Search

In order to assess risks associated with establishing relationships with corporate partners such as clients, suppliers, debtors or contractors, decision makers often turn to business failure prediction models. While a large body of literature has focused on optimizing and evaluating novel methods in terms of classification accuracy, recent research has acknowledged the existence of asymmetric misclassification costs associated with prediction errors and thus, advocates the usage of alternative evaluation metrics. However, these papers often assume a misclassification cost matrix to be known and fixed for both the training and the evaluation of models, whereas in reality these costs are often uncertain. This paper presents a methodological framework based upon heterogeneous ensemble selection and multi-objective optimization for cost-sensitive business failure prediction that accommodates uncertainty at the level of misclassification costs. The framework assumes unknown costs during model training and accommodates varying degrees of uncertainty during model deployment. Specifically, NSGA-II is deployed to optimize cost space resulting in a set of pareto-optimal ensemble classifiers where every learner minimizes expected misclassification cost for a specific range of cost ratios. An extensive set of experiments evaluates the method on multiple data sets and for different scenarios that reflect the extent to which cost ratios are known during model deployment. Results clearly demonstrate the ability of our method to minimize cost under the absence of exact knowledge of misclassification costs.

The goal of sequential testing is to discover the state of a system by testing its components one by one. We consider n-out-of-n systems, which function only if all n components work. The testing continues until the system state (up or down) is identified. The tests have known execution costs and failure probabilities, and are subject to precedence constraints. The objective is to find a sequence of tests that minimizes the total expected cost of the diagnosis. We show how to strengthen the precedence graph without losing all optimal solutions. We examine different formulations for the problem, and propose a dynamic-programming (DP) and a branch-and-price algorithm. Our computational results show that our DP noticeably outperforms the state of the art. Using a novel memory management technique, it significantly increases the size of the instances that can be solved to optimality within given limits on runtime and memory. Keywords: sequential testing, n-out-of-n systems, precedence theorems, dynamic programming, branch-and-price. ispartof: EUROPEAN JOURNAL OF OPERATIONAL RESEARCH vol:274 issue:3 pages:876-885 status: published

International audience; This real-world study with a large European financial services provider combines aggregated customer data including customer demographics, behavior and contact with the firm, with fine-grained transaction data to predict four different customer life events: moving, birth of a child, new relationship, and end of a relationship. The fine-grained transaction data—approximately 60 million debit transactions involving around 132,000 customers to >1.5 million different counterparties over a one-year period—reveal a pseudo-social network that supports the derivation of behavioral similarity measures. To advance decision support systems literature, this study validates the proposed customer life event prediction model in a real-world setting in the financial services industry; compares models that rely on aggregated data, fine-grained transaction data, and their combination; and extends existing methods to incorporate fine-grained data that preserve recency, frequency, and monetary value information of the transactions. The results show that the proposed model predicts life events significantly better than random guessing, especially with the combination of fine-grained transaction and aggregated data. Incorporating recency, frequency, and monetary value information of fine-grained transaction data also significantly improves performance compared with models based on binary logs. Fine-grained transaction data accounts for the largest part of the total variable importance, for all but one of the life events.

This study proposes a decision support framework to help e-commerce companies select the best collaborative filtering algorithms (CF) for generating recommendations on the basis of online binary purchase data. To create this framework, an experimental design applies several CF configurations, which are characterized by different data-reduction techniques, CF methods, and similarity measures, to binary purchase data sets with distinct input data characteristics, i.e., sparsity level, purchase distribution, and item–user ratio. The evaluations in terms of accuracy, diversity, computation time, and trade-offs among these metrics reveal that the best-performing algorithm in terms of accuracy remains consistent regardless of the input-data characteristics. However, for diversity and computation time, the best-performing model varies with the input characteristics. This framework allows e-commerce companies to decide on the optimal CF configuration as a function of their specific binary purchase data sets. They also gain insight into the impact of changes in the input data set on the preferred algorithm configuration.

Financial institutions use credit scorecards for risk management. A scorecard is a data-driven model for predicting default probabilities. Scorecard assessment concentrates on how well a scorecard discriminates good and bad risk. Whether predicted and observed default probabilities agree (i.e., calibration) is an equally important yet often overlooked dimension of scorecard performance. Surprisingly, no attempt has been made to systematically explore different calibration methods and their implications in credit scoring. The goal of the paper is to integrate previous work on probability calibration, to re-introduce available calibration techniques to the credit scoring community, and to empirically examine the extent to which they improve scorecards. More specifically, using real-world credit scoring data, we first develop scorecards using different classifiers, next apply calibration methods to the classifier predictions, and then measure the degree to which they improve calibration. To evaluate performance, we measure the accuracy of predictions in terms of the Brier Score before and after calibration, and employ repeated measures analysis of variance to test for significant differences between group means. Furthermore, we check calibration using reliability plots and decompose the Brier Score to clarify the origin of performance differences across calibrators. The observed results suggest that post-processing scorecard predictions using a calibrator is beneficial. Calibrators improve scorecard calibration while the discriminatory ability remains unaffected. Generalized additive models are particularly suitable for calibrating classifier predictions.

