E-commerce return prediction & return-rate reduction

Free returns are part of European online retail, but every return costs shipping, handling, re-processing and lost margin, and some returned goods are never resold. A return-prediction model scores each line item, basket or order for its return probability, often while the customer is still in the checkout. The score triggers measures that prevent a return before it happens: a size-advice nudge when two sizes of the same item are in the basket, a fix on a product page whose items "run small", or a different delivery or payment option. The same scores also tell the warehouse how many returns will arrive next week. The model itself is a standard gradient-boosting problem. Most of the effort goes into choosing interventions that cut returns without cutting sales.

Problem & value

• Scale: the Bamberg returns research group (Forschungsgruppe Retourenmanagement) estimates about 542 million return parcels in Germany in 2025 at a parcel return rate of 24.2%, and expects about 560 million in 2026. Fashion accounts for just over 80% of return parcels and about 90% of returned items [1].
• Waste: a 2026 report by the same group estimates that 1.3% (DACH) to 3.0% (rest of Europe) of returned items are disposed of. That comes to roughly 18 million items a year, and earlier surveys found about half of them still usable [2].
• Cost: return costs rise with the return rate, so preventing returns usually pays more than processing them faster [3].
• Today: returns are counted in BI reports after the fact. Category managers react to products with high return rates weeks later, and logistics plans return capacity from averages.
• AI value: (1) prevention: target nudges, size advice and checkout steering at high-risk baskets only, so low-risk customers are not affected; (2) product fixes: find articles, size charts and images that cause returns; (3) planning: item-level return probabilities give a better short-term forecast of returns volume than flat rates; (4) pricing and margin: expected return cost per customer or product feeds contribution-margin calculations and CLV (PRED-004).
• Measured effects of a single measure: Fit Analytics reports that retailers using its "Multiple Size Alert" (shown when a second size of the same item is added to the basket) saw 13% fewer multi-size orders and a 6% lower return rate. Its Fit Finder size advisor is quoted with an average return-rate reduction of 2–4% [4][5]. These are vendor figures without a published control-group design.

Variants

• Fashion & footwear (core case): size and fit drive most returns. Key signals are "bracketing" (several sizes or colours of one article in the basket), the article's fit history and the customer's own size history.
• Consumer electronics & home: fewer returns, but each one costs more (testing, refurbishing, data wiping). Signals are product-content gaps, compatibility questions and use-and-return patterns. Combine with VIS-037 condition grading.
• Marketplaces & platforms: score at seller and offer level (misleading listings, poor packaging) and use the scores in seller ranking and quality programmes.
• Pre- vs. post-checkout: pre-checkout scores drive nudges. Post-checkout scores drive logistics (pre-allocating return capacity, re-buy planning) and customer-service measures.
• Return abuse vs. ordinary returns: suspected abuse such as wardrobing or empty-box returns belongs to ANOM-033/ANOM-004. This use case is about ordinary, lawful returns.
• SMB / Shopify-class shops: product-level return analytics and size advice are usually enough; per-item scoring needs enough order history.

How it works

1. orders, line items, returns, reasons, product & customer data (DWH / CDP)
2. label building (returned within N days, per line item; partial returns handled)
3. feature store: basket (multi-size, same-SKU duplicates, discount), product (historic return rate, fit feedback), customer (return ratio, tenure), channel, payment, shipping
4. model: gradient boosting / tabular foundation model
5. calibrated P(return)
6. (online) real-time scoring at add-to-cart / checkout, < 100 ms
7. policy layer: uplift model + business rules pick the intervention (or none)
8. actions: size nudge, multi-size alert, content fix ticket, delivery option, review queue
9. aggregation: expected returns per day/warehouse
10. returns-volume forecast
11. monitoring: A/B holdout, conversion vs. return-rate trade-off, drift

• Prediction: binary classifiers on tabular features (LightGBM, XGBoost, CatBoost) remain the workhorse. Return reasons and reviews are turned into features with a text classifier or an LLM ("too small", "colour differs from picture").
• Decision: a high return probability does not mean an intervention will help. Uplift models (causal-inference) estimate which baskets change behaviour under a nudge, and conversion is watched as a guardrail metric.
• Human-in-the-loop: category managers review the product-level reports and decide on fixes. Any rule that restricts payment or delivery options needs sign-off from legal and the data protection officer, and customers need a manual route (see risks).

