Journal of Forecasting最新文献

This paper analyzes the inflation forecast errors over the period 2021Q1–2022Q3 using forecasts of core and headline inflation from the International Monetary Fund World Economic Outlook for a large group of advanced and emerging market economies. The findings reveal evidence of forecast bias that worsened initially and then subsided towards the end of the sample. There is also evidence of forecast oversmoothing, indicating rigidity in forecast revision in the face of incoming information. Focusing on core inflation forecast errors in 2021, four factors provide a potential ex post explanation: a stronger-than-anticipated demand recovery; demand-induced pressures on supply chains; the demand shift from services to goods at the onset of the pandemic; and labor market tightness. Ex ante, we find that the size of the COVID-19 fiscal stimulus packages announced by different governments in 2020 correlates positively with core inflation forecast errors in advanced economies. This result hints at potential forecast inefficiency, but we caution that it hinges on the outcomes of a few, albeit large, economies.

Wind power has emerged as a successful component within power systems. The ability to reliably and accurately forecast wind speed is of great importance in maintaining the security and stability of the power grid. However, the significance of explaining prediction models has often been overlooked by researchers. To address this gap, this study introduces a novel approach to wind speed forecasting that incorporates a significant decomposition method, attention-based machine learning, and local explanation techniques. The proposed model utilizes grid search variational mode decomposition to decompose the wind speed sequence into different modes while employing gate recurrent unit with an attention mechanism to achieve superior forecasting performance. Experimental evaluations conducted on eight real-world wind speed datasets demonstrate that the proposed approach outperforms other popular models across multiple performance criteria. In two specific experiments, the proposed approach achieved a minimal mean absolute percentage error of 2.74% and 1.70%, respectively. Furthermore, local interpretable model-agnostic explanations (LIME) were employed to assess the influence of factors, highlighting whether they positively or negatively affected the predicted values.

We assess the relationship between model size and complexity in the time-varying parameter vector autoregression (VAR) framework via thorough predictive exercises for the euro area, the United Kingdom, and the United States. It turns out that sophisticated dynamics through drifting coefficients are important in small data sets, while simpler models tend to perform better in sizeable data sets. To combine the best of both worlds, novel shrinkage priors help to mitigate the curse of dimensionality, resulting in competitive forecasts for all scenarios considered. Furthermore, we discuss dynamic model selection to improve upon the best performing individual model for each point in time.

We analyze the out-of-sample predictive power of sentiment for the realized volatility of agricultural commodity price returns. We use high-frequency intra-day data covering the period from 2009 to 2020 to estimate realized volatility. Our baseline forecasting model is a heterogeneous autoregressive (HAR) model, which we extend to include sentiment. We further enhance this model by incorporating various key realized moments such as leverage, realized skewness, realized kurtosis, realized upside (“good”) volatility, realized downside (“bad”) volatility, realized jumps, realized upside tail risk, and realized downside tail risk. In order to setup a forecasting model, we use (i) forward and backward stepwise predictor selection and (ii) a model-based averaging algorithm. The forecasting models constructed through these algorithms outperform both the baseline HAR-RV model and the HAR-RV-sentiment model. We conclude that, for the agricultural commodities studied in our research, realized moments play a more significant role in forecasting realized volatility compared to sentiment.

The accurate forecasting of agricultural futures prices is critical for ensuring national food security. Therefore, this study proposes a text-based deep learning forecasting model. This model first uses the ChineseBERT + a text convolution neural network to classify Weibo text and obtain a raw sentiment index. Then, complete ensemble empirical mode decomposition with adaptive noise, variational mode decomposition, correlation coefficient, and sample entropy are combined to decompose and reconstruct the raw sentiment index and obtain a denoised sentiment index. Subsequently, the neural basis expansion analysis with exogenous variables is improved by designing a weight coefficient and Optuna is used to optimize the designed weight coefficient and the hyperparameters. Finally, the SHapley Additive exPlanations value is used to increase the interpretability of prediction results. Corn futures prices for the Dalian Exchange are used in forecasting to validate the accuracy and stability of the proposed model. Experimental results show that the proposed denoising sentiment index contributes more to the improvement of predictive model performance than the raw sentiment index. The proposed text-based deep predictive model demonstrates strong predictive ability for prediction horizons of 30 and 60 days. SHapley Additive exPlanations value analysis shows that the three features with greater effects on corn futures prices are as follows: “Corn Spot Price of Zhengzhou market,” “CBOT_corn_futures_price,” and “Pork futures price.”

