Philippe du Jardin

Predicting bankruptcy using neural networks and other classification methods: The influence of variable selection techniques on model accuracy

We evaluate the prediction accuracy of models designed using different classification methods depending on the technique used to select variables, and we study the relationship between the structure of the models and their ability to correctly predict financial failure. We show that a neural network based model using a set of variables selected with a criterion that it is adapted to the network leads to better results than a set chosen with criteria used in the financial literature. We also show that the way in which a set of variables may represent the financial profiles of healthy companies plays a role in Type I error reduction.

Predicting corporate bankruptcy using a self-organizing map: An empirical study to improve the forecasting horizon of a financial failure model

The aim of this study is to show how a Kohonen map can be used to increase the forecasting horizon of a financial failure model. Indeed, most prediction models fail to forecast accurately the occurrence of failure beyond 1year, and their accuracy tends to fall as the prediction horizon recedes. So we propose a new way of using a Kohonen map to improve model reliability. Our results demonstrate that the generalization error achieved with a Kohonen map remains stable over the period studied, unlike that of other methods, such as discriminant analysis, logistic regression, neural networks and survival analysis, traditionally used for this kind of task.

Bankruptcy prediction using terminal failure processes

Traditional bankruptcy prediction models, designed using classification or regression techniques, achieve short-term performances (1 year) that are fairly good, but that often worsen when the prediction horizon exceeds 1 year. We show how to improve the performance of such models beyond 1 year using models that take into account the evolution of firm’s financial health over a short period of time. For this purpose, we design models that fit the underlying failure process of different groups of firms. Our results demonstrate that such models lead to better prediction accuracy at a 3-year horizon than that achieved with common models.

A two-stage classification technique for bankruptcy prediction

Ensemble techniques such as bagging or boosting, which are based on combinations of classifiers, make it possible to design models that are often more accurate than those that are made up of a unique prediction rule. However, the performance of an ensemble solely relies on the diversity of its different components and, ultimately, on the algorithm that is used to create this diversity. It means that such models, when they are designed to forecast corporate bankruptcy, do not incorporate or use any explicit knowledge about this phenomenon that might supplement or enrich the information they are likely to capture. This is the reason why we propose a method that is precisely based on some knowledge that governs bankruptcy, using the concept of “financial profiles”, and we show how the complementarity between this technique and ensemble techniques can improve forecasts.

Dynamics of firm financial evolution and bankruptcy prediction

Abstract The optimal forecasting horizon of bankruptcy prediction models is usually one year. Beyond this point, their accuracy decreases as the horizon recedes. However, the ability of models to provide good mid-term forecasts is an essential characteristic for financial institutions due to prudential reasons. This is why we have studied a method of improving their forecasts up to a 5-year horizon. For this purpose, we propose to quantize how firm financial health changes over time, typify these changes and design models that fit each type. Our results show that, whatever the modeling technique used to design prediction models, model accuracy can be significantly improved when the horizon exceeds two years. They also show that when our method is used in combination with ensemble-based models, model accuracy is always improved whatever the forecasting horizon, when compared to traditional models used by financial institutions. The method we propose in this article appears to be a reliable solution that makes it possible to solve a real problem most models are unable to overcome, and it can therefore help financial companies comply with the current recommendations made by the Basel Committee on Banking Supervision. It also provides the scientific community (which is interested in designing reliable failure models) with insights about how the evolution of firms’ financial situations over time can be modeled and efficiently used to make forecasts.

SOM-ELM-Self-Organized Clustering using ELM

This paper presents two new clustering techniques based on Extreme Learning Machine (ELM). These clustering techniques can incorporate a priori knowledge (of an expert) to define the optimal structure for the clusters, i.e. the number of points in each cluster. Using ELM, the first proposed clustering problem formulation can be rewritten as a Traveling Salesman Problem and solved by a heuristic optimization method. The second proposed clustering problem formulation includes both a priori knowledge and a self-organization based on a predefined map (or string). The clustering methods are successfully tested on 5 toy examples and 2 real datasets.

Comparing discriminant analysis and neural network for the determination of sex using femur head measurements

The measurement of the femoral head is usually considered an interesting variable for the sex determination of skeletal remains. To date, there are few published reference measurements of the femoral head in a modern European population for the purpose of sex determination. In this study, 116 femurs from 58 individuals of the South of France (Nice Bone Collection, Nice, France) were studied. Three measurements of the femoral head were taken: the vertical head diameter (VHD), the transversal head diameter (THD) and the head circumference (HC). The results show that: (i) there is no statistical difference between the right and left femurs for each of the three measurements (VHD, THD and HC). Therefore we arbitrarily chose to use the measures from the right femurs (N=58) to pursue our experiments; (ii) the measurements of the femoral head are similar to those of contemporary American populations; (iii) the dimensions of the femoral head place the measurements of the French population somewhere between Germany or Croatia, and Spain; (iv) there is no significant secular trend (in contrast with the femoral neck diameter); (v) the femoral head measurement as a single variable is useful for sex determination: a 96.5% rate of accuracy was obtained using THD and HC measurements with the artificial neural network; and a 94.8% rate of accuracy using VHD, both with the discriminant analysis and the neural network.

Positive identification by X-rays bone trabeculae comparison

Positive (certain, absolute) identification of human remains needs a scientific comparison between ante mortem and post-mortem biologic features, as fingerprint, odontological, radiological or DNA comparisons. X-rays comparison has been extensively used, usually comparing some peculiarities such as outlines of the bones, degenerative evolution or pathological conditions. Trabeculae comparisons are sparsely underlined in the forensic literature. We report on a case of decomposed body where fingerprint, DNA and odontological comparisons were not possible. After dissecting the leg and preparing the bones, comparison of ante mortem and postmortem trabeculae led to a positive identification. It was observed that tens of radiolucencies and radiodensities drawn by the trabeculae were useful for comparison, within a very small part of bone. In the case reported here the positive identity could have been assessed only by the comparison of the first metatarsal. The statement of positive identification needs scientific criteria that will be discussed in this article.

MD-ELM

This paper proposes a methodology for identifying data samples that are likely to be mislabeled in a c-class classification problem (dataset). The methodology relies on an assumption that the generalization error of a model learned from the data decreases if a label of some mislabeled sample is changed to its correct class. A general classification model used in the paper is OP-ELM; it also provides a fast way to estimate the generalization error by PRESS Leave-One-Out. It is tested on two toy datasets, as well as on real life datasets for one of which expert knowledge about the identified potential mislabels has been sought.

Comparison of macroscopic and microscopic (stereomicroscopy and scanning electron microscopy) features of bone lesions due to hatchet hacking trauma

This experimental study examined the lesions produced by a hatchet on human bones (tibiae). A total of 30 lesions were produced and examined macroscopically (naked eye) and by stereomicroscopy. 13 of them were also analyzed using scanning electron microscopy. The general shape of the lesion, both edges, both walls, the kerf floor and the extremities were described. The length and maximum width of the lesions were also recorded. The microscopic analysis of the lesions led to the description of a sharp-blunt mechanism. Specific criteria were identified (lateral pushing back, fragmentation of the upraising, fossa dug laterally to the edge and vertical striae) enabling the forensic expert to conclude that a hacking instrument was used. These criteria are easily identifiable using scanning electron microscopy, but can also be observed with stereomicroscopy. Overall, lateral pushing back and vertical striae visible using stereomicroscopy and scanning electron microscopy signal the use of a hacking tool.