Data preparation

Needed data: order and line-item history with return flags, dates and reason codes, basket composition, product master data (size charts, material, images) and, where legal, customer history. It lives in the shop system, ERP or merchandise system, the returns portal, the carrier or warehouse system and in service e-mails.

Typical gaps: returns booked late or without reason codes, free-text reasons, product variants without consistent size and colour attributes, duplicate customer accounts, and returns booked at order level instead of line level. Personal data creates access and legal-basis limits.

Preparation steps: - Export orders and returns via shop API or ERP (PREP-005). - Match returns to the original line item and fix variant attributes (PREP-011). - Group free-text reasons into a clear category scheme (PREP-019) and harvest labels from existing records (PREP-020). - Remove duplicates and pseudonymise customer IDs (PREP-022); check purpose and legal basis (PREP-025). - Hold back the latest weeks as an evaluation set (PREP-018).

Tools: shop-system export, pandas, a spreadsheet for small shops. Share of total effort: about 30 percent. Master data quality is covered in DATA-007.

Implementation paths

Which path for whom

• smb (A): Shop-system plug-ins and return-management products offer size advice and return scoring at SMB prices, and the process is standard.
• enterprise (B): Return rates drive margin, and scoring must use own order, product and customer data, so cloud APIs or own models integrated in the shop fit.
• developer (B): A tabular model on top of the shop's order export is quickest on cloud services; self-hosting adds effort without clear gain.

Path A — Buy (recommended for most SMBs)

• Size & fit advice with return-prevention features: Fit Analytics Fit Finder with Multiple Size Alert (Berlin-founded), Easysize (fit quiz, return-reduction claims), plus fit tools built into shop platforms. These address the largest return driver directly (REC-012) [4][6].
• Returns platforms with analytics: parcelLab (Munich, DE; EU) with returns, exchanges and fraud/abuse signals [7]; Narvar (Shield, IRIS AI engine) [8]; EU shipping and returns portals such as Sendcloud (NL). They provide return-reason capture, product-level return analytics and exchange-instead-of-refund flows. Item-level prediction is often limited or only in premium tiers.
• Choose A when: you need quick wins on size-driven returns and structured return reasons and have no data-science team. Check EU hosting, a DPA, and whether raw scores can be exported.

Path B — Build on cloud APIs

• Managed AutoML on your order data: Google BigQuery ML / Vertex AI (EU multi-region), Azure Machine Learning AutoML (Germany West Central / EU regions), Amazon SageMaker Canvas or Autopilot (Frankfurt), Snowflake ML or Databricks on an EU region.
• Text features: an LLM API (Mistral, a French EU provider; Claude, GPT or Gemini in EU data zones) classifies free-text return reasons and reviews into a fixed reason taxonomy. Batch is enough here; no real-time LLM is needed.
• Online scoring: deploy the model behind a low-latency endpoint called from the checkout service. Feed the decisions into the shop's experimentation tool for holdout tests.
• Choose B when: you have a cloud data warehouse and want item-level scoring and your own intervention logic without operating ML infrastructure.

Path C — Self-host (open source)

• Models: LightGBM / XGBoost / CatBoost, or AutoGluon-Tabular for automated ensembles. TabPFN (Prior Labs GmbH, Freiburg, DE) is a strong tabular foundation model for smaller datasets. The current TabPFN-3.5 handles up to 1M rows, but the 2.5+ weights carry a non-commercial licence, so check licensing before production use [9].
• Uplift: CausalML or scikit-uplift. Text features: a fine-tuned multilingual encoder (e.g. a German BERT-class model) or a small open LLM (Qwen, Mistral Small) for reason classification.
• Stack: Feast feature store, MLflow, a FastAPI/BentoML scoring service, Kafka for the checkout event stream, Evidently for drift monitoring.
• Hardware: HW1. Training and scoring run on CPU servers. A GPU (HW2) is only needed when a local LLM classifies large volumes of return texts.

