Researchers evaluating cross-country trade efficiency have demonstrated that machine learning algorithms trained on publicly accessible structural indicators can reliably estimate national performance scores under the World Bank’s reformed Logistics Performance Index. In an empirical investigation analyzing data across 139 economies, a team led by Eric Torres Ramirez, Percy J. Rosas, Natalie Gamarra-Vargas, Percy Hugo Quispe-Farfán, and Ricardo Fernando Cosio Borda tested whether open macroeconomic and infrastructure measures could approximate scores generated by proprietary commercial tracking pipelines. Their comparative assessment of eleven distinct regression algorithms established that tree-based ensemble models, particularly the Extra Trees architecture incorporating geographic and income classifications, achieved the highest predictive precision and generalization across international supply chain environments.
The investigation centers on a fundamental shift in international benchmarking that occurred during the 2023 revision of the Logistics Performance Index. For editions published between 2007 and 2018, the benchmark relied on subjective survey assessments completed by international freight forwarders and logistics professionals. While informative, those earlier editions faced academic criticism regarding subjective respondent bias, inconsistent year-to-year ranking movements, survey fatigue, and persistent data gaps in smaller economies. To address these limitations, the 2023 revision replaced qualitative surveys with proprietary, high-frequency commercial tracking data. This new methodology computes key performance indicators from sources including TradeLens ocean container tracking, Universal Postal Union parcel monitoring, and Cargo iQ air freight records, measuring objective parameters such as container dwell times, customs clearance velocity, and multimodal transit delays.
Although the transition to event-based tracking significantly improved measurement objectivity, it introduced new structural constraints regarding accessibility, transparency, and geographic coverage. Because the underlying tracking pipelines depend on commercial logistics platforms, the raw transaction data cannot be released publicly for replication, independent validation, or localized diagnostic modeling by third-party analysts. Furthermore, the World Bank intentionally excluded countries where commercial freight tracking is absent, unrepresentative, or concentrated within a very small number of operators. In response to these analytical barriers, the research team designed a modular, reproducible workflow to determine whether open-access structural data from institutions such as UNCTADstat and World Bank Open Data could serve as effective proxies for proprietary logistics indicators without suffering from circular data dependencies.
To assemble an independent feature space, the investigators selected 23 candidate predictors spanning maritime connectivity, trade volumes, economic sophistication, and institutional capacity for 139 countries. Crucially, none of these features were extracted from the proprietary shipment tracking streams used to calculate the official index. Core predictors included maritime indicators such as the Liner Shipping Connectivity Index and container throughput, trade scale variables encompassing merchandise export and import values, and broader structural metrics such as the Economic Complexity Index, the Productive Capacities Index, inward and outward foreign direct investment stocks, and manufactured goods tariff schedules. The researchers noted that these variables reflect long-term capital accumulation, institutional reliability, and administrative throughput that drive national logistics capability.
Data preprocessing required addressing systematic missingness across international reporting bodies, where roughly 5.9 percent of numerical observations were absent. Rather than discarding incomplete records or applying simple mean or median substitutions that distort underlying variance, the investigators implemented Multivariate Imputation by Chained Equations using an iterative scikit-learn framework driven by a Random Forest estimator. Missing data rates were not distributed at random, appearing with greater frequency across low-income economies and specific regions such as South Asia and Sub-Saharan Africa. The modeling pipeline partitioned data into an 80/20 train-test split, executing imputation and RobustScaler normalization inside the training subset before applying fitted parameters to held-out test records to guard against data leakage.
Evaluating algorithms across an initial benchmark under default hyperparameters revealed substantial dispersion in predictive efficacy across distinct mathematical formulations. Across eleven baseline algorithms, test performance exhibited a 63.2 percent divergence in root mean squared error between the best-performing and worst-performing architectures. Tree-based ensembles populated the top performance tier with test coefficients of determination exceeding 0.80 and root mean squared errors below 0.24, led by Extra Trees, LightGBM, Random Forest, and Gradient Boosting. Intermediate performance was observed in kernel and instance-based models including Support Vector Regression and K-Nearest Neighbors, while regularized linear algorithms showed severe limitations, exemplified by Bayesian Ridge yielding negative test coefficients of determination due to rigid linear assumptions.
To explore the influence of structural context, the authors expanded the feature matrix by introducing two categorical descriptors: World Bank income classifications divided into four tiers and geographic groupings covering seven regions. Incorporating these contextual features substantially improved performance across tree-based ensembles, which naturally capture nonlinear interaction effects through recursive tree partitioning without requiring manually specified interaction terms. For the Extra Trees specification, the test coefficient of determination rose from 0.864 to 0.886 while its root mean squared error dropped by 8.3 percent to 0.198. Subsequent hyperparameter optimization using exhaustive grid search yielded a final Extra Trees model achieving a test coefficient of determination of 0.889, a root mean squared error of 0.195, and a mean absolute error of 0.144.
To understand the mechanisms driving model output, the researchers conducted an interpretability analysis using Shapley Additive Explanations derived from cooperative game theory. Global SHAP summary distributions showed that national logistics predictions were predominantly driven by capability and sophistication indicators rather than gross trade volumes. The overall Productive Capacities Index and its information and communications technology component exerted the strongest positive marginal influence on predicted scores, closely followed by outward foreign direct investment stocks and export diversification indices. Conversely, gross trade volumes such as import concentration metrics contributed minimally to overall variation, reinforcing theoretical models that position institutional, digital, and productive capabilities as primary determinants of operational logistics success.
