基于集成学习的二手车交易价格预测模型

Used Car Transaction Price Prediction Model Based on Ensemble Learning

摘要: 二手车交易价格预测是汽车市场研究的重要问题，对交易平台定价策略与消费者决策具有实际意义。本文基于国内二手车交易平台收集的门店交易数据，构建了一个基于集成学习的二手车交易价格预测模型。首先对原始数据进行异常值剔除、缺失值填充及对数变换，以削弱共线性与异方差性；随后通过特征工程构造了车辆贬值速率、品牌价格统计特征等新变量；在此基础上，设计了包含ExtraTrees、RandomForest、CatBoost、LightGBM、KNeighbors、MLP、XGBoost等七种基学习器的三层stacking集成学习框架，并采用10折交叉验证降低过拟合风险。实验结果表明，该模型在本地验证集上的评估得分为0.645。本文进一步分析了特征重要性，揭示了影响二手车价格的核心因素。本文提出的模型可为二手车定价提供有效的技术参考，而关于销售周期分析与门店选址优化的讨论可作为未来研究方向。

Abstract: Second-hand car transaction price prediction is an important issue in automotive market research, offering practical significance for pricing strategies of trading platforms and consumer decision-making. Based on in-store transaction data collected from a domestic second-hand car trading platform, this paper constructs a price prediction model for second-hand car transactions using ensemble learning. The raw data are first preprocessed by outlier removal, missing value imputation, and logarithmic transformation to mitigate multicollinearity and heteroscedasticity. Subsequently, feature engineering is conducted to create new variables such as vehicle depreciation rate and brand-level price statistics. On this basis, a three-layer stacking ensemble learning framework is designed, incorporating seven base learners: ExtraTrees, RandomForest, CatBoost, LightGBM, KNeighbors, MLP, and XGBoost. A 10-fold cross-validation strategy is adopted to reduce the risk of overfitting. Experimental results show that the proposed model achieves an evaluation score of 0.645 on the local validation set. Furthermore, feature importance analysis is performed to identify the key factors affecting second-hand car prices. The proposed model can serve as an effective technical reference for second-hand car pricing, while discussions on sales cycle analysis and store location optimization are suggested as directions for future research.