Journal of Forecasting最新文献

This paper introduces a novel forecasting approach that addresses a significant challenge in applied research: effectively utilizing high-dimensional and mixed-frequency data from multiple sources to explain and predict variables that respond at high frequency. This approach combines a mixed data sampling model and group variable selection methods, resulting in the development of the Group Penalized Reverse Unrestricted Mixed Data Sampling Model (GP-RU-MIDAS). The GP-RU-MIDAS model is designed to achieve various research objectives, including analyzing mixed-frequency data in reverse, estimating high-dimensional parameters, identifying key variables, and analyzing their relative importance and sensitivity. By applying this model to uncover uncertainties in stock market returns, the following notable results emerge: (1) GP-RU-MIDAS improves the selection of relevant variables and enhances forecasting accuracy; (2) various risks impact stock market returns in diverse ways, with effects varying over time and exhibiting continuous trends, phase shifts, or extreme levels; and (3) stock market volatility and the Euro to RMB exchange rate significantly influence stock market returns over different forecasting periods, with a generally positive and dynamic impact. In conclusion, the GP-RU-MIDAS model demonstrates robustness and utility in complex data analysis scenarios, providing insights into the nuanced realm of stock market risk assessment.

In response to the global need for effective management of carbon emissions and alignment with sustainable development goals, predicting carbon trading prices accurately is critical. This study introduces a multifrequency data fusion carbon price prediction model (MFF-CPPM), addressing the nonlinear characteristics of carbon trading prices and inconsistent feature factor frequencies. The MFF-CPPM consists of a feature-extraction frontend, a multifrequency data fusion transformer, and a fusion regression layer, offering a novel methodological approach in forecasting studies. The model's validity was tested in Guangdong, China's largest carbon trading pilot market. The results demonstrated that the MFF-CPPM outperformed baseline models in terms of carbon price-prediction accuracy and trend forecasting. Additional trials conducted in Hubei and Beijing confirmed the model's robustness and generalization capabilities, providing valuable evidence of its effectiveness and reliability across varying market contexts. This study presents a novel predictive model for carbon trading prices, with a unique capability to harness data at differing frequencies. The MFF-CPPM not only enhances forecasting accuracy but also offers an innovative approach to effectively incorporate multifrequency information. This advancement paves the way for flexible forecasting models in any scenario where data arrive at differing frequencies.

We propose a new momentum-determined indicator-switching (N-MDIS) strategy, harnessing the power of machine learning to enhance the accuracy of equity premium prediction. Specifically, we re-examine the regime-dependent feature of univariate predictive regression relative to the benchmark. Furthermore, we investigate the prediction mechanism of the momentum-determined indicator-switching (MDIS) strategy and validate the significance of market regime information for the MDIS. Our findings demonstrate an overwhelmingly superior ex-post forecasting performance compared with the MDIS. More notably, our empirical results substantiate that machine learning greatly aids in momentum indicator selection. The results show that the N-MDIS with machine learning generates more accurate ex-ante equity premium forecasts than both MDIS strategy and N-MDIS strategy with logistic regression, yielding statistically and economically significant results. Moreover, our new approach exhibits robust forecasting performance across a series of robustness tests.

The combination of the conditional autoregressive value-at-risk (CAViaR) process with the Fissler–Ziegel (FZ) loss function generates a recently emerging framework (CAViaR-FZ) for forecasting value-at-risk (VaR) and expected shortfall (ES). However, existing CAViaR-FZ models typically overlook the presence of long-range dependence, a stylized fact of financial time series. This paper proposes a long-memory CAViaR-FZ model using the cross-sectional aggregation (CSA) method. The CSA method is well-recognized for its ability to generate a long-memory process by aggregating an infinite number of short-memory processes cross-sectionally. The proposed CSA-CAViaR-FZ model flexibly captures long-memory dynamics in both VaR and ES and includes the original short-memory CAViaR-FZ model as a special case. Simulation and empirical results demonstrate that the proposed model outperforms various competing models.

This paper proposes a novel hybrid multiscale decomposition and bootstrap approach for carbon price interval forecasting, aiming to overcome the limitations of traditional carbon price point forecasting. The original carbon price is decomposed into simple modes using the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), and various bootstrap methods are applied to perform a random sampling with a replacement on each mode, generating pseudo datasets forecasted using extreme gradient boosting (XGB). The forecasting values of all modes are then integrated into the original carbon price interval forecasting values. The empirical results, based on samples from China's Guangdong and Hubei carbon markets, provide compelling evidence of the effectiveness of our model. It achieves higher forecasting accuracy, higher interval coverage, and narrower forecasting intervals than currently popular prediction models, instilling confidence in its potential to enhance carbon price forecasting.

In this article, we evaluate the predictive power of multiple machine learning methods using different sets of information, such as location, amenities, socioeconomic characteristics, and available infrastructure nearby, in both residential and commercial real estate markets. This analysis allows us to understand what type of information is the most relevant for each market, which methods are best suited for certain explanatory variables, and the degree of complementarity among different covariates. Our results indicate that the combination of multiple data sources consistently leads to better forecasting and that flexible machine learning models outperform linear regression or spatial methods by taking advantage of the complex interactions between explanatory variables of different sources. From a substantive point of view, we found that residential sale markets have a higher prediction error compared with their rent counterparts, with house sales being the market with the largest estimation error. In terms of the explanatory power of different information sets in different markets, we observe that socioeconomic and location variables have the highest impact on the prediction for sale markets and that, in relative terms, amenities and proximity to places of interest are more important for rental than sale residential markets.

This study aims to forecast the S&P 500 ESG index using the mixture model of complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and long short-term memory (LSTM) prediction models. CEEMDAN enables decomposing the index's original return series into different intrinsic mode functions (IMFs) and a residual series. The decomposed IMFs are then regrouped into aggregate series depicting high frequency and medium frequency, while the residual series represent the trend component. LSTM algorithm is used on the aggregated series to obtain predicted values of the same. The study compares different prediction algorithms to identify their performance and explore the predictive power of the hybrid models.

Was the Depression forecastable? In this paper, we test how effective diffusion indexes are in forecasting the deepest recession in US history: the Great Depression. In a seminal paper, Moore considered the effectiveness of diffusion indexes, though retrospectively and not out-of-sample. We reconstruct Moore's diffusion indexes for this historical period and make our own comparable indexes for out-of-sample predictions. We find that diffusion indexes, including the horizon-specific ones we produce, can nowcast turning points fairly well. Forecasting remains difficult, but our results suggest that the initial downturn in 1929 may be forecastable months before the Great Crash. This is a novel result, as previous authors had generally found the Depression was not forecastable.

We study the short-term predictability of price limit hits. This limit on the trading price is a policy measure imposed with the intention of stabilizing the markets and has been in place for several decades in the Chinese stock markets. We employ feature engineering on past return data and train machine learning models for each individual stock. The results show that a mildly complex model based on ensembling and downsampling the historical information of the majority class (“non-hit” samples) can substantially improve the forecast performance of a naive guess of 50% to about 66% in terms of balanced classification accuracy between true positives and true negatives. We also find that price limit hits of older stocks and of stocks belonging to the tertiary sector are more predictable. We interpret this result with the argument that certain stocks with a longer history are more susceptible to speculative behavior, thus increasing the probability and predictability of such price limit hits.