ABSTRACT Machine learning‐based forecasting has been gaining attention for its flexibility and adaptability. Given the dynamics of temporal data, clustering can be an effective approach for improving forecasting accuracy. While clustering for classification typically considers all observations in the data to effectively divide the data into clusters, clustering for temporal data forecasting may differ. Instead of focusing on dividing the entire data, identifying a single cluster close to the future value may yield better forecasting accuracy. A method to determine the single cluster by dividing a cluster obtained from previous clustering is adopted in this study, thereby improving the forecasting accuracy of clustering methods. Various financial assets are used to validate this approach across different time intervals. The empirical validation demonstrates the improved forecasting accuracy of the method, suggesting its broad applicability in other forecasting contexts.
We propose a state-space approach for a non-parametric tv-GARCH model. Using a number of non-parametric techniques, we estimate slowly time-varying curves for the parameters and combine it with a recursive system based on the Kalman filter as a methodological framework to obtain forecasts from the model. Our results based on Monte Carlo simulations and an application to the Selective Stock Price Index of the Chilean stock market indicate that the non-parametric tv-GARCH model shows the best fit and provides the best forecasting performance, compared with the stationary GARCH model and other techniques, such as tv-GARCH with time-varying parameters determined by deterministic functions. These results have relevant implications for risk management, portfolio diversification, and asset allocation.
Federated learning generally suffers from slow convergence and suboptimal model performance due to client drift caused by data heterogeneity. Although extensive research has been devoted to addressing these issues, adaptive optimization algorithms that combine model training states and are based on client contribution evaluation remain scarce. With the increasing diversity of data distributions and complexity of model architectures, the convergence of federated learning models often requires more training iterations. To verify the application value of adaptive strategies in federated learning scenarios, this study proposes a Learnable Aggregate Federated Learning algorithm with initialization based on contribution evaluation (LAFL-AI). At the initial stage of each training round on the server, the algorithm first quantifies the contribution of each client to the global model by exploring the correlation between local and global gradients, and adaptively initializes the aggregation weights of each local model for the global aggregation process directly based on the results of this contribution evaluation. Meanwhile, to further improve the generalization performance of the global model, a class-balanced proxy dataset is constructed in this study to train and obtain the optimal weights for global aggregation. Experimental results show that the adaptive initialization strategy for aggregation weights based on contribution evaluation proposed in this paper can not only accelerate the training efficiency of aggregation weights but also significantly improve the convergence speed of the global model. Compared with other baseline algorithms, LAFL-AI enables fast convergence of the global model and can achieve equivalent or even better training results with fewer server training rounds.
Traditional econometric models frequently fail to adequately reflect the ability to estimate future values of macroeconomic and financial variables, which is crucial for the implementation of macroeconomic policies. This study aims to tackle this problem by predicting the financial stress index (FSI) of the Eurozone for a sample period spanning from January 1995 to August 2023 using artificial neural networks (ANN) optimized by bio-inspired optimisation algorithms like the Invasive Weed Optimization (IWO), Firefly Algorithm (FA), Particle Swarm Optimization (PSO), Cultural Algorithm (CA), and Artificial Bee Colony (ABC). The prediction error of the ANN is greatly reduced by the ANN optimized by FA (ANN-FA) model, which reduces it by roughly 97.57% for the calibration sample and 97.94% for the validation sample. The second-best model for FSI prediction in the Eurozone is the ANN optimized by PSO (ANN-PSO), which reduces the ANN's prediction error by 94.48% and 95.38% for the calibration and validation samples, respectively. The models are assessed for performance using accuracy metrics, and the metrics are visualized using chord, bump, and Taylor diagrams. These results imply that bio-inspired optimization can significantly enhance predictive reliability in monitoring financial stress conditions.