Cette these ameliore la notation des clients. L’evaluation des clients est importante pour les entreprises dans leurs processus de prise de decision parce qu'elle aide a resoudre des problemes de gestion cles tels que le choix des clients a cibler pour une campagne de marketing ou l'analyse des clients qui sont susceptibles de quitter l'entreprise. La recherche effectuee dans le cadre de cette these apporte plusieurs contributions dans trois domaines de la litterature sur la notation des clients. Premierement, de nouvelles sources de donnees sont utilisees pour evaluer les clients. Deuxiemement, la methodologie pour passer des donnees aux decisions est amelioree. Troisiemement, la prediction des evenements courants du client est proposee comme une nouvelle application de la notation des clients. Tous les resultats presentes dans cette these sont issus de donnees reelles et sont non seulement d'une grande valeur academique, mais aussi d'une grande pertinence commerciale.

International audience; This study investigates the value added by incorporating textual data into customer churn prediction (CCP) models. It extends the previous literature by benchmarking convolutional neural networks (CNNs) against current best practices for analyzing textual data in CCP, and, using real life data from a European financial services provider, validates a framework that explains how textual data can be incorporated in a predictive model. First, the results confirm previous research showing that the inclusion of textual data in a CCP model improves its predictive performance. Second, CNNs outperform current best practices for text mining in CCP. Third, textual data are an important source of data for CCP, but unstructured textual data alone cannot create churn prediction models that are competitive with models that use traditional structured data. A calculation of the additional profit obtained from a customer retention campaign through the inclusion of textual information can be used by practitioners directly to help them make more informed decisions on whether to invest in text mining.

International audience; In this paper we propose a heuristic approach that computes the order in which patients will be treated in an ambulatory chemotherapy center. Each patient follows an individual treatment plan that fixes dates for series of drug injections separated by recovery periods. The daily care process has three steps: consultation with the oncologist, drug preparation in the pharmacy and drug injection in medical beds. The facility closes after the last injection. As drug injection varying considerably in duration-from 15 minutes to 6 hours-bad schedules lead to excessive overtime. In addition, after the consultation the oncologist may decide to cancel the injection because of a weak patient's health condition. In the current setting of the chemotherapy facility we work with, First Come First Served policy controls the care process. In this study, we propose to compute a common priority list of patients for consultation and injection phases. A unique list of patients is a simple tool used by nurses to manage the flow of patients and to react to uncertain events. A GRASP algorithm is developed to compute optimized list of patients in few seconds as the operating planning context requires. Two objectives are considered; the closing time and the overworking time of the facility. Numerical experiments show that our GRASP is able to quickly reach near optimal solutions and that list of patients policy performance is comparable to more complex scheduling policies. Benchmark data sets are built based on historical data of the French chemotherapy facility, ICL in Saint-´ Etienne.

International audience; The popularity of the big data domain has boosted corporate interest in collecting and storing tremendous amounts of consumers’ textual information. However, decision makers are often overwhelmed by the abundance of information, and the usage of text mining (TM) tools is still at its infancy. This study validates an extended technology acceptance model integrating information quality (IQ) and top management support. Results confirm that IQ influences behavioral intentions and TM tools usage, through perceptions of external control, perceived ease of use, and perceived usefulness; top management support also has a key role in determining the usage of TM tools.

Nous situons ce chapitre dans le cadre de l’utilisation de methodes qualitatives et quantitatives, notamment celles ayant pour but de faire emerger des concepts et des theories. Plus precisement, nous nous appuyons sur une synthese de la litterature sur la creativite et les neurosciences pour proposer des outils de developpement des competences en conceptualisation. Nous retenons notamment l’importance des differents stades de la creativite, l’adaptation du mode de pensee par rapport aux specificites de la tâche, les trois vues de l’attention (diffuse, concentree et flexible) et quand les utiliser, le role de l’attention interne et externe dans la maintenance de representations mentales complexes sur une longue periode de temps, le role de la meditation et finalement, cinq habitudes des personnes tres creatives. Enfin, nous ouvrons la discussion sur le futur des approches methodologiques au vu des recentes avancees technologiques comme l’intelligence artificielle ou la representation holographique de pensees.

Decision trees and logistic regression are two very popular algorithms in customer churn prediction with strong predictive performance and good comprehensibility. Despite these strengths, decision trees tend to have problems to handle linear relations between variables and logistic regression has difficulties with interaction effects between variables. Therefore a new hybrid algorithm, the logit leaf model (LLM), is proposed to better classify data. The idea behind the LLM is that different models constructed on segments of the data rather than on the entire dataset lead to better predictive performance while maintaining the comprehensibility from the models constructed in the leaves. The LLM consists of two stages: a segmentation phase and a prediction phase. In the first stage customer segments are identified using decision rules and in the second stage a model is created for every leaf of this tree. This new hybrid approach is benchmarked against decision trees, logistic regression, random forests and logistic model trees with regards to the predictive performance and comprehensibility. The area under the receiver operating characteristics curve (AUC) and top decile lift (TDL) are used to measure the predictive performance for which LLM scores significantly better than its building blocks logistic regression and decision trees and performs at least as well as more advanced ensemble methods random forests and logistic model trees. Comprehensibility is addressed by a case study for which we observe some key benefits using the LLM compared to using decision trees or logistic regression.

