Selçuk Bayracı

Used car auction platforms face unique challenges in personalized recommendation due to extreme data sparsity, high inventory turnover, and real-time operational constraints. This study develops and evaluates a hybrid recommendation system combining Word2Vec embeddings for categorical vehicle attributes with standardized numerical features, applying temporal decay weighting to prioritize recent user interactions. Deployed on Azure infrastructure, the system was evaluated using 12 months of transaction data from a Turkish B2B auction platform comprising 5,322 users, 24,987 vehicles, and 1.87 million interactions. Offline evaluation demonstrates superior performance over baselines (Hit Rate@10: 0.456 vs 0.234 popularity baseline, 94.9% improvement). Production deployment over six months (April–September 2025) generated 977 recommendation-driven sales representing 15.26% of total platform transactions and 17.27M TL in commission revenue. Quasi-experimental analysis revealed a 26.7% increase in monthly purchase frequency among active users, yielding 420 incremental transactions. Results demonstrate how interpretable temporal-weighted embedding models generate measurable commercial value in high-turnover, data-sparse B2B marketplaces.

Accurately predicting sales conversion in automotive CRM systems is critical for optimizing marketing spend and sales team efficiency. This study presents a calibrated ensemble framework combining XGBoost, Gradient Boosting, and Random Forest classifiers to predict lead conversion probability in automotive dealership operations. Using 62,859 real-world leads collected between July 2024 and July 2025, we developed a systematic pipeline encompassing behavioral feature engineering, statistical feature selection, ensemble modeling, and probability calibration via Platt scaling. The calibrated ensemble achieved an AUC of 0.841, Brier score of 0.146, and 19% improvement in top-decile precision over baseline logistic regression. The framework provides actionable lead segmentation into four priority tiers, directly supporting sales resource allocation and marketing campaign optimization. Results confirm that probability calibration is essential for automotive CRM applications where predicted scores inform operational decisions.

Customer retention and vehicle renewal prediction remain critical challenges in premium automotive markets. This study presents a comprehensive data-driven framework for predicting BMW customer renewal probability using historical transactional and behavioral data from Borusan Otomotiv's enterprise systems. We developed a hybrid machine learning model that integrates Random Forest feature selection with Binary Logistic Regression to achieve interpretability while maintaining predictive accuracy. The model leverages customer demographics, service engagement metrics, and ownership patterns to generate individual-level renewal probability scores. Evaluated on 1,211 holdout observations through temporal validation, the model achieved 77% overall accuracy and an AUC-ROC of 0.80, demonstrating strong discriminatory power in distinguishing between renewal and non-renewal customers. Model outputs are transformed into five operational risk grades (G1-G5) and seamlessly integrated into Salesforce CRM, enabling proactive customer relationship management and targeted retention strategies. Key empirical findings indicate that service expenditure patterns, time since last purchase, and multi-vehicle ownership significantly influence renewal likelihood. The framework bridges predictive analytics with operational deployment through automated data pipelines and continuous model monitoring, representing a practical approach to data-driven customer retention in the automotive sector.