ABSTRACT Accurately forecasting European carbon prices is essential for effective climate policy and market risk management yet remains challenging due to the coexistence of exogenous macroeconomic shocks, endogenous market momentum, and pronounced spatial and temporal dependencies across related markets. Conventional forecasting models often fail to accommodate such spatial–temporal heterogeneity, leading to unstable predictive performance and limited economic interpretability. To address these challenges, we introduce a heterogeneous ensemble framework that integrates the strengths of variable selection, frequency‐domain decomposition, and graph‐based spatial learning. Using 41 variables, we employ LASSO to extract endogenous market momentum, implement the MEMD‐ARIMAX‐mLSTM framework to model temporal dynamics, and adopt GWnet‐attn to capture spatial interdependencies. Through the ensemble approach, the framework effectively combines these distinct market dynamics. Empirical results demonstrate that the proposed model significantly outperforms recent state‐of‐the‐art benchmarks. Ablation studies and robustness tests demonstrate that integrating spatial and temporal information substantially reduces forecasting errors and confers strong robustness, while highlighting the criticalness of the energy–carbon nexus, offering critical insights for market design and risk management.
Predicting current and near-term macroeconomic developments using linear indicator models and factor models, together with Economic Tendency Survey data, is standard nowcasting practice. In the current article, it is investigated whether machine learning (ML) methods, when used together with a limited set of tendency survey confidence indicators, can improve the forecasts of Swedish quarterly GDP growth compared with linear indicator models and factor models. The results indicate that ML methods generally perform relatively well. In particular, gradient-boosted regression trees, random forests, and multilayer perceptron models are identified as some of the best performing models. The results indicate that ML models can be fruitfully applied in macroeconomic forecasting without employing vast amounts of data. One factor contributing to this could be the ability of ML methods to capture nonlinearities. Results also indicate that, when implementing ML models, care should be taken in determining how often central model parameters should be tuned and estimated.
Evaluating credit risk is crucial for financial institutions because it determines whether loans should be granted to applicants, impacting potential returns and losses for the institution. However, modeling the heterogeneous information of applicants while simultaneously balancing institutional gains and losses faced by the institution is a complex challenge. To address the issues, this study proposes a credit risk evaluation framework with heterogeneous information: combining three-way decision (TWD) theory and graph sample and aggregate (GraphSAGE) learning. This framework (i) integrates a fuzzy similarity relation based on fuzzy theory, constructing a GraphSAGE model for graph-structured data, (ii) combines the results of GraphSAGE with the TWD theory to divide loan applications into three regions: positive, negative, and boundary, and (iii) adds additional information for instances of the boundary for final evaluations, providing new perspectives for research in credit risk evaluation. The experiment results demonstrate the excellent effectiveness of the method in credit risk evaluation, coping well with data imbalances while balancing losses and gains.
An Internet of Things (IoT)-based intelligent energy management system enables monitoring and control of renewable energy generation. It lowers waste and increases energy efficiency via the use of sophisticated forecasting and load optimization techniques. However, challenges such as data inconsistency processing and integration of diverse energy sources still need to be addressed. In this paper, an optimized IoT-based intelligent energy management system for enhanced renewable generation through advanced forecasting and load strategies (IOT-IEMS-RGFLS-GPTPINN) is proposed. Firstly, the input data is gathered from the National Solar Radiation Database. Then the input data is preprocessed using an adaptive higher-order singular value decomposition clutter filter (AHOSVDC) for normalization. The preprocessed data is then fed into the prediction segment by using generative pretrained physics-informed neural networks (GPT-PINN) to predict the accurately forecast renewable energy generation and load demand. The battlefield optimization algorithm (BFOA) is used for enhancing GPT-PINN parameters. The proposed IOT-IEMS-RGFLS-GPTPINN technique is executed in Python. The proposed method's performance was evaluated using performance indicators such as mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE). The IOT-IEMS-RGFLS-GPTPINN achieves an MAE of 0.011%, outperforming the other existing methods SAT-IRES-IOT for 0.052%, RA-DSEM-PSO for 0.034%, and EDSM-SIMS-LEO for 0.063%. This includes existing methods such as Enhanced demand-side management for solar-based isolated microgrid systems: Load prioritization and energy optimization (EDSM-SIMS-LEO), Recent advancement in demand-side energy management systems for optimal energy utilization (RA-DSEM-PSO), and Smart agriculture technology: An integrated framework of renewable energy resources, IoT-based energy management, and precision robotics (SAT-IRES-IOT).
