Predictive and decision models in finance are typically validated under assumptions of distributional stability over the evaluation window. In deployment, those assumptions often fail: the data-generating process undergoes structural changes—breaks, regime transitions, and drift—that can invalidate conditional relationships, degrade calibration, and amplify tail risk precisely when decisions are most consequential. Despite a large literature, results remain hard to reconcile across econometrics, statistical monitoring, and machine learning due to divergent terminology and incompatible evaluation protocols. This survey addresses this fragmentation through a taxonomy-based methodological synthesis of financial non-stationarity, linking drift signatures to design trade-offs in representation, detection, adaptation, and evaluation. We contribute: (1) a unified taxonomy of drift and regime change; (2) a structured, literature-grounded synthesis of drift-aware representation, change detection, and continuous adaptation methods; and (3) scenario-based evaluation guidance to support future comparative studies under non-stationarity in financial time series. We align terms such as structural breaks, regimes, concept drift, and dataset shift, and organize them through a taxonomy that combines temporal, statistical, spatial, ontological, and causal dimensions to describe real drift scenarios consistently. Using this lens, we review drift-aware representations, change detection methods, and continuous adaptation strategies—from classical sequential monitoring and segmentation to Bayesian, multivariate, and embedding-based out-of-distribution approaches. We then consolidate evaluation guidance spanning detection delay, false-alarm control, computational cost, and finance-specific utility. Finally, we highlight emerging directions (foundation models, multimodal context, parameter-efficient adaptation) and open challenges in benchmark design and reliable online calibration.

In the evolving landscape of software engineering, accurate identification of defect-inducing commits is critical to improving software quality and reducing development costs. This paper revisits the widely adopted SZZ algorithm, which is utilized for labeling commits as clean or defect-inducing, to address one of its main limitations, i.e., its reliance on outdated corrective commits identification strategies. We propose an innovative approach that integrates the semantic understanding capability of the GPT OpenAI model into the SZZ flow to better interpret commit messages. Our experiments reveal, for some projects, a large number of commits incorrectly interpreted as defect-fixing, consequently, leading to the misclassification of commits as defect-inducing. As an example, for the Postgresql dataset, the number of defect-inducing commits was reduced in 21% when compared to the original SZZ. Furthermore, results of our experiments strongly suggest that, as a result of the proposed SZZ labeling process, the JIT-SDP problem has been shown to be more challenging than originally reported by previous works.

Just-in-time Software Defect Prediction (JIT-SDP) aims to detect potential defects early, helping to prevent risky code from entering the repository during development. This study evaluates JIT-SDP using pre-trained language models in different architectures and settings. It compares open-source fine-tuned models such as CodeT5+ and UniXCoder with closed LLMs such as GPT and Gemini. This is the first known study to compare trainable open and prompt-based closed decoder-only models for JIT-SDP. The main results show that fine-tuned open models outperform closed models in zero-shot and few-shot scenarios without advanced prompt engineering techniques, and in cross-project tasks, CodeT5+ and UniXCoder surpass previous state-of-the-art results. The findings underscore the value of model architecture, fine-tuning, and expert features for effective defect prediction. Finally, we introduce CodeFlowLM - to our knowledge, the first framework for continual JIT-SDP using pre-trained language models.

This work introduces CodeFlowLM, an incremental learning framework for Just-In-Time Software Defect Prediction (JIT-SDP) that leverages pre-trained language models (PLMs). Unlike traditional online learners, CodeFlowLM employs continual fine-tuning to address concept drift, class imbalance, and verification latency without retraining from scratch. We evaluated encoder-only and encoder-decoder PLMs (notably CodeT5+ and UniXCoder) in JIT-SDP scenarios within and between projects, comparing them with the incremental baseline BORB. The results show that CodeFlowLM achieves up to 68

Legislative documents present substantial obstacles to summarization due to their complex argument structures and specialized terminology. This research investigates the application of Large Language Models (LLMs) in summarizing Brazilian legislative proposals from the Chamber of Deputies, examining a dataset of over 56 thousand texts from 2013 to 2023. The paper explores three main summarization methodologies: extractive, abstractive, and hybrid, with an emphasis on abstractive summarization using LLMs. The performance of the LLM LLAMA2-13b is assessed using metrics such as ROUGE, BLEU, METEOR, BERTScore, and BERTopic, compared against reference summaries. The results show that LLMs can generate coherent and informative summaries, with positive evaluation metric results. Notably, the study reveals that traditional summary evaluation metrics may not be adequate for evaluating LLMs in summarization tasks. On the other hand, metrics based on pre-trained models like BERT provide a more effective evaluation of this innovative automatic summarization approach.

