Machine Learning-Based Procurement Price Benchmarking Through Historical Ranking and Market Evidence

Abstract: Historical procurement records can support price benchmarking, but selecting a suitable reference is challenging when similar descriptions represent different brands, models, specifications, commercial units, quantities, or procurement contexts. This paper examines the historical candidate-reranking component developed for procurement price decision support in CivicEngage. For each query, an upstream retrieval stage supplies ten historical candidates. The proposed reranker combines semantic similarity, structured-attribute matching, retrieval signals, and historical price- consensus features using a LightGBM binary classifier with a nested-cross-validation-based decision gate. It either retains the original Rank-1 candidate or replaces it with a candidate from the supplied set; the returned benchmark is the selected candidate's original unit price, not a newly generated or adjusted price. On the documented benchmark of 10,560 queries and 105,600 candidate rows, the gated reranker achieved 38.99% Accuracy@10, compared with 37.57% for the original Rank-1 baseline, an improvement of 1.42 percentage points. On a clean subset of 9,417 queries whose target descriptions did not contain explicit price strings, it achieved 39.70%, compared with 38.36% for the baseline. These results show a measurable but limited improvement for the evaluated candidate pool. They do not establish general current-market price accuracy, and the result remains constrained by whether a suitable historical candidate is retrieved.

Keywords: Procurement Price Benchmarking; Historical Procurement Data; Candidate Reranking; LightGBM; Binary Classification; Semantic Similarity; Price Consensus; Decision Support