We develop an efficient algorithm for computing lead-lag signature features derived from rough path theory and evaluate its application to financial return prediction. The path signature provides a universal nonlinear feature map for sequential data, but na & iuml;ve computation is intractable as truncation depth grows. We exploit the algebraic structure of Chen's identity to derive incremental updates for depth-2 signatures, achieving substantially lower computational complexity, and prove that the resulting algorithm is numerically stable. The lead-lag transformation embeds univariate series into a two-dimensional path, recovering realized quadratic variation as the L & eacute;vy area-a quantity lost in one-dimensional signatures. In an empirical evaluation on a panel of financial assets spanning equities, fixed income, currencies, commodities, and cryptocurrency, signature features yield their largest gains over technical indicators on long-duration Treasuries, where bond convexity creates path-dependent return dynamics; they offer little incremental advantage on equities, where simple momentum features perform best. We further document a divergence between statistical predictability and economic profitability: a signature-based gold strategy attains a high backtested Sharpe ratio despite a negative unconditional correlation between forecasts and realized returns.
The European Union Emissions Trading System has emerged as a cornerstone of climate policy, with carbon allowance prices exhibiting complex dynamics influenced by policy reforms, market shocks, and structural transitions. This study examines European Union Allowances (EUA) spot-price dynamics through four canonical processes and their regime-switching hidden Markov model (HMM) extensions, estimated using rolling windows to capture time-varying parameters. Our framework integrates in-sample fitting, multi-horizon forecasting, density calibration, and hedging performance evaluation, revealing that regime switching systematically enhances model flexibility and accuracy, particularly for jump-diffusion specifications. By incorporating maturity-matched risk-free rates and statistical tests, we demonstrate that regime-switching models better capture state-dependent behaviors and provide more reliable derivatives pricing and risk management insights. These findings offer practical value for market participants and policymakers in navigating carbon market risks and designing effective hedging strategies.
This paper proposes a convolutional neural network (CNN) model that utilizes EUA chart images to predict carbon price trends. An imaging approach is adopted to convert EUA price and trading volume data into pixel images across four different time horizons as model inputs, enabling predictions for both the next-day price direction and the n-day cumulative trend. Results demonstrate that the image-based CNN model achieves superior performance across various prediction metrics and time horizons, outperforming all traditional machine learning models reliant on time-series data. Furthermore, our forecasting approach exhibits robustness within China's carbon market. This methodology provides carbon market participants with an effective predictive tool, contributing to the market's healthy operation.
Predicting carbon allowance prices has grown more crucial in relation to carbon market regulation, financial strategy, and environmental policy development. This study examines a hybrid forecasting system that combines deep learning with ensemble machine learning models to forecast the price fluctuations of EU Emissions Allowance (EUAs) within the European Union Emissions Trading System (EU ETS). By leveraging a dataset that includes past EUA prices alongside macroeconomic factors like exchange rates, stock indices, natural gas, and crude oil prices, we evaluate the forecasting capabilities of long short-term memory (LSTM) neural networks, random forest (RF), and extreme gradient boosting (XGBoost) models. These models are assessed using commonly recognized metrics: root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). The findings suggest that LightGBM and XGBoost outshine Random Forest and LSTM in performance, with XGBoost emerging as the best predictor model. XGBoost specializes in capturing complex connections within structured data by managing nonlinear connections and interactions. Considering that carbon markets operate in phases, with each phase bringing its own set of reforms, this study offers novel results. In an integrated market prediction of evolving asset class series, LSTM models that are considered more adept at handling sequential data like time series might not yield superior forecasting performance. Our results emphasize the capability of ensemble learning approach forecasting systems to improve prediction accuracy in emissions trading markets, providing important information for policymakers, financial analysts, and sustainability strategists. The hybrid method discussed here demonstrates how AI-based analytics can enhance more resilient and data-informed environmental decision-making.
