Journal of Forecasting最新文献

The aim of this paper is to utilize the generalized autoregressive conditional heteroscedasticity–mixed data sampling (GARCH-MIDAS) framework to predict the daily volatility of state-level stock returns in the United States (US), based on the weekly metrics from the corresponding broad economic conditions indexes (ECIs). In light of the importance of a common factor in explaining a large proportion of the total variability in the state-level economic conditions, we first apply a dynamic factor model with stochastic volatility (DFM-SV) to filter out the national factor from the local components of weekly state-level ECIs. We find that both the local and national factors of the ECI generally tend to affect state-level volatility negatively. Furthermore, the GARCH-MIDAS model, supplemented by these predictors, surpasses the benchmark GARCH-MIDAS model with realized volatility (GARCH-MIDAS-RV) in a majority of states. Interestingly, the local factor often assumes a more influential role overall, compared with the national factor. Moreover, when the stochastic volatilities associated with the local and national factors are integrated into the GARCH-MIDAS model, they outperform the GARCH-MIDAS-RV in over 80% of the states. Our findings have important implications for investors and policymakers.

This article studies the estimation of the constrained factor models for high-dimensional time series. The approach is based on the eigenanalysis of a nonnegative definite matrix constructed from the autocovariance matrices. The convergence rate of the estimator for loading matrix and the asymptotic normality of the estimated factor score are explored under regularity conditions set for the proposed model. Our estimation for the constrained factor models can achieve the optimal rate of convergence even in the case of weak factors. The finite sample performance of our approach is examined and compared with the existing methods by Monte Carlo simulations. Our methodology is illustrated and supported by a real data example.

Recently, many researchers have shown keen interest in natural gas price prediction using machine learning and hybrid architectures. Our research forecasts natural gas future prices with different hybrid machine learning models using over a hundred technical indicators. The hybrid deep cross-network model outperformed the single-stage deep cross-network regression and hybrid support vector machine models with 33% and 46% lower mean absolute error and 22% and 1.2 times better directional hit rate during 11 months of turbulent market circumstances due to the Russia–Ukraine crisis. The hybrid deep cross-network model is 14, 5, and 6 times more profitable than the hybrid support vector machine, the benchmark passive buy-and-hold strategy, and the single-stage deep cross-network regression models. The hybrid deep cross-network model is resilient during low- and high-volatility periods. Deep cross-network algorithm technical indicator interactions are more statistically significant than support vector machine polynomial kernel interactions. Energy traders and policymakers can exploit our findings.

In a complex and volatile macroeconomic environment, precious metals, which have the functions of preservation, appreciation, and hedging, play an important role in investment risk management. Therefore, this study adopts the extended GARCH-MIDAS model to investigate the underlying connection between gold price volatility and different uncertain shocks. In this paper, we consider five uncertainty indicators and then decompose them into different states to capture their shock sizes. Next, we introduce uncertainty shocks into the MIDAS structure to test whether they contain relevant and valid information about gold price volatility forecasts. Specifically, parameter significance suggests a positive association between uncertain indicators and gold price volatility, but variability in the influence of their shock sizes on gold price volatility. Out-of-sample results present that the extended model that includes asymmetric shock sizes outperforms other competitive models. Besides, the model that includes large shock sizes exhibits better predictive performance than the model that includes small shocks. Finally, based on the empirical analyses, this paper provides new insights for the gold industry, futures exchanges, government regulators, and investors engaged in futures hedging to achieve risk control and financial stability in response to uncertain shocks.

Between 2001 and 2023, Central Bank forecasts were patently inaccurate. In this paper, we argue that many of such forecast failings were already present during the earlier years of inflation targetry. Central Banks normally adjust monetary policy so that inflation hits the Inflation Target (IT) within two years. Since a central bank must believe its policy stance is appropriate to achieve this goal, its inflation forecast at the two-year horizon should generally be close to target. We examine whether this has held for three main Central Banks, Bank of England, ECB, and Fed. Although over the IT period prior to 2020, both forecasts and outcomes were commendably close to target, we found that this was due to a sizeable forecast underestimate of the effects of policy and inherent resilience to revive inflation after the GFC crisis hit, largely offset by an overestimate of the effect of monetary policy to restore inflation to target during the more normal times. We attribute such latter overestimation to an unwarranted belief in forward-looking, “well anchored”, expectations amongst households and firms, and to a failure to recognize the underlying disinflationary trends, especially in 2010–2019. We outline a novel means for assessing whether these latter trends were primarily demand driven, e.g. secular stagnation, or supply shocks, a labor supply surge. Finally, we examine how forecasts for the uncertainty of outcomes and relative risk (skew) to the central forecast have developed by examining the Bank of England's fan chart, again at the two-year horizon.

