Journal of Forecasting最新文献

Influenced by various complex factors, the price series of agricultural futures exhibit nonstationarity. Existing research often presumes that the relationship between inputs and outputs remains stable throughout the training process. This assumption makes it challenging to dynamically adjust the weights of various models based on data characteristics. Furthermore, existing studies focus only on modeling variable dependencies, overlooking the impact of variable independence on model robustness. Therefore, this paper proposes a two-stream ensemble forecasting model that integrates a dynamic sentiment index. Initially, ChineseBERT and textCNN are employed to classify the sentiment of news texts, calculating the sentiment scores. Subsequently, weight factors are designed based on daily price fluctuations to adjust these sentiment scores, ensuring they accurately reflect the impact of news sentiment on market prices. In the model construction phase, multivariate time series data are input into two distinct models: one model is dedicated to capturing temporal dependencies, while the other focuses on capturing intervariable dependencies, thereby providing diverse yet complementary predictive insights. An online convex optimization approach is then utilized to learn the optimal combination weights. During the testing phase, reinforcement learning is applied to dynamically adjust the prediction weights of these two models. The effectiveness of the proposed methods is validated using soybean and corn futures prices. Experimental results demonstrate that the proposed two-stage sentiment index (TPSI) exhibits strong predictive capability for agricultural futures prices, achieving high accuracy in short-term and medium-term price forecasts.

This study examines the effectiveness of different predictors to forecast volatility of E7 emerging markets. First, we employ the economic uncertainty factors information to find out the economic impact on stock market volatility. Second, we used the geopolitical uncertainty factor information to explore the influence of geopolitical events on stock market volatility. Third, we investigate the climate risk factors to forecast the stock market volatility. Fourth, we examine the energy market–related uncertainty factors' impact on equity market volatility (EMV). The out-of-sample results show that among all predictors, economic policy uncertainty (EPU), EMV, geopolitical risk (GPR), and the environmental social and governance (ESG) index have the better forecasting ability to predict stock market volatility. Additionally, we find evidence during different business conditions, recession and expansion, and the most recent COVID-19 pandemic. The results are identical during recession periods and COVID-19 pandemic. Notably, these indices proved their superior predictive performance for volatility estimation. Finally, to ensure the robustness of our findings, we use different forecasting window method. Considering a range of uncertainty factors allows for a more comprehensive understanding of market fluctuations. These results suggest that policymakers and investors should consider the volatility dynamics of uncertainty factors in financial decision-making and policy realms.

With Covid-19 and Russian invasion of Ukraine triggering vast uncertainty, non-traditional (potentially big) data sources could provide timely information on the state of the economy. We investigate the usefulness of big data during these events in a bottom-up nowcasting exercise for Polish food inflation. We show that information embedded in roughly 250 million web-scraped prices makes it possible to closely track official inflation and provides the most accurate nowcasts in times of elevated uncertainty and volatility, while competing model-based frameworks skew nowcasts towards historical patterns, which leads to much lower accuracy. In turn, during normal times, time series models with web-scraped prices provide slightly more accurate nowcasts but the difference in predictive quality is negligible in comparison to nowcasts based purely on online prices. We also extensively experiment on how to optimally aggregate daily online prices. For practitioners we formulate four guidelines. First, let big data speak, especially during uncertain times. Second, model-based frameworks with online prices are not necessary to obtain precise nowcasts of price developments, even on long samples. Third, simplest aggregation methods of individual product prices lead to the most accurate nowcasts. Fourth, for nowcasting quality the gain from observing prices daily is marginal compared to weekly frequency of data collection.

The rapid expansion of Internet infrastructure and artificial intelligence (AI) has significantly advanced intelligent transportation systems (ITS), which are considered as essential for automating traffic monitoring and management in smart cities. Among ITS applications, traffic flow and density prediction are considered as important problem for optimizing transportation planning and reducing congestion. In recent years, deep learning models, particularly recurrent neural networks (RNNs) and graph neural networks (GNNs), have been widely utilized for traffic forecasting. These models can support to effectively capture temporal and spatial dependencies in traffic data, as a result enabling more accurate forecasting. Despite advancements, recently proposed RNN-GNN-based forecasting models still face challenges related to the capability of preserving rich structural and topological features from traffic networks. The complex spatial dependencies inherent in road connections and vehicle movement patterns are often underrepresented; therefore, limiting the forecasting accuracy. To address these limitations, in this paper, we propose SGL4TF, a structure-enhanced graph learning model that integrates graph convolutional networks (GCN) with a sequence-to-sequence (seq2seq) framework. This architecture enhances the ability to jointly model spatial relationships and long-term temporal dependencies, hence can lead to more precise traffic predictions. Our approach introduces a deeper graph-structural learning mechanism using nonlinear transformations within GNN layers, which can effectively assist to improve structural feature extraction while mitigating over-smoothing issues. The seq2seq component further refines temporal correlations, enabling long-term traffic state predictions. Extensive experiments on real-world datasets demonstrate our proposed SGL4TF model's superior performance over state-of-the-art traffic forecasting techniques.