Kernel-based extreme learning machines (KBELMs) have demonstrated great potential in predicting bankruptcy because of their rapid learning capability and the ability to model nonlinear financial relationships. However, the predictive performance of KBELM is very sensitive to the choice of hyperparameters, and the traditional tuning strategies tend to have the shortcomings of early convergence and insufficient search space exploration. To overcome these problems, this study proposes a quantum-inspired chimp optimization algorithm (QICHOA) for effective hyperparameter optimization of KBELM, and the resultant hybrid model is called KBELM-QICHOA. The proposed optimizer is an improvement of the classical chimp optimization algorithm by adding quantum-inspired mechanisms for better global search capability and balancing exploration and exploitation. The performance of KBELM-QICHOA is evaluated using two real-world bankruptcy datasets, that is, Wieslaw dataset and Japanese bankruptcy dataset, under a nested cross-validation framework. The proposed model is compared with five benchmark approaches, which are conventional KBELM, KBELM-HFDO, KBELM-HAOA, KBELM-RCGWO, and KBELM-IPBBO. Experimental results show that KBELM-QICHOA is significantly better than competing models in terms of prediction accuracy, robustness and stability in terms of RMSE, Nash-Sutcliffe efficiency (NSEF), and bias. The results show that combining quantum-inspired optimization with kernel-based learning has a significant impact on improving the performance of bankruptcy prediction. The proposed KBELM-QICHOA framework therefore constitutes a reliable and economically meaningful early warning tool for the financial risk assessment and decision support.
Time series forecasting is a fundamental task in scientific and engineering disciplines. Quantile regression (QR) has gained substantial popularity due to its flexibility in modeling conditional distributions without stringent parametric assumptions. However, traditional QR approaches often overlook the geometric dependencies and structural relationships among predictors, leading to suboptimal performance in complex forecasting scenarios. To address this gap, this paper proposes a novel framework that integrates graphical regularization with sparse quantile regression, enhanced by an -greedy reinforcement learning (RL) strategy for efficient parameter tuning. Our model incorporates a graph Laplacian matrix to preserve spatial structures among predictors while maintaining the robustness of QR. The resulting optimization problem is solved efficiently using the Limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) algorithm. Empirical studies on real-world electricity market datasets (Belgian and American markets) demonstrate that our approach, designated as RL-RG-SCAD(0.5), significantly outperforms state-of-the-art statistical and deep learning models. These improvements are rigorously validated by Diebold-Mariano tests, Giacomini-White tests, and Hansen's superior predictive ability test. The framework offers enhanced accuracy and interpretability by effectively leveraging predictor structures.
This study aims to explain the transmission mechanism of the COVID-19 pandemic through short-term forecasts based on the logistic smooth transition (LST) model. The model simulation was applied to data from the first seven waves in Japan, as well as the first waves of the other G7 countries. The model provided forecasts with significant precision, with an average of 59 days ahead of the actual number of COVID-19 cases. The model effectively estimated transition speed and identified critical thresholds for each pandemic wave, facilitating early predictions of peak timings. This framework not only enhances our understanding of COVID-19 transmission dynamics but also offers a replicable method for managing future pandemics.
Volatility, as a measure of uncertainty, plays a crucial role in various financial activities such as risk management. Existing methods, including GARCH models and neural network (NN) approaches have their limitations. In this paper, we introduce a novel "decomposed volatility modeling" (DVM) framework, which decomposes the raw volatility signals into two distinct components: the GARCH component and the GARCH Remainder component, which are independently modeled using GARCH and NN techniques. The GARCH component preserves stylized facts signals without distortion by NN, ensuring their retention in the forecast. Simultaneously, the GARCH Remainder, whose stationarity is improved by filtering out the stylized facts signals from the raw volatility series, is modeled using NN. DVM leverages the strengths of both GARCH and NN, which enhances predictive performance. We validate the proposed framework using a comprehensive dataset comprising six financial assets, covering stock indices, cryptocurrencies, and macroeconomic data. The experimental results substantiate that DVM significantly enhances the predictive performance compared with using GARCH or NN model alone, highlighting the efficacy of disentangling the stylized facts signals from the raw volatility series.
