Hamed Yazdanpanah

Automated credit risk assessment plays an important role in agricultural lending. However, credit risk assessment in the agricultural domain has unique challenges due to the impact of weather, pest outbreaks, commodities market dynamics, and other volatile forces that drive risk. Training a model to account for these factors requires immense data assets that are challenging to obtain. Indeed, even the best credit risk assessment models in this domain are trained using data from single-institutions that often focus on dedicated geographical regions, or singular commodities. Hence, most agricultural credit risk models exhibit poor out-of-domain performance. In this paper, we use a novel dataset describing nearly 100 thousand historical loans, sourced from 9 large agricultural lenders to train a Bayesian network model for loan delinquency classification. The proposed model exhibited improved calibration (relative improvement in Expected Calibration Error) in out-of-domain performance tests when compared to three state-of-the-art credit risk scoring approaches: Logistic regression (81 ± 15% improvement), XGBoost (80 ± 14% improvement), and an Artificial Neural Networks (7 ± 2% improvement). We conclude that Bayesian networks provide better modeling of agricultural credit risk by combining (limited) data assets with expert domain knowledge. Our approach is likely to generalize to any credit risk assessment task where small sample sizes is of concern.

Credit risk analysis is a process that involves a wide range of complex cognitive abilities. Automating the credit risk analysis process using Large Language Models can bring transformative changes to the finance industry, but not without appropriate measures to ensure trustworthy responses. In this work, we propose a novel prompt-engineering method that enhances the ability of Large Language Models to generate reliable credit risk reports - Labeled Guide Prompting (LGP). LGP consists of: (1) providing annotated few-shot examples to the LLM that denote sets of tokens in an exemplary prompt that are of greater importance when generating sets of tokens in the exemplary response and (2) providing text in the prompt that describes the direction, pathways and interactions between variables from a Bayesian network used for credit risk assessment, thus promoting abductive reasoning. Using data from 100 credit applications, we demonstrate that LGP enables LLMs to generate credit risk reports that are preferred by human credit analysts (in 60-90% of cases) over alternative credit risk reports created by their peers in a blind review. Additionally, we found a statistically significant improvement (p-value < 10− 10) in the insightfulness of the responses generated using LGP when compared to identical prompts without LGP components. We conclude that Labeled Guide Prompting can enhance LLM performance in complex problem-solving tasks, achieving a level of competency comparable to or exceeding human experts.

In this work, we propose a new frequency-domain adaptive filtering algorithm within the set-membership filtering (SMF) paradigm. Algorithms employing the previously proposed frequency-domain SMF (FD-SMF) concept are not capable of performing linear convolutions (i.e., linear filtering operations), but only circular convolutions. Here, we overcome such limitation by generalizing the FD-SMF concept to allow for linear convolutions, and we propose a new algorithm following this extended FD-SMF concept. Synthetic and real-life simulation results confirm that the proposed algorithm can perform quite better than its non-SMF counterpart, achieving faster convergence and requiring much fewer computations.

Recently, the improved simple set‐membership affine projection (IS‐SM‐AP) algorithm has been proposed in order to exploit sparsity in system models. Although its update equation resembles that of the set‐membership affine projection (SM‐AP) algorithm, the IS‐SM‐AP algorithm has two fundamental advantages over the SM‐AP algorithm: (i) it can exploit sparsity in system models and (ii) its computational complexity is lower. Up to now, the properties of this algorithm have been addressed only through numerical simulations, and no analytical study has been presented. To fill this gap, in this study, the authors analyse the steady‐state mean squared error of the IS‐SM‐AP algorithm using the energy conservation method. Furthermore, some important implementation issues are addressed, and a time‐varying parameter for the discard function is proposed. Finally, the authors present numerical results corroborating the theoretical analysis and the effectiveness of the proposed time‐varying parameter.

We have been witnessed a growing research activity to advance new strategies to detect and exploit underlying sparsity in the parameters of physical models. In many cases, the sparsity is not explicit in the relations among the parameter coefficients requiring some suitable tools to reveal the potential sparsity. This work proposes a family of adaptive filtering algorithms, aimed at exposing some hidden features of the unknown parameters. Although the basic idea applies to any algorithm, we will concentrate the work in the LMS-type algorithms, giving rise to a family collectively named as Feature LMS (F-LMS) algorithms. These algorithms increase the convergence speed and reduce the steady-state mean-squared error, in comparison with the classical LMS solution. The main idea is to apply linear transformations, through the so-called feature matrices, to reveal the sparsity hidden in the coefficient vector, followed by a sparsity-promoting penalty function to exploit the exposed sparsity. For illustration, a few F-LMS algorithms for lowpass, bandpass, and highpass systems are introduced by using simple feature matrices that require either only simple operations or can learn the features. Simulations and real-life experiments demonstrate that the F-LMS algorithms bring about several performance improvements whenever the unknown sparsity of parameters is exposed.

Recently, a great effort has been observed to exploit sparsity in physical parameters; however, in many cases, sparsity is hidden in relations between parameters, and some appropriate tools should be utilized to expose it. In this paper, a family of algorithms called feature normalized least-mean-square (F-NLMS) algorithms is proposed to exploit hidden sparsity. The major key of these algorithms is a so-called feature matrix that transforms non-sparse systems to new systems containing sparsity; then the revealed sparsity is exploited by some sparsity-promoting penalty function. Numerical results demonstrate that the F-NLMS algorithms can reduce the steady-state mean-squared-error (MSE) and/or improve the convergence rate of the learning process significantly.

Epitheliod hemangioendothelioma (EHE) is a rare spinal neoplasm with vascular provenance, which can present as a primary lesion or metastatic lesion with th ...