Applied Stochastic Models in Business and Industry最新文献

Fairness is a key requirement for artificial intelligence applications. The assessment of fairness is typically based on group-based measures, such as statistical parity, which compares the machine learning output for the different population groups of a protected variable. Although intuitive and simple, statistical parity may be affected by the presence of control variables, correlated with the protected variable. To remove this effect, we propose to employ Shapley values, which measure the additional difference in output specifically due to the protected variable. To remove the possible impact of correlations on Shapley values, we compare them across different subgroups of the most correlated control variables, checking for the presence of Simpson's paradox, for which a fair model may become unfair when conditioning on a control variable. We also show how to mitigate unfairness by means of a propensity score matching that can improve statistical parity, building a training sample that matches similar individuals in different protected groups. We apply our proposal to a real-world database containing 157,269 personal lending decisions and show that both logistic regression and random forest models are fair when all loan applications are considered, but become unfair for high loan amounts requested. We show how propensity score matching can mitigate this bias.

The supermarket industry faces fierce competition in many geographic areas, with various brands competing for a market share that represents a significant portion of a country's GDP. For these businesses, understanding the factors that influence their sales and how they do so is crucial: location, pricing, quality, customer loyalty, and more. While much has been written about how these elements can affect revenue, there is a lack of existing models quantifying the monetary impact of having competitor stores within an own establishment area of influence. This article introduces a hierarchical Bayesian model that quantifies this impact for each product category within the brand's catalog. The model requires careful parameterization to adapt to the specific context, and therefore both its application and design represent a methodological novelty in this area. The main result is a matrix of coefficients that indicate the monetary effect of the presence of every competitor for each product category. The findings revealed by the model show that the influence of the proximity of each competitor does indeed vary depending on these categories. This information is valuable for the company when making decisions, such as choosing new store locations or determining the product assortment to offer in each establishment. The use of the results derived from this model provides a competitive advantage over others and helps to better understand the market by identifying which supermarket brands are the perceived leaders in each product category.

Bootstrap procedures represent a straightforward approach to assessing the uncertainty around estimates of interest in statistical models. However, with the rising prevalence of massive datasets in statistical problems, the computational cost of bootstrap methods can quickly become prohibitive in many settings. To this end, this paper proposes the Averaged Robbins-Monro Bootstrap (ARM-B), a scalable tool for estimating parameter variability via multiple chains of Robbins-Monro updates. The method is illustrated in large-scale Poisson regression and logistic regression settings and compared with the alternative scalable method given by the bag of little bootstraps (BLB). Some simulation experiments and an illustrative analysis on a large-scale dataset show that ARM-B has comparable accuracy with ordinary bootstrap, but, at the same time, it is significantly less computationally demanding and quite competitive with BLB.

Discrete Choice Experiments (DCEs) investigate the attributes that affect individual choices among different options and are widely applied across numerous fields. Past DCEs provide clear evidence that the presentation order of the profiles within a choice set can impact the respondents' choices. Ignoring such order effects can produce severely biased estimates, as we illustrate using a product packaging DCE performed for Procter & Gamble in Mexico. Currently, the most common approach to address profile order effects is to randomize the profile order. While this method is relatively easy to implement in online surveys, it can be considerably cumbersome in offline experimental settings. To address this, we suggest incorporating an order covariate in the model to measure the effect of profile order, and propose a Bayesian optimal Balanced Profile Order Design (BPOD) that accounts for this order effect. Our simulation experiments reveal that our Bayesian optimal BPOD achieves accurate parameter estimates comparable to those obtained through randomization in both the multinomial logit model and the panel mixed logit model. Beyond DCEs, this design strategy contributes to broader efforts in experimental design by providing a generalizable framework for addressing structural sources of bias in applied statistical research.

Machine learning (ML) will most likely play a large role in many processes in the future, also in the insurance industry. However, ML models are at risk of being attacked and manipulated. A model compromised by a backdoor attack loses its integrity and can no longer be deemed trustworthy. Ensuring the trustworthiness of ML models is crucial, as compromised models can lead to significant financial and reputational damage for insurance companies. In this work the robustness of Gradient Boosted Decision Tree (GBDT) models and Deep Neural Networks (DNN) within an insurance context is evaluated. Therefore, two GBDT models and two DNNs are trained on two different tabular datasets from an insurance context. Past research in this domain mainly used homogeneous data and there are comparably little insights regarding heterogeneous tabular data. The ML tasks performed on the datasets are claim prediction (regression) and fraud detection (binary classification). For the backdoor attacks different samples containing a specific pattern were crafted and added to the training data. It is shown, that this type of attack can be highly successful, even with a few added samples. The backdoor attacks worked well on the models trained on one dataset but poorly on the models trained on the other. In real-world scenarios the attacker will have to face several obstacles but as attacks can work with very few added samples this risk should be evaluated. Therefore, understanding and mitigating these risks is essential for the reliable deployment of ML in critical applications.