS model to assess the effects of various explanatory variables on the demand for freight ( 3 , 10 , 11 , 13 , 14 , 16 , 19 , 20 , 22 , 25–31). The OLS models have been enhanced by applying spatial regression models to capture spatial dependencies ( 11 , 26 ). Various studies have utilized machine (deep) learning models to examine freight transportation and logistics ( 63 ); only a few have explored their application in freight (trip) generation modeling ( 16 ). Lim et al. ( 22 ) provided a comparative study demonstrating that ML models tailored to specific industries significantly outperform traditional regression models in predicting FG, highlighting the importance of domain-specific customization in freight modeling. Venkadavarahan and Marisamynathan ( 64 ) focused on developing FTG models that incorporate ML-based establishment classification to improve predictions. The comparative strengths of different ML architectures depend on the modeling context. For instance, Al-Deek ( 65 ) found that neural networks outperform multiple regression models in freight planning at seaports, particularly when data exhibit multicollinearity and nonlinear trends. Similarly, Lim et al. ( 22 ) demonstrated that random forest and eXtreme gradient boosting (XGBoost) models deliver higher accuracy and robustness than linear models because of their capacity to manage outliers and heterogeneous data. Zhang et al. ( 66 ) introduced a deep learning architecture that integrates pseudo people flow with agent-based modeling, showing the potential of hybrid models to bridge human activity data with freight logistics. Nampalli and Adusupalli ( 67 ) implemented gradient boosting and ensemble models for route optimization and demand prediction in road and rail freight, noting their effectiveness in handling large-scale operational data. These studies indicate a growing trend toward the use of deep learning and ensemble approaches that can learn high-dimensional, nonlinear relationships. However, deep learning architecture is more powerful; however, it requires larger data sets and longer training times, which may be limiting in some contexts. Therefore, model choice often hinges on the availability of high-quality data and the need for interpretability versus accuracy.
Recently, various methods have shown success in predicting passenger transportation demand, including classical time series models, such as moving average and autoregressive integrated moving average, ML techniques such as XGBoost ( 68 ), and deep learning architectures including convolutional neural networks (CNN), long short-term memory (LSTM) networks, feedforward neural networks (FNN), and graph neural networks (GNN) ( 54 , 69 ).
Deep learning models, such as LSTM and CNN, are feasible candidates for predicting freight demand. With the rapid growth of big data, historical freight demand values across different time stamps can be the input for deep learning models to make short-term predictions. The LSTM is particularly suitable for modeling sequences with long-range dependencies, which is important in freight demand analysis where patterns vary over time intervals spanning hours to months. However, CNNs excel at capturing spatial patterns in data. Combining LSTM and CNN can capture complex interactions between time-varying freight demand patterns and spatial factors ( 70 ).
Understanding local and global spatial dependencies is important for accurate freight demand prediction. The CNN models can effectively capture local spatial dependencies, influenced by nearby zones’ trip generation and attraction roles. In contrast, models, such as transformers, are good at capturing global spatial dependencies in freight transportation, including regional economic conditions and citywide transportation policies ( 71 ). As data availability and computational resources improve, these architectures will probably become the standard in freight demand modeling because of their superior predictive performance and scalability.
Development of FTG Models With Robust Temporal and Spatial Transferability
Developing FTG models requires significant investments in data collection, analytical expertise, and time, which makes their transferability an essential consideration. Transferability refers to the ability of a model developed in one context to be applied to another with minimal recalibration while maintaining predictive accuracy ( 72 ). In freight demand modeling, transferability is typically assessed along two dimensions: (1) spatial transferability, which evaluates applicability between regions; and (2) temporal transferability, which examines stability over time ( 73 , 74 ).
Spatial Transferability
Spatial transferability studies assess whether models calibrated in one region remain valid in another by examining the stability of model parameters. Research has demonstrated that employment-based models tend to transfer more effectively than those relying on land use or business area data because employment is a relatively stable predictor of FTG ( 73 , 75 ). In contrast, land use differences often introduce substantial uncertainty in freight activity patterns ( 76 ).
Several approaches have been proposed to evaluate spatial transferability, including direct model application in a new region, re-estimation using pooled regional data, and Bayesian updating techniques. Direct transfers frequently perform poorly unless regional characteristics such as freight regulations, logistics infrastructure, and commercial density, are accounted for ( 75 ). Segmenting models by industry sector further improves adaptability, particularly in regions with distinct economic structures.