The vast application of optimization algorithms has been observed in the field of machine learning for improving the accuracy of the model. A typical setback observed in optimization algorithms is the issue of premature convergence which might occur due to the absence of extraordinary exploration of the complex and continuous search space. Therefore, the present study attempts to introduce an imperative improvement in conventional horse herd optimization (HHO) which significantly boosts the optimizing capability by resolving the issue of premature convergence. This improved HHO (iHHO) is the integration of conventional HHO with the hunting and encircling characteristics of the conventional grey wolf optimization (GWO) algorithm. The performance of the proposed iHHO algorithm has been evaluated on standard benchmark functions and compared with conventional HHO, GWO, and other variants of basic variants of HHO. The statistical significance of the results is verified using the Friedman ranking and Wilcoxon pairwise tests, while computational analysis is demonstrated using Big-O analysis. Later, the practical applicability of the proposed iHHO algorithm has been analyzed by optimizing the hyperparameters of random forest (RF) for developing the effective electric load forecasting model. The error matrices evaluation, stability test, and the Diebold-Mariano test have been performed to highlight the superiority of the iHHO optimized RF model. Furthermore, the performance validation has been extended by comparing the present state of art with the other recently developed models.
We propose PRISM-Now, a novel ensemble forecasting system for near-term GDP projection. Recognizing that relevant economic information evolves over time, we treat forecasts from multiple base models as draws from a mixture distribution of "good" and "bad" estimates, whose composition changes continuously and cannot be identified ex ante. To improve forecasting accuracy, PRISM-Now adaptively selects an aggregation quantile using contemporaneous ensemble distributional information, including changes in central tendency, dispersion, and skewness. Empirical results show that PRISM-Now outperforms alternative ensemble methods, including simple averaging and approaches that rely on backward-looking information. Using Korean GDP data, we further find that conventional models perform relatively well for nowcasting () when near-complete data are available, while big data and machine learning models exhibit stronger performance for one-quarter-ahead forecasts () in the absence of realized information. Models incorporating text and sentiment data are particularly effective during the COVID-19 period. Overall, these findings highlight the value of dynamic ensembling in adapting to rapidly changing economic conditions.
This paper proposes the mixture of hidden Markov factor analyzers (MHMFA), a unified framework for jointly forecasting value-at-risk (VaR) and expected shortfall (ES) in digital asset portfolios. The model integrates regime-switching dynamics via a hidden Markov chain, a latent factor structure capturing systematic co-movements, and regime-specific Gaussian mixtures to flexibly accommodate non-Gaussian features in both common and idiosyncratic components. Parameters are estimated using an expectation-maximization algorithm, and joint VaR-ES forecasts are generated through Monte Carlo simulation under a probabilistic soft-assignment scheme. In a large-scale out-of-sample evaluation on a six-cryptocurrency portfolio, the MHMFA consistently outperforms all competing models across multiple portfolio strategies and confidence levels according to the Patton-Ziegel-Chen joint loss criterion. The model achieves strong Basel III compliance at and generates a time-varying ES/VaR ratio that adapts to market conditions, reflecting increased tail risk during turbulent periods. From an economic perspective, the risk-averse strategy based on MHMFA delivers robust performance after transaction costs, highlighting the practical relevance of the proposed approach for risk management and portfolio allocation.
Industry experts as well as academic scholars have directed substantial attention toward researching stock market volatility. Most of the past research has focused on using singular features such as closing prices, opening prices, or news stories to predict stock movement. In recent times, there has been a growing interest in using multimodal graph neural networks which can analyze a variety of features. Most of these methods focus on node aggregation to extract relevant features from related stocks to further improve model accuracy. The current state-of-the-art model-ML-GAT: Multilevel Graph Attention Model-constructs a graph network between the stocks using Wikidata relations. It uses multiple layers of graph attention to aggregate features such as historical price features and current news. However, the high number of intercompany relationships in ML-GAT may include irrelevant and noisy edges. It also uses individual attention coefficients for each layer, leading to inefficient utilization of computational resources. To overcome these challenges, a Sparse Graph Attention Network for Stock Prediction (SGAT-SP) is proposed in this paper. SGAT-SP uses a single set of attention coefficients to reduce training and inference time. It assigns a binary mask to every edge which represents whether they will be used for node aggregation to reduce noisy edges. The proposed approach achieves an average accuracy score of 0.83, a slight improvement over ML-GAT, which has an accuracy score of 0.827. Additionally, it significantly reduces inference time by 85%, resulting in faster results and decreased computational expenses.
