Adrian Costea

A retraining neural network technique for glass manufacturing data forecasting

This paper advances a retraining-neural-network-based forecasting mechanism that can be applied to complex prediction problems, such as the estimation of relevant process variables for glass manufacturing. The main purpose is to obtain a good accuracy of the predicted data by using an optimal feedforward neural architecture and well-suited delay vectors. The artificial neural networks (ANNs) ability to extract significant information provides a valuable framework for the representation of relationships present in the structure of the data. The evaluation of the output error after the retraining of an ANN shows that the retraining technique can substantially improve the achieved results.

An integrated two-stage methodology for optimising the accuracy of performance classification models

In this paper we propose a two-stage methodology to classify the non-banking financial institutions (NFIs) based on their financial performance. The first stage of the methodology consists of grouping the companies in similar financial performance classes (e.g.: “good”, “average”, “poor” performance classes). We optimise the allocation of the observations within the performance clusters by applying an enhanced version of an observation re-allocation procedure proposed in our previous work. Next, based on the result of the grouping phase, we construct a performance class variable by attaching a performance label to each data row. Then, in the second phase of our methodology, we propose a feed-forward neural-network classification model that maps the input space to the newly-constructed performance class variable. This model allows us to forecast the performance of new companies as data become available.

Applying Fuzzy Logic and Machine Learning Techniques in Financial Performance Predictions

Abstract In this article we apply a fuzzy logic technique, namely Fuzzy C-Means clustering, and artificial intelligence algorithms for evaluating comparatively the financial performance of non-banking financial institutions (NFIs) in Romania. The NFIs’ performance dataset consists of indicators that define the capital adequacy, assets’ quality and profitability performance dimensions. The class performance variable is obtained by applying on the performance dataset the Fuzzy C-Means algorithm and obtaining clusters with similar performance. We attach to each input dataset observation a performance class depending on which cluster contains the observation given the characterization and hierarchy of the clusters in “good”, “medium” and “poor” performance clusters. Finally, we apply artificial neural networks (ANNs) trained with genetic algorithms in order to find a function that maps the input performance space on the newly constructed performance class variable. The classification model obtained can be used by different beneficiaries (e.g.: the Supervision Department of National Bank of Romania) to classify new NFIs as having a “good” or “poor” performance so that the limited resources of the supervision authority to be better allocated.

A Statistical-Based Approach to Assessing Comparatively the Performance of Non-Banking Financial Institutions in Romania

In this paper we construct a framework that enables us to make class predictions about the performance of non-banking financial institutions (NFIs) in Romania. Our objective is to create a classification model in the form of a logistic regression function that can be used to assess the performance of NFIs based on different performance dimensions, such as capital adequacy, assets’ quality and profitability. Our methodology consists of two phases: a clustering phase, in which we obtain several clusters that contain similar data-vectors in terms of Euclidean distances, and a classification phase, in which we construct a class predictive model in order to place the new row data within the clusters obtained in the first phase as they become available. Our goal is two-fold: to validate the dimensionalities of the map used to represent the performance clusters and the quantisation error associated with it and to use the obtained model to analyze the movements of three largest NFIs during the period 2007–2010. Using our validation procedure that is based on a bootstrap technique, we are now able to find the proper map architecture and training–testing dataset combination for a particular problem. At the same time, using the visualization techniques employed in the study, we understand how different financial factors can and do contribute to the companies’ movements from one group/cluster to another. Furthermore, the classification model is validated based on high training and testing accuracy rates.