Executive Summary

^{Author: Fabio Albuquerque — Advisor: Wagner Dos Anjos Carvalho}

^{Summary prepared by the ResumeAI tool, an artificial intelligence solution developed by Instituto Pecege focused on synthesis and writing.}

This work develops and compares three machine learning predictive models – multinomial logistic regression, decision trees, and random forests – to determine the approach with the highest accuracy in allocating a jewelry inventory. The objective is to propose a methodological framework to optimize inventory management, reduce product turnover time, and maximize revenue. The research seeks to answer which approach, between explanatory and purely predictive models, offers greater strategic value for the challenge of allocating unique items in the second-hand luxury retail sector, where the correct allocation of each product is a critical success factor. The study aims to create a replicable framework to automate and improve this decision, generating a direct impact on operational efficiency and profitability.

The luxury market in Brazil, the third largest among emerging nations (MCKINSEY, 2014), motivates this research. The sector grew 30% in 2023 (VALOR ECONÔMICO, 2025). Simultaneously, the trade of used products increased by 48.5% between 2020 and 2021 (SEBRAE, 2021), driven by the search for affordable prices and sustainability, which is expected to influence between 20% and 30% of jewelry purchases in the coming years (MCKINSEY, 2021).

In this context, inventory management is the main operational challenge for second-hand brands (THREDUP, 2024). Unlike traditional retail, the inventory is composed of unique and non-replicable items, acquired from individuals, making allocation a complex and high-impact decision. The Brazilian luxury consumer values the immediate availability of the product in physical stores (MCKINSEY, 2014), and inadequate allocation results in lost sales, increased turnover time, and tied-up capital. Excellence in allocation is, therefore, fundamental to accelerating the sales cycle.

The application of machine learning is a strategic solution for this complexity. Predictive models transform large volumes of historical data into optimized decisions (Provost and Fawcett, 2016). By analyzing jewelry characteristics (type, brand, stone, price) and the sales history of each store, algorithms predict the ideal destination for each new item, replacing intuition with a systematic, scalable process with greater potential for accuracy, aligning inventory with the consumption profiles of each point of sale.

The main contribution of this study is a replicable methodological framework for inventory optimization, materialized in an automation script. The framework values interpretability, because, as Provost and Fawcett (2016) argue, analyzing the internal structure of a model can generate more strategic value than its isolated predictions, by revealing the “why” behind the recommendations. By comparing models of different complexities, the work offers an adaptable solution that can be applied to other retail sectors facing similar challenges with unique item inventories, providing a practical guide for implementing data intelligence focused on results.

The methodology is quantitative, designed to evaluate the effectiveness of predictive models. The process begins with the construction of a database of simulated data that mirrors the reality of the Brazilian second-hand jewelry market. After a data wrangling stage, Multiple Correspondence Analysis (MCA) is used to explore consumption profiles. The core of the analysis is the development and comparison of three algorithms: Multinomial Logistic Regression (MLR), Decision Trees, and Random Forests. The performance of each model is evaluated to determine the optimal solution, balancing accuracy and interpretability.

The database was simulated via a Python script in Google Colab, allowing for a controlled and realistic scenario. The data construction was based on three pillars: the characteristics of the jewelry and prices were defined from research on the e-commerces of the main jewelry stores in Brazil; sales seasonality was modeled based on public quarterly reports from the country’s largest jewelry store, compiling data from 2023 and 2024 (Vivara S. A., 2023a, 2023b, 2023c, 2024, 2025) for a 24-month historical series; and distinct consumption profiles were created for four hypothetical stores, inducing statistical affinities by product category and price range. The analysis and modeling were conducted in Python with the Pandas, Statsmodels, and Scikit-learn libraries.

Before modeling, the data underwent “data wrangling”. The continuous price variable was converted into the categorical variable “fx_preco” with five ranges based on percentiles (10%, 25%, 50%, 75%), allowing classification models to identify consumption profiles more granularly. The date column was removed, as the focus was on the product profile. To process the categorical variables, the “one-hot encoding” technique was applied. To avoid perfect multicollinearity, the first “dummy” column of each original variable was removed, a recommended practice to ensure the stability of linear models (Géron, 2021). This data structure was used in all three models to ensure comparability of results.

