NB: authors are listed alphabetically.
NB: authors are listed alphabetically.
A neural network approach to the environmental Kuznets curve
Mikkel Bennedsen, Eric Hillebrand, Sebastian Jensen
Energy Economics, 2023, Volume 126, 106985 (19 pages)
We investigate the relationship between per capita gross domestic product and per capita carbon dioxide emissions using national-level panel data for the period 1960-2018. We propose a novel semiparametric panel data methodology that combines country and time fixed effects with a nonparametric neural network regression component. Globally and for the regions OECD and Asia, we find evidence of an inverse U-shaped relationship, often referred to as an environmental Kuznets curve (EKC), in production-based emissions. For OECD, the EKC-shape disappears when using consumption-based emissions data, suggesting the EKC-shape observed for OECD is driven by emissions exports. For Asia, the EKC-shape becomes even more pronounced when using consumption-based emissions data and exhibits an earlier turning point.Â
Nowcasting U.S. CO2 emissions using machine learning
Sebastian Jensen
I investigate the use of machine learning methods for nowcasting the yearly growth rate of United States carbon dioxide emissions over the period 2000-2019, using a high-dimensional panel of macroeconomic variables sampled at mixed frequencies. To handle the problem of mixed frequencies, I propose to use the frequency alignment transformation from the mixed data sampling regression (MIDAS) literature. I find that neural networks, and to a lesser extent tree-based machine learning methods (random forest, bagging, and gradient boosting), are able to utilize the stream of macroeconomic information that becomes available through the target year to produce repeatedly more accurate nowcasts of U.S. CO2 emissions that generally outperform forecasts from univariate time series models and nowcasts from MIDAS models.
Apocalypse Now? Projecting CO2 Emissions with Neural Networks
Mikkel Bennedsen, Eric Hillebrand, Sebastian Jensen
We project carbon dioxide emissions through 2100 using a reduced-form model and national level scenarios for per capita gross domestic product from the Shared Socioeconomic Pathways (SSPs). We propose a novel neural network-based panel data model that combines country fixed effects with a long short-term memory (LSTM) recurrent neural network regression component that takes into account time implicitly by letting model predictions depend on the income path of a country. Only for scenarios with low socioeconomic challenges for mitigation, SSP1 and SSP4, do our emissions projections appear consistent with baseline projections from structural integrated assessment models (IAMs) that are meant to describe future developments in absence of new climate policies.
The Conditional Akaike Information Criterion for Time-varying Parameter Models
Sebastian Jensen, Siem Jan Koopman, and Jan van den Brakel
We propose using the conditional Akaike Information Criterion (cAIC) for model selection in time-varying parameter models, where the penalty term is computed as the trace of the projection matrix. We provide an efficient procedure for calculating the trace of the projection matrix from the model's state space form using the Kalman filter and a light "disturbance" smoothing algorithm. We demonstrate that the effective number of parameters in most time-varying parameter models can be linked to the parameter variances, and provides a natural generalization of the parameter count in fixed parameter models.
Machine Learning and Filtering Methods for Long-term ENSO Prediction
Sebastian Jensen, Siem Jan Koopman, and Desislava Petrova
We predict El Nino and the Southern Oscillation (ENSO) at long lead times by embedding machine learning methods in a statistical unobserved components time series framework. The framework consists of unobserved components (trend, seasonal, and cycles) and a nonlinear machine learning regression function based on a high-dimensional set of predictor variables, including subsurface temperatures from the Pacific Ocean. We develop a novel estimation procedure that combines machine learning methods with state space methods and filtering techniques (including the Kalman filter) for joint estimation of the unobserved component parameters and the machine learning parameters.
Bayesian Neural Networks for Panel Data
Mikkel Bennedsen, Eric Hillebrand, Sebastian Jensen
We propose a class of Bayesian neural networks for panel data that accounts for country and time fixed effects and a variational Bayes learning framework. The class of model solves two problems often encountered in practice: uncertainty quantification and model selection. By learning a parameter distribution, Bayesian inference naturally accounts for parameter (epistemic) uncertainty. We account for data (aleatoric) uncertainty by modeling the variance conditional on regressors in addition to the mean. We demonstrate that model selection can be carried out efficiently in this framework using the widely applicable information criterion (WAIC).
Use of Machine Learning in Climate Econometrics
Sebastian Jensen
This dissertation consists of three self-contained chapters on the use of machine learning in climate econometrics and is particularly concerned with how tools and ideas from the fields of econometrics and machine learning can be combined to shed new light on the relationship between macroeconomic activity and carbon dioxide (CO2) emissions.