Long-term load forecasting (LTLF) has been a fundamental least-cost planning tool for electric utilities. In the past, utilities were monopolies and paid less attention to uncertainty in their LTLF methodologies. Nowadays, such casualness is pricey in competitive markets because utilities need to examine the financial implications of forecast uncertainty for survival. Hence, the aim of this paper is to critique the LTLF research trends with a focus on uncertainty quantification (UQ). For this purpose, we examined 40 LTLF articles published between January 2003 and February 2021. We found that UQ is a nascent area of LTLF research. Our review found two approaches to UQ in LTLF: probabilistic scenario analysis and direct probabilistic methods. The former approach is more helpful to risk analysts but has major caveats in addressing interdependencies of socioeconomic and climate scenarios. We identified very little LTLF research that examines uncertainties associated with climate extremes, distributed generation resources, and demand-side management. Lastly, there is enormous potential for mitigating financial risks by embracing asymmetric cost functions in LTLF research. Future LTLF researchers can work on these identified gaps to help utilities in risk estimation, cost-reliability balancing, and estimation of reserve margin under climate change.

The rise of inflation in 2021 and 2022 surprised many macroeconomists who ignored the earlier surge in money growth because of past instability in the demand for simple-sum monetary aggregates. We find that the demand for more theoretically based Divisia aggregates can be modeled and that these aggregates provide useful information about nominal GDP. Unlike M2 and Divisia-M2, whose velocities do not internalize shifts in liabilities across commercial and shadow banks, the velocities of broader Divisia monetary aggregates are stable and can be empirically modeled through the Covid-19 pandemic. In the long run, these velocities depend on regulation and mutual fund costs that affect the substitutability of money for other financial assets. In the short run, we control for swings in mortgage activity and use vaccination rates and the stringency of government pandemic restrictions to control for the unusual pandemic effects. The velocity of broad Divisia money declines during crises like the Great and COVID Recessions but later rebounds. In these recessions, monetary policy lowered short-term interest rates to zero and engaged in quantitative easing of about $4 trillion. Nevertheless, broad money growth was more robust in the COVID Recession, reflecting a less impaired banking system that promoted rather than hindered deposit creation. Our framework implies that nominal GDP growth and inflation rebounded more quickly from the COVID Recession versus the Great Recession. Our different scenarios for future Divisia money growth and the unwinding of the pandemic have different implications for medium-term nominal GDP growth and inflationary pressures.

The prediction and early warning of soybean futures prices have been even more crucial for the formulation of food-related policies and trade risk management. Amid increasing geopolitical conflicts and uncertainty in trade policies across countries in recent years, there have been significant fluctuations in global soybean futures prices, making it necessary to investigate fluctuations in soybean futures prices, reveal the price determination mechanism, and accurately predict trends in future prices. Therefore, this study proposes a comprehensive and interpretable framework for soybean futures price forecasting. Specifically, this study employs a set of methodologies. Using a snow ablation optimizer (SAO), this study improves the parameters of a time fusion transformer (TFT) model, an advanced interpretable predictive model based on a self-attention mechanism. Besides, it addresses the factors influencing soybean futures prices and constructs effective fusion features through a feature fusion method. To explore volatility trends, the original soybean futures price series are decomposed using variational mode decomposition (VMD). This study also enhances the accuracy of soybean futures price predictions by introducing global geopolitical risk coefficients and trading volumes as predictors. The empirical findings suggest that the VMD-SAO-TFT model enhances prediction accuracy and interpretability, offering implications for decision-makers to achieve accurate predictions and early warning of agricultural futures prices.

Forecasting the realized volatility of oil and gas firms is of interest to investors and practitioners trading on the energy spot and derivative markets. In this paper, we assess whether several machine learning (ML) techniques can offer superior forecasts compared to HAR models for predicting realized volatility at the firm level. Moreover, we investigate whether economically motivated variables and technical indicators contain valuable information for forecasting firm volatility beyond those contained in various volatility factors previously identified in the literature. Our results demonstrate that certain ML techniques provide superior forecasting accuracy compared to the benchmark model. Additionally, we identify variables such as the 1-month treasury bill and the aggregate VIX index as significant drivers of realized firm volatility in the oil and gas industry.

We propose a nonstationary functional time series forecasting method with an application to age-specific mortality rates observed over the years. The method begins by taking the first-order differencing and estimates its long-run covariance function. Through eigendecomposition, we obtain a set of estimated functional principal components and their associated scores for the differenced series. These components allow us to reconstruct the original functional data and compute the residuals. To model the temporal patterns in the residuals, we again perform dynamic functional principal component analysis and extract its estimated principal components and the associated scores for the residuals. As a byproduct, we introduce a geometrically decaying weighted approach to assign higher weights to the most recent data than those from the distant past. Using the Swedish age-specific mortality rates from 1751 to 2022, we demonstrate that the weighted dynamic functional factor model can produce more accurate point and interval forecasts, particularly for male series exhibiting higher volatility.