Evaluating Predictive Models for Reducing E-Commerce Returns and Associated Carbon Emissions

Ritesh Kalidindi, Leelavathy Narkedamilly, Uma Meghana Saladi

Abstract: Product returns generate a disproportionate share of e-commerce?s environmental footprint through repeated reverse-haul transportation, repackaging, and, for a meaningful share of items, outright disposal, which makes predictive models for identifying high-return-risk orders a potentially valuable lever for cutting both return volume and the carbon emissions that follow from it. This paper evaluates seven predictive models - logistic regression, decision tree, random forest, gradient boosting, SVM, k-nearest neighbours, and a neural network - for checkout-stage return-risk prediction, extending an earlier real- data evaluation on the BADS German fashion-retailer dataset (best model ROC-AUC = 0.574) with an independent re-evaluation of the full pipeline - feature engineering, model benchmarking, a feature-group ablation study, feature-importance analysis, and a literature-grounded carbon-emissions estimator - on a second, structurally different Kaggle dataset: a 10,000-row synthetic e-commerce returns dataset in which each customer places exactly one order. On this dataset, every predictive model evaluated produced a held-out ROC-AUC between 0.482 and 0.498, statistically indistinguishable from chance (0.50), and the ablation study confirmed that no feature group discriminated meaningfully above chance either. Rather than a weakness, this outcome functions as a negative control on the evaluation pipeline itself: it shows that the same models and procedure do not manufacture spurious predictive skill when no genuine return- risk signal is present in the underlying fields, which strengthens confidence that the earlier positive evaluation on real BADS order data reflected a genuine, learnable relationship rather than a pipeline artefact. A full carbon-emissions reduction calculator is maintained, and reported to ensure methodological completeness, and translates model recall into projected CO2-equivalent and packaging- waste savings over a set of intervention-effectiveness and emissions conditions, but the output is explicitly assigned the caveat of being an illustrative, but not a validated deployable projection of this dataset, until it is combined with a demonstrably-beating model. This paper has a fully reproducible Google Colab notebook that downloads the dataset and recreates all figures and tables reported end-to-end.

Keywords: Product Return Prediction, Machine Learning, E-Commerce Sustainability, Ablation Study, Negative Control, Reproducibility

How to Cite?: Ritesh Kalidindi, Leelavathy Narkedamilly, Uma Meghana Saladi, "Evaluating Predictive Models for Reducing E-Commerce Returns and Associated Carbon Emissions", Volume 15 Issue 7, July 2026, International Journal of Science and Research (IJSR), Pages: 1472-1481, https://www.ijsr.net/getabstract.php?paperid=SR26717182049, DOI: https://dx.doi.org/10.21275/SR26717182049

Download Citation: APA | MLA | BibTeX | EndNote | RefMan