In this work, we develop time series regression models for long-memory count processes based on the generalized linear Gegenbauer autoregressive moving average (GLGARMA) framework. We present a comprehensive Bayesian formulation that addresses both in-sample and out-of-sample forecasting within a broad class of generalized count time series regression models. The GLGARMA framework supports various count distributions, including Poisson, negative binomial, generalized Poisson, and double Poisson distributions, offering the flexibility to capture key empirical characteristics such as underdispersion, equidispersion, and overdispersion in the data. We connect the counting process to a time series regression framework through a link function, which is associated with a stochastic linear predictor incorporating the family of long-memory GARMA models. This linear predictor is central to the model's formulation, requiring careful specification of both the GLGARMA Bayesian likelihood and the resulting posterior distribution. To model the stochastic error terms driving the linear predictor, we explore two approaches: parameter-driven and observation-driven frameworks. For model estimation, we adopt a Bayesian approach under both frameworks, leveraging advanced sampling techniques, specifically the Riemann manifold Markov chain Monte Carlo (MCMC) methods implemented via R-Stan. To demonstrate the practical utility of our models, we conduct an empirical study of open interest dynamics in US Treasury Bond Futures. Our Bayesian models are used to forecast speculative pressure, a crucial concept for understanding market behavior and agent actions. The analysis includes 136 distinct time series from the US Commodity Futures Trading Commission (CFTC), encompassing futures-only and futures-and-options data across four US government-issued fixed-income securities. Our findings indicate that the proposed Bayesian GLGARMA models outperform existing state-of-the-art models in forecasting open interest and speculative pressure. These improvements in forecast accuracy directly enhance portfolio performance, underscoring the practical value of our approach for bond futures portfolio construction. This work advances both the methodology for modeling long-memory count processes and its application in financial econometrics, particularly in improving the forecasting of speculative pressure and its impact on investment strategies.

Current studies have designed many credit scoring models with high performance, but they are often weak in interpretability with obvious “black box” features. This makes them difficult to meet the requirements of the regulators about the model's interpretability. This paper presents a novel credit scoring model as the IWSL model, which is data feature driven with interpretable features. The IWSL model first calculates the representative eigenvectors of default and nondefault samples according to their spatial distribution characteristics, as well as the eigenvector located in the middle of these two types of eigenvectors in the sample space. It then calculates the weighted distance between each sample and each eigenvector to divide the training dataset into three subsets and constructs sublogistic models separately. In the absence of prior information about the optimal weight setting of each attribute, the swarm intelligence algorithm is applied to back-optimize the weights according to the model's generalization ability in the validation stage. The empirical results show that the IWSL model outperforms 12 leading credit scoring models on three public consumer credit scoring datasets with statistical significance. Model component validity testing confirms the effectiveness of the IWSL model's core settings, while sensitivity analysis validates its ability to maintain robust results.

In financial distress prediction (FDP), it is very important to ensure the quality of the data for developing effective prediction models. Related studies often apply feature selection to filter out some unrepresentative features from a set of financial ratios, or data re-sampling to re-balance class imbalanced FDP training sets. Although these two types of data pre-processing methods have been demonstrated their effectiveness, they have not often been applied at the same time to develop FDP models. Moreover, the performances of various feature selection algorithms, which can be divided into filter, wrapper, and embedded methods, and data re-sampling algorithms, which can be divided into under-sampling, over-sampling, and hybrid sampling methods, have not been fully investigated in FDP. Therefore, in this study several feature selection and data re-sampling methods, which are employed alone and in combination by different orders are compared. The experimental results based on nine FDP datasets show that executing data re-sampling alone always outperforms executing feature selection alone to develop FDP models, in which hybrid sampling is the better choice. In most cases, better prediction performances can be obtained by performing feature selection first and data re-sampling second. The best combined algorithms are based on the decision tree method for feature selection and Synthetic Minority Over-sampling Technique-Edited Nearest Neighbors (SMOTE-ENN) for hybrid sampling. This combination allows the random forest classifier to produce the highest rate of prediction accuracy. On the other hand, for the Type I error, where crisis cases are misclassified into the non-crisis class, the lowest error rate is produced by executing under-sampling alone using the ClusterCentroids algorithm combined with the random forest classifier.