The selection of models sought to balance predictive power and explanatory capacity, a central dilemma in applied data science (Shmueli, 2010). Multinomial Logistic Regression (MLR) was chosen as the base model due to its high interpretability, allowing for the testing of the statistical significance of each characteristic (Favero and Belfiore, 2024). The Decision Tree was employed to capture non-linear relationships through an easily visualized rule-based structure, using the Gini Index (Kotsiantis, 2007; Géron, 2021). Finally, the Random Forest model, an “ensemble” method, was selected to maximize predictive accuracy (Breiman, 2001). Performance was evaluated with confusion matrix metrics, such as Accuracy, Sensitivity, Specificity, and F1-Score, which is useful in scenarios with imbalanced classes (Favero and Belfiore, 2024; Géron, 2021).

The results phase began with Multiple Correspondence Analysis (MCA). A prior chi-square test confirmed (p-value < 0.05) the existence of statistically significant associations between the variables, validating the use of MCA. The technique decomposed the total inertia of the data (Favero and Belfiore, 2024), projecting the associations onto a perceptual map. The first two dimensions explained 35.87% of the total variance, allowing for the identification of clear profiles. The map revealed that Store D had a strong association with entry-level products (un-stoned jewelry, price range 0-1630), while Store A was the destination for high-value products (diamond necklaces, range 13441-25200). Stores B and C presented similar profiles, suggesting niche overlap.

The coordinates of each jewel in the two dimensions of MCA were incorporated into the dataset as new synthetic “features”. The first model, Multinomial Logistic Regression (MLR), was implemented with the original variables to establish a “baseline”. The global significance of the model was confirmed by the likelihood ratio test (low p-value). The analysis of the coefficients revealed profiles consistent with MCA: Store D showed a strong affinity for unstouned (coefficient of 4.26) and national brand (coefficient of 2.96) jewels compared to the reference store (Store A). It was inferred that Store A was the focus for imported brand products with diamonds, necklaces, and bracelets.

The RLM analysis was deepened by visualizing the impact of price on allocation probability. A graph of predicted probabilities showed that jewelry below R$5,000 had a higher probability of being allocated to Store D, while the probability for Store A grew exponentially with price. Stores C and B occupied intermediate price niches, with peaks between R$4,000-R$6,000 and R$9,000-R$12,000, respectively. The model achieved an accuracy of 62.9% in training and 61.6% in testing. The proximity between the values indicated good generalization, without overfitting, establishing a solid performance baseline.

Next, a Decision Tree model was developed. Initially, trained with ACM coordinates and with complexity cost pruning to avoid overfitting, the model presented an accuracy of 61.60% and an F1-score of 61.00% on the test set. However, a second tree model, adjusted only with the original variables, presented slightly superior performance, with an accuracy of 62.3% on the test set. This counterintuitive result suggested that ACM summarization might have hidden nuances that the model captured directly from product features. The greater interpretability of the second model, with clear business rules such as “if ring and national brand and no stone, then Store D”, made it more valuable in practice, reinforcing that complexity does not always translate into better performance (Shmueli, 2010).

The final stage of modeling sought the accuracy limit with a Random Forest model. With default hyperparameters, it achieved 62.6% accuracy on the test set. After optimization with “GridSearchCV”, the optimal combination found was “maxdepth” = 10, “nestimators” = 500, and “max_features” = 3. The final optimized model recorded an accuracy of 62.4% and a weighted F1-score of 62.3% on the test set. This result, although robust, did not represent a significant gain compared to simpler models.

The final comparative analysis of all models revealed the study’s most important finding. The Multinomial Logistic Regression model emerged as the solution with the highest predictive power, achieving a weighted F1-score of 62.9% and an accuracy of 63.1%. More complex models, such as Random Forests (F1-score of 62.3%), and feature engineering approaches, such as the Decision Tree with ACM coordinates (F1-score of 61.0%), did not outperform the RLM. The incorporation of ACM coordinates proved detrimental, suggesting that, for this dataset, the information contained in the original variables was richer and more directly usable.

The main conclusion is that, with similar performance between models, interpretability and simplicity should be the tie-breaking criteria. In this sense, both RLM, which offers clear statistical insights, and the Decision Tree (without ACM coordinates), which generates intuitive business rules, present themselves as the solutions with the greatest practical value. They outperform “black-box” models such as Random Forests, which, in this case, did not bring performance gains that justified the loss of transparency. The study demonstrates that the pursuit of complexity and marginal accuracy can be less valuable than the adoption of simpler models that provide actionable strategic insights.