International audience

Investigating linkages between credit and equity markets, we consider daily aggregate U.S. CDS spreads as well as well-chosen equity market and implied volatility indexes over ten years. We describe such robust (to spurious correlation) relationship with the quantile cointegrating regression approach. Such approach handles extreme quantiles/CDS values and their behavior with respect to the equity market's influence. Heteroskedastic patterns such as time-varying variance, but also autocorrelation, skewness and leptokurtosis are captured. Thus, the sensitivity of aggregate CDS spreads to equity market price and volatility channels is accurately measured across quantiles and spreads. Such quantile-dependent sensitivity exhibits asymmetric responses to equity market shocks. A sub-period analysis investigates potential regime shifts in estimated quantile cointegrating regressions. Quantile cointegrating coefficients vary over time and quantiles, and exhibit different magnitudes across sub-periods and spreads. Therefore, the relationship is unstable over time. We also propose a scenario analysis and risk signaling application for credit risk management prospects. Under specific risk levels, credit risky situations are described conditional on the equity market's information over time, and related expected aggregate CDS spreads are computed. Estimated conditional quantiles/CDS spreads act as credit alert triggers.

International audience; In the recent financial crisis period, when traditional economic organizations were not able to meet stakeholders’ expectations, not-for profit organizations such as cooperatives became an optimal solution as they are expected to serve social and economic performance simultaneously. This theoretical assumption is questioned based on the downturn pressures that may weaken cooperatives’ social performance in favor of economic performance. This degeneration process (Cornforth, et al., 1998) is countered by some traditional cooperatives that have developed regeneration dynamics. The aim of our study is to explore how small and medium cooperatives face degeneration and develop regeneration dynamics in periods of crisis. To fulfil the objective of the paper four small and medium Basque and Breton cooperatives are studied in depth.

Data preparation is a process that aims to convert independent (categorical and continuous) variables into a form appropriate for further analysis. We examine data-preparation alternatives to enhance the prediction performance for the commonly-used logit model. This study, conducted in a churn prediction modeling context, benchmarks an optimized logit model against eight state-of-the-art data mining techniques that use standard input data, including real-world cross-sectional data from a large European telecommunication provider. The results lead to following conclusions. (i) Analysts better acknowledge that the data-preparation technique they choose actually affects churn prediction performance; we find improvements of up to 14.5% in the area under the receiving operating characteristics curve and 34% in the top decile lift. (ii) The enhanced logistic regression also is competitive with more advanced single and ensemble data mining algorithms. This article concludes with some managerial implications and suggestions for further research, including evidence of the generalizability of the results for other business settings. We study the impact of data preparation on customer churn prediction performance.Effective data preparation improves AUC up to 14.5% and top decile lift up to 34%.Optimized logistic regression is competitive with advanced data mining algorithms.

This article focuses on changes in leadership’s discourse about the “triple bottom line” in Ben & Jerry’s ice cream from its founding days through to its acquisition by and integration into Unilever. For this study, we analyzed CEO claims about “who we are” from their letters in annual reports (what we label projected identity). A sample of employees (both long-service and relative newcomers) were interviewed about their perceptions of Ben & Jerry’s over the 30 years covered. Findings reveal that successive CEOs stressed different “logics” about the business and what would make it successful over the years with the founders emphasizing a strong linkage between the economic, product, and social components of the company’s triple bottom line and their next three successors decoupling these components and pushing, each in different ways, for stronger financial returns. As a result, organization members were “whipsawed” between their CEOs’ different logics and identity claims. The CEO letters exhibit a progression over time from a more normative to utilitarian tone familiar in the organizational identity literature. The messaging shifts, however, when a fifth CEO takes charge and reintegrates the firm’s triple bottom line. Thus, the firm’s projected identity evolved in a U pattern starting with an integrated triple bottom line logic, shifting to a more linear logic where the economic mission dominates, and then reintegration where multiple bottom lines are embraced once again. Here, we explore both the strategic (external) and personal (internal) challenges informing the different CEOs’ messages over years, the whipsaw effect on staff, and the longer term evolution of projected identity in the company and reemergence of its integrated triple bottom line. This study contributes to the corporate social responsibility and organization identity literatures by documenting how CEOs (and their company) must struggle with maintaining an integrated triple bottom line in the context of commercial challenges and major changes involved in mergers and acquisitions. It also speaks to the practical matters of keeping normative traditions alive amid competing pressures for change.

This study creates a hybrid recommendation system for online offer personalization of an e-commerce company. The system goes beyond existing literature by combining four different data sources, i.e. customer data, product data, implicit and explicit behavioral data, in a single algorithm. Factorization machines are employed as model-based algorithm and have as advantage that the four data sources are incorporated in a single model by feature combination. Results show that hybridization of the four distinct data sources improves accuracy compared to (i) factorization machines based on a single data source and (ii) a real-life company benchmark model using collaborative filtering.

International audience