Temporal Transferability
Temporal transferability presents additional challenges in freight demand modeling because of fluctuations in economic conditions, regulatory changes, and technological developments. Longitudinal analyses have shown that variables, such as employment and revenue remain, relatively stable over time; however, others including shipment frequency and vehicle utilization, are highly sensitive to economic cycles and external shocks ( 74 ). For instance, models calibrated during stable economic periods often fail to predict freight activity accurately during downturns or regulatory shifts.
Methods for evaluating temporal transferability include time series econometric modeling, longitudinal parameter comparisons, and structural equation modeling. The findings suggest that models developed before major disruptions, such as economic crises, typically require substantial recalibration to maintain predictive accuracy ( 74 ). These limitations highlight the need for dynamic modeling approaches capable of adapting to evolving freight patterns rather than relying on static calibration.
Evaluation and Limitations
Transferability is typically assessed using statistical and econometric validation techniques, such as R2, RMSE, and MAPE, to evaluate predictive accuracy ( 43 , 73 , 74 ). Parameter stability tests and comparative error analysis are also commonly employed to assess deviations between transferred and locally calibrated models. To enhance transferability, Bayesian updating and combined transfer estimation methods have been proposed, which allow refinement by incorporating limited local data ( 43 ).
Despite these advances, limitations persist. The direct transfer of models often underperforms without adjustments for regional or temporal variability, and the stability of key parameters cannot always be guaranteed. In addition, the heavy reliance on static models poses challenges in contexts where freight activity is subject to rapid structural or regulatory changes. These issues underscore the need for flexible modeling frameworks that can adapt to varying contexts while minimizing the costs of recalibration.
Conclusions
This study addresses the challenge of understanding and modeling the complex issues associated with freight transportation, which is critical because of its significant economic, social, and environmental effect. There is a close correlation between the number of passenger trips generated and the corresponding vehicle trips because of relatively fixed vehicle occupancies (1.25–1.50 passengers per car) in passenger transportation ( 77 ), the relationship between demand and traffic is much weaker in freight transportation. This is because of wide variations in the freight industry, which is influenced by different factors, such as commodity type and shipment sizes. Therefore, it is crucial to treat demand and traffic as distinct concepts as emphasized in the literature ( 3 ). Therefore, an extensive literature review was conducted on the scope of freight demand modeling, containing the FG and FTG models within urban areas. Different modeling techniques were critically investigated, and variables and data sets were utilized in FG and FTG modeling with methodologies of evaluating the performance of models to understand the potential research opportunities.
Three future research directions were identified that could enhance FTG modeling. The first direction is integrating traditional survey data with emerging and innovative data sets. Freight demand exhibits seasonal fluctuations, causing variations in production and attraction. Utilizing emerging data sets (e.g., vehicle telematics data), which offer frequent and extensive temporal and spatial coverage, can provide deeper insights into the dynamics of freight demand. The second direction involves the application of deep learning models. Deep learning has shown promise in demand modeling across various transportation modes; however, its application to freight demand remains underexplored. Several deep learning architectures are candidates for predicting zonal-based freight demand. In addition, combining different deep learning approaches, such as CNNs, LSTM, GNNs, and transformers, may enhance the modeling of complex spatiotemporal dependencies in freight demand. The third research direction is transferability of models. Recent progress has been made in this section; challenges remain in ensuring that models perform consistently when applied across different regions or time periods. The instability of key parameters and the dependence on static formulations limit their robustness, particularly in environments where freight systems are rapidly evolving because of structural or regulatory shifts. Future research should focus on developing adaptable and context-sensitive modeling approaches that enhance reliability while reducing the effort and cost associated with recalibration.
This study provides a comprehensive narrative review that lays the groundwork for future meta-analyses and systematic reviews in the field. The insights and identified research trends highlighted in this study can guide researchers in designing robust methodologies and refining freight modeling approaches. Practically, this study sheds light on the two concepts of FG and FTG, helping policymakers and industry stakeholders achieve a more accurate estimation of the number of trips and demand in freight transportation. This thorough understanding could enhance the result of four-step travel demand modeling and transportation master plans, which could ultimately mitigate congestion and environmental concerns.
Footnotes
Author Contributions
The authors confirm contribution to the paper as follows: study conception and design: A.A., M.D; data collection: A.A., P.A., M.D; analysis and interpretation of results: A.A., P.A., M.D; draft manuscript preparation: A.A., P.A., M.D. All authors reviewed the results and approved the final version of the manuscript.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work is supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) Discovery Grant (RGPIN-2022-03037) and the NSERC Alliance Advantage Grant (ALLRP-2024-598529), which the Alberta Ministry of Transportation and Economic Corridors partially funds.