The main finding of this study is that, for the jewelry stock allocation problem, methodological simplicity outperformed complexity. The Multinomial Logistic Regression model, the simplest and most interpretable, presented the best predictive performance, with an F1-score of 62.9%. More complex approaches, such as Random Forests and feature engineering with Principal Component Analysis, did not bring accuracy gains. This result validates a decision framework that, in the absence of significant predictive gains, should prioritize models offering greater interpretability and strategic value. The research reinforces that the intrinsic characteristics of the products are the most robust variables and that a model’s clarity can be a more valuable asset than marginal performance gains.

As a limitation, the study was based on simulated data which, although faithfully constructed, may not capture all the nuances of a real market. Future research could apply this framework to real sales data to validate the findings. It is also suggested to explore other variables not contemplated, such as the specific brand of the product and the weight of the jewelry, to potentially enrich the predictive power of the models. It is concluded that the objective was achieved: it was demonstrated that, for the problem of used jewelry stock allocation, a simpler and more interpretable multinomial logistic regression model outperformed more complex machine learning approaches in predictive performance.

References:
^{Breiman, L. 2001. Random forests. Machine Learning 45(1): 5-32.
Favero, L. P.; Belfiore P. Manual de Análise de Dados – Estatística e Machine Learning com Excel®, SPSS®, STATA®, R® e Python®. 2 ed. Gen Grupo Editorial Nacional, 2024.
Géron, A. Mãos à Obra: Aprendizado de Máquina com Scikit-Learn, Keras & TensorFlow: Conceitos, Ferramentas e Técnicas Para a Construção de Sistemas Inteligentes. 2. ed. Alta Books, 2021.
Kotsiantis, S. B. 2007. Supervised machine learning: a review of classification techniques. Informatica 31(3): 249-268
MCKINSEY. 2014. Capturing the hearts of Brazil’s luxury consumers. Available at: <https://www. mckinsey. com/capabilities/growth-marketing-and-sales/our-insights/capturing-the-hearts-of-brazils-luxury-consumers>. Accessed on: 12 Mar 2025.
MCKINSEY. 2021. In fine jewelry and watches, sustainability shines through. Available at: <https://www. mckinsey. com/industries/retail/our-insights/in-fine-jewelry-and-watches-sustainability-shines-through>. Accessed on: 12 Mar 2025.
Provost, F.; Fawcett, T. Data Science para Negócios: O que você precisa saber sobre mineração de dados e pensamento analítico de dados. 1. ed. Alta Books, 2016.
Serviço Brasileiro de Apoio às Micro e Pequenas Empresas [SEBRAE]. 2021. Mercado de segunda mão: um nicho bilionário da moda. Available at <https://sebrae. com. br/sites/PortalSebrae/artigos/mercado-de-segunda-mao-um-nicho-bilionario-da-moda,06c74a08ce761810VgnVCM100000d701210aRCRD>. Accessed on: 12 Mar 2025.
Shmueli, G. 2010. To explain or to predict?. Statistical Science 25(3): 289-310.
VALOR ECONÔMICO. 2025. Mercado de luxo cresce no Brasil, enquanto consumo global desacelera. Available at: <https://valor. globo. com/empresas/noticia/2025/02/20/mercado-de-luxo-cresce-no-brasil-enquanto-consumo-global-desacelera. ghtml>. Accessed on: 12 Mar 2025.
Vivara S. A. [VIVA3]. 2023a. Release de Resultados 1T23. Available at: <https://ri. vivara. com. br/central-de-resultados/>. Accessed on: 23 Feb. 2025.
Vivara S. A. [VIVA3]. 2023b. Release de Resultados 2T23. Available at: <https://ri. vivara. com. br/central-de-resultados/>. Accessed on: 23 Feb. 2025.
Vivara S. A. [VIVA3]. 2023c. Release de Resultados 3T23. Available at: <https://ri. vivara. com. br/central-de-resultados/>. Accessed on: 23 Feb. 2025.
Vivara S. A. [VIVA3]. 2024. Release de Resultados 4T23. Available at: <https://ri. vivara. com. br/central-de-resultados/>. Accessed on: 23 Feb. 2025.
Vivara S. A. [VIVA3]. 2025. Dados Históricos – Ano de 2024. Available at: <https://ri. vivara. com. br/central-de-resultados/>. Accessed on: 23 Feb. 2025.}

---

Executive summary from the Final Project of the Specialization in Data Science and Analytics from the MBA USP/Esalq

_{Learn more about the course; click here:}