Modelos baseados em BERT vêm sendo largamente utilizados, tornando-se o estado da arte para muitas tarefas de Processamento de Linguagem Natural e também para Recuperação de Informação. A arquitetura Sentence-BERT permitiu que esses modelos fossem facilmente utilizados para a busca semântica de documentos, já que ela gera embeddings contextuais que podem ser comparados através de medidas de similaridade. Para melhor investigar a aplicação de modelos baseados em BERT para Recuperação de Informação, este trabalho avaliou 12 modelos Sentence-BERT, disponíveis publicamente, para a recuperação de documentos no cenário legislativo brasileiro. Duas variantes do algoritmo BM25 foram utilizadas como baseline: Okapi BM25 e BM25L. O BM25L alcançou melhores resultados, com significância estatística, mesmo no cenário em que os documentos não foram pré-processados, enquanto que apenas um dos modelos de linguagem, ajustado usando dados legislativos brasileiros, obteve um desempenho similar para uma das três bases de dados utilizadas

A time series is a sequence of numerical data arriving sequentially as a data stream. Typically, data streams’ generating distributions change over time (i.e., concept drifts), potentially degrading predictive model performance. Despite extensive literature on time series forecasting and concept drift handling separately, few studies address both problems together. Recently, Particle Swarm Optimization (PSO) has been applied to dynamic optimization problems and training neural networks for forecasting. Building on these PSO applications, we propose the first approach using dynamic swarm intelligence to address concept drift in time series forecasting. Our approach includes three strategies: (i) using swarm models to detect concept drifts; (ii) managing swarm models to adapt to new concepts; and (iii) utilizing past swarm models for recurrent drifts. Experiments with six synthetic and four real-world datasets demonstrated that the proposed approach was competitive in concept drift detection, achieving the best forecasting errors compared to existing methods. This suggests dynamic swarm intelligence is a promising method for handling concept drift in time series forecasting.

Just-in-Time Software Defect Prediction (JIT-SDP) operates in an online scenario where additional training data is received over time. Existing online JIT-SDP studies used online Oza ensemble learning methods with Hoeffding Trees as base learners to learn and update JIT-SDP models over time in this scenario. However, it is unknown how these approaches compare against offline learning approaches adapted to operate in online scenarios, and how the use of any other online or offline base learners would affect online JIT-SDP in terms of predictive performance and computational cost. We therefore propose a new approach called Batch Oversampling Rate Boosting (BORB) that is able to use offline base learners in an online JIT-SDP scenario. Based on 10 open source projects, we provide a comprehensive evaluation of BORB with 5 different base learners and the existing online approach Oversampling Rate Boosting with 4 different base learners, both in within-project and cross-project online JIT-SDP scenarios. The results show that offline learning can lead to better predictive performance than the top performing online learning approaches considered in our study, at a higher computational cost. Cross-project data was helpful to improve predictive performance both for offline and online learning, but especially for online learning.

The increasing use of artificial intelligence methods in the legal field has sparked interest in applying Natural Language Processing techniques to handle legal tasks and reduce the workload of these professionals. However, the availability of legal corpora in Portuguese, especially for the Brazilian legal domain, is limited. Existing resources offer some legal data but lack comprehensive coverage. To address this gap, we present Ulysses Tesemõ, a large corpus specifically built for the Brazilian legal domain. The corpus consists of over 3.5 million files, totaling 30.7 GiB of raw text, collected from 159 sources encompassing judicial, legislative, academic, news, and other related data. The data was collected by scraping public information from governmental websites, emphasizing contents generated over the past two decades. We categorized the obtained files into 30 distinct categories, covering various branches of the Brazilian government and different types of texts. The corpus retains the original content with minimal data transformations, addressing the scarcity of Portuguese legal corpora and providing researchers with a valuable resource for advancing in the research area.

Symbolic boxplot data can be considered as a particular case of the numerical multi-valued variable. This kind of symbolic data is an useful exploratory tool with a simple structure for summarizing groups of numerical data. However, in the literature of symbolic data analysis it has been little explored. In this paper, we propose a new prediction method for extracting knowledge from boxplot data. A parametrized regression approach automatically extracts the best reference points from the regressor variables. These reference points are then used to build five linear regression models based on values of the boxplot: minimum (m), lower quartile (Q1), median (Q2), upper quartile (Q3) and maximum (M). A strategy based on BoxCox transformation is applied to the response variable in order to guarantee the mathematical coherence of the predictions and build the boxplot. Experimental evaluation with synthetic and real boxplot datasets illustrates the advantages of the proposed method. Moreover, the present work also focuses in the development of an application for predicting temperature data based on boxplot in the Brazilian Electric Sector.

