International Journal of Forecasting最新文献

This paper introduces a novel stochastic process with signed integer values. Its autoregressive dynamics effectively captures persistence in conditional moments, rendering it a valuable feature for forecasting applications. The increments follow a Generalized Poisson distribution, capable of accommodating over- and under-dispersion in the conditional distribution, thereby extending standard Poisson difference models. We derive key properties of the process, including stationarity conditions, the stationary distribution, and conditional and unconditional moments, which prove essential for accurate forecasting. We provide a Bayesian inference framework with an efficient posterior approximation based on Markov Chain Monte Carlo. This approach seamlessly incorporates inherent parameter uncertainty into predictive distributions. The effectiveness of the proposed model is demonstrated through applications to benchmark datasets on car accidents and an original dataset on cyber threats, highlighting its superior fitting and forecasting capabilities compared to standard Poisson models.

Emergency department (ED) crowding is a significant threat to patient safety and it has been repeatedly associated with increased mortality. Forecasting future service demand has the potential to improve patient outcomes. Despite active research on the subject, proposed forecasting models have become outdated, due to the quick influx of advanced machine learning models and because the amount of multivariable input data has been limited. In this study, we document the performance of a set of advanced machine learning models in forecasting ED occupancy 24 h ahead. We use electronic health record data from a large, combined ED with an extensive set of explanatory variables, including the availability of beds in catchment area hospitals, traffic data from local observation stations, weather variables, and more. We show that DeepAR, N-BEATS, TFT, and LightGBM all outperform traditional benchmarks, with up to 15% improvement. The inclusion of the explanatory variables enhances the performance of TFT and DeepAR but fails to significantly improve the performance of LightGBM. To the best of our knowledge, this is the first study to extensively document the superiority of machine learning over statistical benchmarks in the context of ED forecasting.

In this paper, we evaluate the marginal predictive content of a variety of new business conditions (BC) predictors as well as nine economic uncertainty indexes (EUIs) constructed using these predictors. Our predictors are defined as observable variables and latent factors extracted from a high-dimensional macroeconomic dataset, and our EUIs are functions of predictive errors from models that incorporate these predictors. Estimation of the predictors is based on a number of extant and novel machine learning methods that combine dimension reduction, variable selection, and shrinkage. When predicting 14 monthly U.S. economic series selected from eight different groups of economic variables, our new indexes and predictors are shown to result in significant improvements in forecast accuracy relative to predictions made using benchmark models. In particular, inclusion of either BC predictors or EUIs often yields forecast accuracy improvements, while even greater predictive gains accrue when including both BC predictors and EUIs when forecasting real economic activity-type variables at shorter forecast horizons.

Alternative data are timely but messy. They can provide policymakers with real-time information, but their use is constrained by the complexity of their relationship with official statistics. Data from credit card transactions, search engines, or traffic have been made available since only recently, which makes it more difficult to precisely gauge their relationship with national accounts. This paper aims at solving this problem by compensating their short history with their large country coverage. It introduces a heterogeneous panel model approach where a neural network learns the relationship between Google Trends data and GDP growth from data pooled from 46 countries. The resulting “OECD Weekly Tracker” yields real-time estimates of weekly GDP, which are proven to be accurate using forecast simulations. It is a valuable compass for policymaking in turbulent waters.

The expectations-augmented Phillips curve (PC) is a cornerstone of many macroeconomic models. We consider the extent to which professional forecasters’ inflation and unemployment rate forecasts are ‘theory consistent’, and find much heterogeneity. Perceptions about the responsiveness of inflation to the unemployment rate are shown to depend on whether the respondent was active earlier or later during the period 1981–2019, and on whether the respondent happened to forecast at times of tight labour markets.

Theory consistency is related to more accurate forecasts at the shortest horizon but not significantly so at longer horizons. At longer horizons PC-model heterogeneity accounts for the lion’s share of the observed disagreement in reported inflation forecasts.

Macroeconomic forecasting has recently started embracing techniques that can deal with large-scale datasets and series with unequal release periods. Mixed-data sampling (MIDAS) and dynamic factor models (DFMs) are the two main state-of-the-art approaches to modeling series with non-homogeneous frequencies. We introduce a new framework, called the multi-frequency echo state network (MFESN), based on a relatively novel machine learning paradigm called reservoir computing. Echo state networks (ESNs) are recurrent neural networks formulated as nonlinear state-space systems with random state coefficients where only the observation map is subject to estimation. MFESNs are considerably more efficient than DFMs and can incorporate many series, as opposed to MIDAS models, which are prone to the curse of dimensionality. All methods are compared in extensive multistep forecasting exercises targeting U.S. GDP growth. We find that our MFESN models achieve superior or comparable performance over MIDAS and DFMs at a much lower computational cost.

Market integration connects multiple autarkic electricity markets and facilitates the flow of power across areas. More often than not, market integration can increase social welfare for the whole power system with a clear spatial dependence structure among area electricity prices, which motivates a new perspective on electricity price forecasting. In this paper, we construct a model to forecast the day-ahead electricity prices of Nord Pool with spatial dependence. First of all, we convert the electricity prices into graph data. Then, we propose an STGNN (Spatial-Temporal Graph Neural Network) model to exploit spatial and temporal features. In particular, the STGNN model can accurately forecast electricity prices for multiple areas, where the adjacency matrix representing the spatial dependence structure is pre-captured by the R-vine (regular vine) copula. Our results show that the spatial dependence structure described by the R-vine copula can perfectly reflect the physical characteristics of the electricity system; moreover, the forecasting performance of the proposed STGNN model is significantly better than the existing models in terms of overall accuracy and hourly accuracy within a day.