The study also examined the geographic distribution of model errors, analyzing residuals to identify structural limitations. Baseline specifications displayed systematic overestimation tendencies across lower-middle and upper-middle income countries, alongside an upward residual skew in Latin America and the Caribbean. Incorporating income and regional categories mitigated these biases, producing balanced error distributions centered near zero across diverse groups. In holdout evaluations, the optimized Extra Trees architecture matched official values exactly in countries such as Bangladesh, Sri Lanka, and Uzbekistan, while maintaining errors within 0.1 points for economies including Brazil and Chile. However, notable discrepancies emerged in dynamic or shock-sensitive contexts, overpredicting Mexico by 0.4 points while underpredicting Denmark by 0.4 and the United Arab Emirates by 0.3.
As an applied policy demonstration, the researchers used the finalized model to generate provisional logistics performance estimates for thirteen countries excluded from the official 2023 World Bank release due to sparse shipment tracking data. The resulting model-derived scores ranged from 2.44 for Mozambique to 2.92 for Barbados, offering comparative diagnostic benchmarks for omitted economies such as Ecuador, Jordan, Kenya, Morocco, Pakistan, Senegal, and Tunisia. The authors stressed that these predicted scores are intended as provisional reference estimates for trend monitoring, diagnostic screening, and international technical assistance prioritization, rather than direct substitutes for empirical tracking measurements. They noted that future refinements could integrate multi-year panel datasets, formal conformal prediction intervals, and multimodal indicators covering air freight corridors.
During initial exploratory analysis, the investigators identified significant linear associations across several candidate predictors, such as a 0.85 correlation linking gross domestic product to merchandise export and import values, a correlation exceeding 0.80 between inward and outward foreign direct investment stocks, and a 0.73 correlation connecting liner shipping connectivity to container throughput. Despite these strong dependencies, the authors opted against dimensionality reduction techniques, noting that tree-based ensemble architectures inherently accommodate multicollinearity without degrading predictive accuracy. Sensitivity testing regarding reference years for indicators lacking 2023 values also demonstrated stability, altering the coefficient of determination by less than 0.01. Furthermore, specific tuning constraints in the final Extra Trees architecture, such as restricting tree depth to four levels and setting the minimum leaf sample threshold to two, eliminated the gap between training and testing performance, reducing the divergence from 0.114 to 0.007.
Repeated k-fold cross-validation further established model generalizability under varied data partitions, revealing that single train-test splits introduced an optimistic evaluation bias. Across cross-validation folds, Extra Trees maintained consistent superiority with an average test coefficient of determination of 0.769 and a root mean squared error of 0.274, narrowly outperforming Random Forest at 0.767 and 0.275, while LightGBM recorded 0.744 and 0.287. When evaluating out-of-sample projections for omitted nations, the researchers examined feature-space density to contextualize prediction reliability. Economies such as Morocco, Tunisia, and Ecuador occupied densely populated feature regions corresponding closely to the training distribution, supporting greater confidence in their model estimates. In contrast, countries such as Kenya and Mozambique fell along the outer boundaries of the feature space regarding foreign direct investment and trade volumes, meaning fewer neighboring training observations were available to anchor their projections.
The authors also detailed several technical boundaries within their current modeling framework. Specifically, the candidate indicators captured trade volume almost entirely through maritime statistics, omitting specific metrics on air cargo flows and failing to separate air shipments from ocean freight. Furthermore, the cross-sectional dataset was restricted to 2023 observations, preventing the analysis of temporal dynamics over time. Finally, the researchers noted that while their modular pipeline could theoretically be transferred to other composite benchmarks, such as the Global Competitiveness Index or the Environmental Performance Index, empirical validation remains necessary to verify whether open indicators can reliably estimate those targets.
Subject of Research: Machine learning prediction of the World Bank Logistics Performance Index
Article Title: Predicting the Logistics Performance Index (LPI) with machine learning methods
Article References: Ramirez, E. T., Rosas, P. J., Gamarra-Vargas, N., Quispe-Farfán, P. H., & Borda, R. F. C. (2026). Predicting the Logistics Performance Index (LPI) with machine learning methods. Journal of Shipping and Trade, 11(1), Article 32. https://doi.org/10.1186/s41072-026-00248-0
Image Credits: AI Generated
DOI: 10.1186/s41072-026-00248-0
Keywords: Logistics Performance Index, Machine Learning, Extra Trees, World Bank, Predictive Modeling, Global Supply Chains, Cross-Country Indicators, Trade Infrastructure, Economic Complexity Index, Multivariate Imputation, Random Forest, SHAP Values, LightGBM, Trade Logistics, Container Throughput, Public Datasets, Empirical Economics
Cite Scienmag News
Teresa Odom. (November 18, 2026). Machine Learning Models Predict Logistics Performance Index Scores Using Open Indicators. Scienmag. https://scienmag.com/machine-learning-models-predict-logistics-performance-index-scores-using-open-indicators/
Teresa Odom. "Machine Learning Models Predict Logistics Performance Index Scores Using Open Indicators." Scienmag, 18 November 2026, https://scienmag.com/machine-learning-models-predict-logistics-performance-index-scores-using-open-indicators/. Accessed 18 November 2026.
Teresa Odom. "Machine Learning Models Predict Logistics Performance Index Scores Using Open Indicators." Scienmag. November 18, 2026. https://scienmag.com/machine-learning-models-predict-logistics-performance-index-scores-using-open-indicators/