The proper functioning of judicial and legislative institutions requires the efficient retrieval of legal documents from extensive datasets. Legal Information Retrieval focuses on investigating how to efficiently handle these datasets, enabling the retrieval of pertinent information from them. Relevance Feedback, an important aspect of Information Retrieval systems, utilizes the relevance information provided by the user to enhance document retrieval for a specific request. However, there is a lack of available corpora containing this information, particularly for the legislative scenario. Thus, this paper presents Ulysses-RFCorpus, a Relevance Feedback corpus for legislative information retrieval, built in the real-case scenario of the Brazilian Chamber of Deputies. To the best of our knowledge, this corpus is the first publicly available of its kind for the Brazilian Portuguese language. It is also the only corpus that contains feedback information for legislative documents, as the other corpora found in the literature primarily focus on judicial texts. We also used the corpus to evaluate the performance of the Brazilian Chamber of Deputies' Information Retrieval system. Thereby, we highlighted the model's strong performance and emphasized the dataset's significance in the field of Legal Information Retrieval.

Deep learning methods have emerged as state-of-the-art techniques for learning handwritten signature feature representations. However, successful results in deep learning require a significant amount of training data. The GPDS-960 dataset used to be the largest publicly available dataset of offline handwritten signatures for training deep models. However, due to data protection regulatory issues, the GPDS-960 dataset is no longer publicly available. This way, new investigators starting research in this field have suffered from the absence of a largescale real signature dataset and have resorted to adopt the GPDSsynthetic dataset, a large-scale set of synthetically generated signatures for training models. Nevertheless, we have found a difference in verification performance between models trained using real and synthetic signature data. To deal with this problem, we apply a method based on datafree knowledge transfer learning. Firstly, we generate inverted examples with the same distribution as the real examples. Then, we complement a feature space based on real data using synthetic data while minimizing the divergence in distribution between the representations provided by these two different types of data sources. This is achieved through continual learning based on knowledge distillation. We evaluated models obtained with the proposed method in terms of the equal error rate on GPDSsynthetic, GPDS-300, CEDAR, and MCYT-75 datasets in a writer-dependent verification approach. Experiments demonstrated that the proposed method provides a more robust model for writer-dependent verification when considering real and synthetic signature datasets. Inverted data is available for download at https://github.com/ tallesbrito/continual sigver.

Named Entity Recognition (NER) is an important task in Natural Lan-guage Processing, as it is a key information extraction sub-task with numerous ap-plications, such as information retrieval and machine learning. However, resources are still scarce for some languages, as it is the case of Portuguese. Thus, the objective of this research is to map NER techniques, methods and resources for the Portugue-se language. Manual and automated searches were applied, retrieving 447 primary studies, of which 45 were included in our review. The growing number of studies reveal a greater interest of researchers in the area. 21 studies focused on the compa-rative analysis between techniques and tools. 24 new or updated NER corpora were mapped, in several domains. The most used text pre-processing techniques were to-kenization, embeddings, and PoS Tagging, while the most used methods/algorithms were based on BiLSTM, CRF, and BERT models. The most relevant researchers, institutions and countries were also mapped, as well as the evolution of publications.

Este trabalho investiga frameworks de recuperação de informações para lidar com as dificuldades existentes no processo de elaboração de leis da Câmara dos Deputados do Brasil. Dois frameworks de código aberto foram escolhidos. Além disso, foram avaliadas diferentes técnicas de pré-processamento, incluindo stemmers e modelos de linguagem n-gram. Dois corpora legislativos da Câmara foram usados para construir e validar os experimentos. Os resultados foram comparados com um baseline utilizado pela Câmara dos Deputados. O baseline apresentou o melhor resultado, alcançando um Recall para 20 documentos de 0,7376.

Deep learning models often require large datasets to perform well from scratch. Transfer learning methods solve this issue by using a pre-trained source network to improve a target network training. Recent approaches involve using feature maps from the source network to guide the target network training. The latest transfer learning methods use meta-networks to enhance the knowledge transfer process. These meta-networks bridge the source and target networks, deciding which pairs of feature map layers and channels should be matched for optimal knowledge transfer. This paper improves this approach by using pixel-level information, in addition to layers and channels, for better knowledge transfer. Our experiments on multiple datasets show that the proposed approach outperforms previous baselines in scenarios with limited labels per class. The source code is available at https://github.com/lucasdelimanogueira/L2T-www.

