Amazon 第三方賣家物流成本效益分析－以 Hour Loop 為例

Abstract

在 Amazon 平台運作下，第三方賣家需於訂單成立初期即選擇物流路徑，常見選項包括透過 Amazon Warehousing and Distribution（AWD）模式進行二段式配送，或採用 Fulfillment by Amazon（FBA）模式將商品直送至 Amazon 指定之倉儲中心。由於兩種路徑成本結構複雜且 Amazon 並未提供完整運費查詢機制，致使賣家難以即時做出具成本效益之決策。

本研究以 HourLoop 公司歷史訂單為研究對象，整合供應商、倉庫與商品尺寸等資料，建構小包裹（SPD）與卡車運輸（TL）之運費預測模型，並進一步建立分類模型以推薦每筆訂單最適物流路徑。模型結果顯示，Random Forest 在預測效能上表現最佳，SHAP 分析亦指出總體積與總重量為最具影響力之成本決策因子。敏感度測試結果顯示推薦模型具穩定性，僅在少數臨界訂單上產生路徑翻轉。

最終模擬結果顯示，若企業全面依推薦模型選擇路徑，整體運輸成本可較原始決策下降 16.57%。本研究提供一套數據驅動的物流決策工具，協助第三方賣家於訂單成立初期即做出有效且可解釋的路徑選擇，具備高度實務應用潛力。

Under the operational framework of Amazon, third-party sellers are required to determine their logistics path at the time of purchase order (PO) creation. The two main options include using Amazon Warehousing and Distribution (AWD) for two-stage delivery or Fulfillment by Amazon (FBA) for direct shipment to Amazon’s fulfillment centers (FCs). Due to the complexity of cost structures and the lack of transparent freight estimation tools, sellers face challenges in making cost-efficient decisions in real time.

This study focuses on historical order data from HourLoop Inc., integrating vendor addresses, warehouse locations, and product dimension information to develop freight cost prediction models for small parcel delivery (SPD) and truckload (TL) shipping. A classification model was further built to recommend the optimal logistics path for each PO. Experimental results show that the Random Forest model provides the best predictive performance, with SHAP analysis indicating that total volume and weight are the most influential cost factors. Sensitivity analysis reveals that the recommendation logic is stable under varying input conditions, with only marginal flips near decision boundaries.

Simulation results demonstrate that following the model’s recommendations can reduce total logistics costs by 16.57% compared to actual decisions. This research provides a data-driven decision support tool that enables third-party sellers to make cost-effective and explainable logistics choices at the point of order creation, offering strong potential for real-world application.