The reformed Basel framework has left value at risk (VaR) as a basic tool of validating risk models. Within this framework, VaR independence tests have been regarded as critical to ensuring stability during periods of financial turmoil. However, until now, there is no consent among researchers regarding the choice of the appropriate test. The available procedures are either inaccurate in finite samples or need to rely on Monte Carlo simulations. To remedy these problems, we propose a new method for testing VaR models, based on the distribution of the number of runs. It outperforms the existing methods in two main aspects: First, it is exact in finite samples and thus allows for perfect control over the Type 1 error; second, its distribution is available in a closed form, so it does not require simulations before implementation. We show that it is the most adequate current procedure for testing low-level VaR series, which corresponds to today's regulatory standards.

Electronic and digital trading models have made stock trading more accessible and convenient, leading to exponential growth in trading data. With a wealth of trading data available, researchers have found opportunities to extract valuable insights by uncovering patterns in stock price movements and market dynamics. Deep learning models are increasingly being employed for stock price prediction. While neural networks offer superior computational capabilities compared with traditional statistical methods, their results often lack interpretability, limiting their utility in explaining stock price volatility and investment behavior. To address this challenge, we propose a causality-based method that incorporates a multivariate approach, integrating news event attention sequences and sentiment index sequences. The goal is to capture the intricate and multifaceted relationships among news events, media sentiment, and stock prices. We illustrate the application of this proposed approach using a Global Database of Events, Language, and Tone global event database, demonstrating its benefits through the analysis of attention sequences and media sentiment index sequences for news events across various categories. This research not only identifies promising directions for further exploration but also offers insights with implications for informed investment decisions.

In recent years, the semiparametric methods for the joint estimation and prediction of value at risk (VaR) and expected shortfall (ES) have triggered great interests and attention. Compared to existing literature which usually incorporates realized volatility (RV) into the dynamic semiparametric risk models, this paper considers three more robust proxies (medRV, BPV, and RK) of intraday volatility in the models to verify whether high-frequency information can improve the joint prediction ability of risk measures. To strengthen the persuasion of conclusions, four international stock indices (S&P500, Nikkei225, GDAXI, and DJIA) are applied to these models to estimate and forecast VaR and ES at different probability levels (1%, 2.5%, 5%, and 10%). Then, the predicted VaR and ES are backtested by several methods individually, and the popular score function FZ0 and MCS test are used to compare the effects of jointly predicting risk measures. Our results confirm that these semiparametric models containing intraday information outperform the benchmark models for four stocks and various probability levels, and medRV is the best volatility measure in improving the effects of models.

This study presents a novel and fully probabilistic approach for combining model-based forecasts with surveys or other judgmental forecasts. In our method, survey forecasts are integrated as penalty terms for the model parameters, facilitating a probabilistic exploration of additional insights obtained from surveys. We apply this approach to estimate a growth-at-risk model for real GDP growth in the United States. The results reveal that this additional shrinkage significantly improves prediction performance, with the information from surveys even exerting an influence on the lower tails of the distribution.

New fashion product sales forecasting is a challenging problem that involves many business dynamics and cannot be solved by classical forecasting approaches. In this paper, we investigate the effectiveness of systematically probing exogenous knowledge in the form of Google Trends time series and combining it with multi-modal information related to a brand-new fashion item, in order to effectively forecast its sales despite the lack of past data. In particular, we propose a neural network-based approach, where an encoder learns a representation of the exogenous time series, while the decoder forecasts the sales based on the Google Trends encoding and the available visual and metadata information. Our model works in a non-autoregressive manner, avoiding the compounding effect of large first-step errors. As a second contribution, we present VISUELLE, a publicly available dataset for the task of new fashion product sales forecasting, containing multimodal information for 5,577 real, new products sold between 2016 and 2019 from Nunalie, an Italian fast-fashion company. The dataset is equipped with images of products, metadata, related sales, and associated Google Trends. We use VISUELLE to compare our approach against state-of-the-art alternatives and several baselines, showing that our neural network-based approach is the most accurate in terms of both percentage and absolute error. It is worth noting that the addition of exogenous knowledge boosts the forecasting accuracy by 1.5% in terms of Weighted Absolute Percentage Error (WAPE), revealing the importance of exploiting informative external information. The code and dataset are both available online (at https://github.com/HumaticsLAB/GTM-Transformer).