Path D — Custom training

Training your own model is the normal case once you go beyond Path A, because the labels (your own returns) and features are specific to the business. Requirements: at least several months of line-item order and return history, ideally a full seasonal cycle, plus logged nudges with a holdout group to train uplift models. Deep sequence models over customer histories or image-based fit features are only worthwhile for large retailers with millions of orders.

State of the art (2026)

• Models: gradient-boosted trees on carefully built basket and history features remain the standard. Bamberg research showed early on that relatively simple techniques can forecast returns well when the features are good [3]. Tabular foundation models (TabPFN, Nature 2025 [10]) make strong baselines quick to set up on smaller datasets.
• Shift from prediction to prevention: the focus has moved from accurate scores to causally evaluated interventions (multi-size alerts, fit advice, exchange-first flows). This is evaluated with holdout tests, because nudges can also lower conversion.
• GenAI contributions: LLMs turn unstructured return reasons, reviews and service chats into product-level defect signals ("sleeves too short in size M") and draft fix tickets for the product data. Virtual try-on (GEN-010) and AI size advice (REC-012) attack the root cause.
• Regulatory push: the ESPR destruction ban for unsold apparel and footwear (large companies from 19 July 2026) also covers products returned under the right of withdrawal. Preventing returns is therefore increasingly a compliance and sustainability topic as well [11].
• Limits: labels arrive late, because returns come in days to weeks after the order. Partial returns and exchanges blur the labels. Nudges change the behaviour the model learned from, which is a feedback loop. Scores for new products and new customers are weak (cold start). A model that learns "customer returns a lot" invites restrictive treatment, which carries legal and reputational risk.

Shared resources & combinations

• Same tabular stack (tabular-ml + causal-inference + feature store): PRED-003 churn, PRED-004 CLV (expected return cost lowers CLV), PRED-005 lead scoring and PRED-021 payment behaviour can share one feature store, model registry and CPU scoring service (HW1).
• Prevention partners: REC-012 size & fit recommendation (the main intervention), GEN-010 virtual try-on, LANG-001 product descriptions (fixing content gaps), REC-001 recommendations (avoid recommending high-return items).
• Downstream: PRED-001 demand forecasting (net demand after returns), OPT-005 inventory replenishment (returned stock re-enters availability), OPT-006 warehouse planning and PRED-011 staffing for the returns centre, VIS-037 condition grading of the returned items.
• Abuse boundary: ANOM-033 retail returns fraud, ANOM-004 platform and promo abuse and ANOM-001 payment fraud share features (customer return history), but they serve a separate decision and need separate governance.
• Text analytics: LANG-013 review and feedback analysis shares the return-reason classifier.

Effort, cost & KPIs

• Path A: days to weeks to add size advice and a returns portal. Pricing is typically per order, per return or by revenue tier.
• Path B/C: 2–4 months for a first model plus product-level reporting. Real-time checkout integration and a clean A/B setup take another 1–3 months. Compute costs are low; the cost drivers are data engineering and experimentation capacity.
• KPIs: return rate (items and value) vs. holdout, conversion rate and revenue per visitor as guardrails, net margin after returns, share of multi-size baskets, model AUC and calibration, forecast error of weekly return volumes, and returned items disposed of.

People & change

Involve management for goals (margin versus conversion), then the e-commerce or customer-service lead who owns the process. IT connects shop and ERP; the DPO is essential because the score profiles customers. Customer service staff (end users) should see why an order is flagged.

Works council: usually no, since the score concerns customers, not employees. It applies if the data also evaluates service-agent or warehouse performance; then §87(1) no. 6 BetrVG matters, and §90(1) no. 3 requires information in good time (GUIDE-004).

Data protection: scoring individual customers is profiling. Check the legal basis, run a DPIA check under Art. 35 GDPR, inform customers in the privacy notice, and conclude AVVs with providers (GUIDE-005). Never block customers solely on an automated score (Art. 22 GDPR); prefer product-level fixes and neutral nudges.

AI literacy: support the team with a short briefing on what a probability score means (GUIDE-006). Pilot on one product category with an A/B test and measure return rate, conversion and complaints. Keep humans in control: staff can override interventions, and sensitive measures such as shipping restrictions need approval. Typical concern: "customers will be treated unfairly". Address it with transparent rules and regular fairness checks (GUIDE-008, GUIDE-009).