Context: Process Mining (PM) aims to discover processes and their characteristics from event logs recorded by information systems. There are dozens of general-purpose tools. The Brazilian judiciary wants to make the technology available to magistrates with little or no knowledge of the field of PM. Problem: The usability and comprehensibility of the available tools prevent their adoption by laypeople. In fact, these are two of the eleven challenges proposed by the IEEE Task Force on PM that are still little explored in the context of non-specialists. Methodology: Applied qualitative research, using User Centered Design (UCD) principles to guide the construction of the JuMP tool, conceived considering anthropological and sociological aspects of the Brazilian judiciary. For a year and six months, a team of PM specialists worked with representatives of the judiciary sector to produce, evaluate and evolve the new product. An experiment was performed to evaluate the usability and understandability of JuMP. Results: The study demonstrated that the use of a PM tool oriented to the application domain is fundamental for domain experts with little or no knowledge in PM to be able to make good use of it. Contribution: Demonstration that, unlike the prevailing practice in the area of PM, which prioritizes the provision of general purpose tools, the design of tools oriented to the application domain are prerequisites to improve usability and comprehensibility attributes. Finally, the work proposes a generic and adaptable methodology for developing specific-purpose PM tools and the experience of its implications to the design of the JuMP tool.

In spite of recent advances in computer vision, the classic problem of offline handwritten signature verification still remains challenging. The signature verification task has a high intra-class variability because a given user often shows high variability between its samples. Besides, signature verification is harder in the presence of skilled forgeries. Recently, in order to tackle these challenges, the research community has investigated deep learning methods for learning feature representations of handwritten signatures. When mapping signatures to a feature space, it is desired to obtain dense clusters of signature’s representations, in order to deal with intra-class variability. Besides, not only dense clusters are required but also a larger separation between different user’s clusters in the feature space. Finally, it is also desired to move away feature representations of skilled forgeries in relation to the respective dense cluster of genuine representations. This last property is hard to achieve in the real-world scenario because skilled forgeries are not readily available during training. In this work, we hypothesize that such properties can be achieved by means of a multi-task framework for learning handwritten signature feature representations based on deep contrastive learning. The proposed framework is composed of two objective-specific tasks. The first task aims to map signature examples of the same user closer within the feature space, while separating the feature representations of signatures of different users. The second task aims to adjust the skilled forgeries representations by adopting contrastive losses with the ability to perform hard negative mining. Hard negatives are examples from different classes with some degree of similarity that can be applied for training. We evaluated models obtained with the proposed framework in terms of the equal error rate on GPDSsynthetic, CEDAR and MCYT-75 datasets in writer-dependent and writer-independent verification approaches. Using synthetic and real signature datasets, Friedman tests with Bonferroni–Dunn post hoc tests were performed to compare the proposed multi-task contrastive models against the popular SigNet model as a baseline. Experiments demonstrated an statistically significant improvement in signature verification with a multi-task contrastive model based on the Triplet loss. Implementation of the method is available for download at https://github.com/tallesbrito/contrastive_sigver.

Ensemble Effort Estimation (EEE) techniques combine several individual software estimation methods in order to address the weaknesses of individual methods for prediction tasks. A systematic review published in 2016 analyzed empirical studies on EEE techniques published between 2010 and (January) 2016. The research on EEE has continuously evolved over the past five years (2016–2020), generating new findings that should be aggregated to the existing body of evidence on the subject. The goal of this paper is to update the systematic review from 2016 with new findings from studies published between 2016 (full year) and 2020 (inclusive). To conduct our review update, we followed existing guidelines for updating systematic reviews in software engineering and other fields. We started with an appraisal of the background and methods of the 2016 review, which resulted in the updated review protocol used to conduct our study. We retrieved 3,682 papers using automatic searching techniques, from which we selected 30 papers for data extraction and analysis. Our findings reinforce the results of the previous review in that machine learning is still the technique most common to construct EEE and that the ensemble techniques have outperformed the individual models. We added new evidence showing that there is no clear superiority of an EEE model over the others. Also, we found that ensemble dynamic selection is still little used in Software Effort Estimation (SEE). This review adds new evidence about the use of EEE techniques in software development which reinforces previous findings and also shows research opportunities in constructing more effective EEE. Besides, ensemble dynamic selection appears as a promising area of research which still is underexplored. • Updated Systematic review of 30 selected studies on Ensemble Effort Estimation (EEE). • Studies analyzed based on six review questions and eight assessment quality criteria. • EEE techniques yield better estimation accuracy than single models. • There is no evidence to point out which EEE method is best. • Dynamic ensemble selection is still scarcely used in EEE.

A large amount of legal and legislative documents are generated every year with highly specialized content and significant repercussions on society. Besides technical, the produced information is not semantically standardized or format structured. Automating the document analysis, categorization, search, and summarization is essential. The Named Entity Recognition (NER) task is one of the tools that have the potential to extract information from legal documents with efficiency. This paper evaluates the state-of-the-art NER models BiLSTM+CRF and BERT+Fine-Tunning trained on Portuguese corpora through finetuning in the legal and legislative domains. The obtained results (F1-scores of 83.17% and 88.27%) suggest that the BERT model is superior, achieving better average results.