IDWSDS 2024 - S54: Statistics Helping the Growth of Business

Kaplan-Meier estimates and Markov models in health economic analysis: a statistical approach to business decisions
Hugo Pereira, University of Lisbon
Helena Mouriño, University of Lisbon
Raquel Fonseca, University of Lisbon
Liver transplantation (LT) is primary curative option for patients with hepatocellular carcinoma (HCC). Due to the scarcity of cadaveric donor livers, selection criteria have been established, but they are very restrictive. This study compares a new criterion, HepatoPredict (ClassI and ClassII), against existing ones (Milan Criteria (MC), UCSF, Up-to-7, AFP Model, and MetroTicket 2.0) using a cost-effectiveness analysis from the U.S. healthcare system perspective to determine which criteria is better. A Markov model was used to simulate the health status of patients with HCC who underwent LT over five years. Transition probabilities, costs, and utility were obtained from published data. Recurrence probabilities, calculated using Kaplan-Meier estimators, were based on a cohort of 149 patients from Portugal and Spain. We analysed the recurrence-free survival, life years gained, quality of life and the incremental cost-effectiveness ratio (ICER) relative to the MC. HepatoPredict offers the best benefit but has a higher cost. The ICER of HepatoPredict-ClassI and HepatoPredict-ClassII relative to the MC was $16 085.43/QALY and $39 407.58/QALY, respectively, both below the cost-effectiveness threshold (U.S. GDP per capita, $81 632.25/QALY), which means that HepatoPredict is acceptable in the U.S. healthcare system. It is the most cost-effective criterion and optimized organ allocation although deceased donor liver scarcity, with significant advantages for healthcare system.

Integrating Statistics and Machine Learning to Forecast New Products Sales: the Case of a Portuguese Brewery
Ricardo Galante, SAS Portugal and University of Lisbon
The introduction of new products is an important point of growth for any brewery, yet the inherent uncertainty of consumer preferences poses a significant challenge. Traditional forecasting methods may struggle to accurately predict demand in this dynamic market. This research explores the potential for machine learning (ML) systems to enhance demand forecasting for new products within the Portuguese brewery industry. We propose a framework utilizing historical sales data, market trends, and relevant external factors such as seasonality and economic indicators. A suite of ML algorithms, including cluster analysis, regression models, decision trees, and potentially neural networks, will be evaluated for their predictive performance. The study aims to: • Identify the most important predictors of demand for new brewery products. • Compare the accuracy of various ML algorithms in this forecasting context. • Develop a practical ML-based forecasting system tailored to the Portuguese brewery sector. This research provides breweries with data-driven insights into demand for new products, aiding in decision-making, production planning and, ultimately, improving resource allocation and profitability. Keywords: New Product Forecasting, Cluster Analysis, Gradient Boosting, Demand Forecasting, Machine Learning.

Credit Risk as a Tool for the Leverage of Investment in Business: Evaluating the Probability of Default for Medium Size Companies by Mariana Franco, Novobanco
Granting credit is one of the main banking activities and an essential factor for economic growth. However, a poor or careless risk assessment can have serious negative consequences, the most recent case being the so-called subprime crisis of 2008 Since then, the banking authorities have been tightening the regulations for granting credit, especially regarding the solvability ratio which has to be greater than 8%. The contribution of credit risk to this ratio is very important and is calculated on the basis of three parameters: the Probability of Default (PD), the Exposure at Default (EAD) and the Loss Given Default (LGD). We propose an evaluation method for the Probability of Default for medium-size companies,.on the basis of which the enterprise will be classified as compliant or non-compliant. As default is a binary variable, we use logistic regression with explanatory.variables.extracted.from a wide range of information about the situation of the enterprises, related to its balance sheet and other qualitative information.
After a careful treatment of the explanatory variables, including. the elimination of highly correlated variables and the grouping of different levels of some categorical variables, several logistic regression models were constructed, using different statistical approaches. All the models showed good prediction capabilities but we selected the approach that, while keeping a good predictive capability, produces the most parsimonious model.