Risks, regulation & compliance

• EU AI Act: minimal risk in the typical form, as product and basket risk scoring for operational decisions is not an Annex III use. The boundary is creditworthiness: if the score feeds the decision on whether a natural person may pay on invoice or by retailer-financed BNPL, check Annex III point 5(b) (creditworthiness / credit scoring of natural persons, high-risk) and keep the two models separate. AI literacy (Art. 4) applies [12].
• GDPR: customer-level return histories are personal data. You need a legal basis (usually legitimate interest with a documented balancing test), transparency in the privacy notice and data minimisation. Excluding a customer from payment methods or delivery options solely on the score can be a decision with similarly significant effects under Art. 22. The CJEU's SCHUFA ruling (C-634/21) reads Art. 22 broadly. Keep a human review route and explain the logic on request [13].
• Consumer law: the 14-day right of withdrawal (Directive 2011/83/EU) cannot be restricted by a score [14]. Nudges must not become dark patterns (UCPD; the planned Digital Fairness Act). Personalised prices based on automated decision-making must be disclosed (Art. 6(1)(ea) CRD).
• Fairness and reputation: proxies such as postcode, age or gender can produce discriminatory treatment (German AGG for access to goods and services). Exclude or audit them.
• Circular economy: destruction of returns conflicts with the German KrWG duty of care (Obhutspflicht) and, for apparel and footwear, the ESPR ban, so measure disposal as a KPI [2][11].

Getting started

1. Build the baseline: return rate by category, product, size, channel and payment method, plus structured return reasons (fix the reason codes first if they are free text or missing).
2. Quick wins: activate a size advisor with a multiple-size alert, and send the top 50 high-return products to the category managers with their reason analysis.
3. Train a line-item return model on 6–12 months of history, check its calibration, and use it first for product reports and volume forecasting, where no customer is affected.
4. Run one intervention (e.g. size nudge for high-risk multi-size baskets) as an A/B test, with conversion as a guardrail.
5. Before any payment or delivery steering, document the legal basis, the Art. 22 assessment and the manual review route with the DPO.

Sources

1. Forschungsgruppe Retourenmanagement, University of Bamberg — Updated estimate of return parcels 2026 — https://www.retourenforschung.de/info-aktualisierte-schaetzung-bezueglich-der-anzahl-an-retourenpaketen-im-jahr-2026.html
2. Forschungsgruppe Retourenmanagement — Report on disposal of returns and overstock (2026) — https://www.retourenforschung.de/info-neuer-forschungsbericht-veroeffentlicht-warum-werden-retouren-und-%C3%9Cberbestaende-entsorgt-was-kann-man-dagegen-tun.html
3. Forschungsgruppe Retourenmanagement — Publications (incl. Asdecker & Karl 2018, "Big data analytics in returns management – are complex techniques necessary to forecast consumer returns properly?"; Asdecker 2015 on return rate vs. cost) — https://www.retourenforschung.de/publikationen.html
4. Fit Analytics — Multiple Size Alert reduces returns by 6% — https://www.fitanalytics.com/blog/multiple-size-alert-reduces-returns
5. Fit Analytics — Fit Finder product page — https://www.fitanalytics.com/
6. Easysize — size & fit quiz — https://www.easysize.me/
7. parcelLab — returns, exchanges and fraud & abuse protection — https://parcellab.com/ and https://docs.parcellab.com/docs/retain/fraud-and-abuse
8. Narvar — returns & exchanges (Shield) and IRIS — https://corp.narvar.com/
9. Prior Labs — TabPFN repository (versions, limits, licences) — https://github.com/PriorLabs/TabPFN
10. Hollmann et al. (2025), "Accurate predictions on small data with a tabular foundation model", Nature — https://www.nature.com/articles/s41586-024-08328-6
11. Regulation (EU) 2024/1781 (Ecodesign for Sustainable Products Regulation, destruction of unsold consumer products) — https://eur-lex.europa.eu/eli/reg/2024/1781/oj
12. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
13. Regulation (EU) 2016/679 (GDPR) — https://eur-lex.europa.eu/eli/reg/2016/679/oj
14. Directive 2011/83/EU (Consumer Rights Directive) — https://eur-lex.europa.eu/eli/dir/2011/83/oj