nep-for New Economics Papers
on Forecasting
Issue of 2021‒06‒14
eight papers chosen by
Rob J Hyndman
Monash University

  1. Forecasting Macroeconomic Variables in Emerging Economies: An Application to Vietnam By Le Ha Thu; Roberto Leon-Gonzalez
  2. Macroeconomic Forecasting and Variable Ordering in Multivariate Stochastic Volatility Models By Jonas E. Arias; Juan F. Rubio-Ramirez; Minchul Shin
  3. Forecasting UK GDP growth with large survey panels By Anesti, Nikoleta; Kalamara, Eleni; Kapetanios, George
  4. Censored Density Forecasts: Production and Evaluation By James Mitchell; Martin Weale
  5. Forecasting CPI Inflation Components with Hierarchical Recurrent Neural Network By Oren Barkan; Jonathan Benchimol; Itamar Caspi; Allon Hammer; Noam Koenigstein
  6. Forecasting of cohort fertility by educational level in countries with limited data availability: the case of Brazil By Ewa Batyra; Tiziana Leone; Mikko Myrskylä
  7. Blurred Crystal Ball: investigating the forecasting challenges after a great exogenous shock By Marcelo A. T. Aragão
  8. GARCHNet - Value-at-Risk forecasting with novel approach to GARCH models based on neural networks By Mateusz Buczyński; Marcin Chlebus

  1. By: Le Ha Thu (National Graduate Institute for Policy Studies, Tokyo, Japan); Roberto Leon-Gonzalez (National Graduate Institute for Policy Studies, Tokyo, Japan)
    Abstract: Forecasting macroeconomic variables in the rapidly changing macroeconomic envi- ronments faced by developing and emerging countries is an important task for central banks and policy-makers, yet often presents a number of challenges. In addition to the structural changes in the economy, the time-series data are usually available only for a small number of periods, and predictors are available in different lengths and frequencies. Dynamic model averaging (DMA), by allowing the forecasting model to change dynamically over time, permits the use of predictors with different lengths and frequencies for the purpose of forecasting in a rapidly changing economy. This study uses DMA to forecast inflation and growth in Vietnam, and compares its forecast- ing performance with a wide range of other time-series methods. Some results are noteworthy. First, the number and composition of the optimal predictor set changed, indicating changes in the economic relationships over time. Second, DMA frequently produces more accurate forecasts than other forecasting methods for both the inflation and the economic growth rate of Vietnam.
    Keywords: Bayesian, dynamic model averaging, forecasting macroeconomic variables, Vietnam
    Date: 2021–06
  2. By: Jonas E. Arias; Juan F. Rubio-Ramirez; Minchul Shin
    Abstract: We document five novel empirical findings on the well-known potential ordering drawback associated with the time-varying parameter vector autoregression with stochastic volatility developed by Cogley and Sargent (2005) and Primiceri (2005), CSP-SV. First, the ordering does not affect point prediction. Second, the standard deviation of the predictive densities implied by different orderings can differ substantially. Third, the average length of the prediction intervals is also sensitive to the ordering. Fourth, the best ordering for one variable in terms of log-predictive scores does not necessarily imply the best ordering for another variable under the same metric. Fifth, the best ordering for variable x in terms of log-predictive scores tends to put the variable x first while the worst ordering for variable x tends to put the variable x last. Then, we consider two alternative ordering invariant time-varying parameter VAR-SV models: the discounted Wishart SV model (DW-SV) and the dynamic stochastic correlation SV model (DSC-SV). The DW-SV underperforms relative to each ordering of the CSP-SV. The DSC-SV has an out-of-sample forecasting performance comparable to the median outcomes across orderings of the CSP-SV.
    Keywords: Vector Autoregressions; Time-Varying Parameters; Stochastic Volatility; Variable Ordering; Cholesky Decomposition; Wishart Process; Dynamic Conditional Correlation; Out-of-sample Forecasting Evaluation
    JEL: C8 C11 C32 C53
    Date: 2021–06–03
  3. By: Anesti, Nikoleta (Bank of England); Kalamara, Eleni (King’s College London); Kapetanios, George (Bank of England)
    Abstract: By employing large panels of survey data for the UK economy, we aim at reviewing linear approaches for regularisation and dimension reduction combined with techniques from the machine learning literature, like Random Forests, Support Vector Regressions and Neural Networks for forecasting GDP growth at monthly frequency for horizons from one month up to two years ahead. We compare the predictive content of surveys with text based indicators from newspaper articles and a standard macroeconomic data set and extend the empirical evidence on the contribution of survey data against text indicators and more traditional macroeconomic time series in predicting economic activity. Among the linear models, the Ridge and the Partial Least Squares models report the largest gains consistently for most of the forecasting horizons, and for the non‑linear machine learning models, the SVR performs better at shorter horizons compared to the Neural Networks and Random Forest that seem to be more appropriate for longer‑term forecasting. Text based indicators appear to favour more the use of non‑linear models and the expansion of the information set with macroeconomic time series does not appear to add much more predictive power. The largest forecasting gains are overwhelmingly concentrated at the shorter horizons for the majority of models and datasets which provides further empirical support that non‑linear machine learning models appear to be more useful during the Great Recession.
    Keywords: Forecasting; survey data; text indicators; machine learning
    JEL: C53 C55
    Date: 2021–05–28
  4. By: James Mitchell; Martin Weale
    Abstract: This paper develops methods for the production and evaluation of censored density forecasts. Censored density forecasts quantify forecast risks in a middle region of the density covering a specified probability, and ignore the magnitude but not the frequency of outlying observations. We propose a new estimator that fits a potentially skewed and fat-tailed density to the inner observations, acknowledging that the outlying observations may be drawn from a different but unknown distribution. We also introduce a new test for calibration of censored density forecasts. An application using historical forecast errors from the Federal Reserve Board and the Monetary Policy Committee at the Bank of England illustrates the utility of censored density forecasts when quantifying forecast risks after shocks such as the global financial crisis and the COVID-19 pandemic.
    Keywords: Forecast uncertainty; Outliers; Fan charts; Skewed densities; Best critical region; Density forecasting; Censoring; Forecast evaluation
    JEL: C24 C46 C53 E58
    Date: 2021–05–27
  5. By: Oren Barkan (Ariel University); Jonathan Benchimol (Bank of Israel); Itamar Caspi (Bank of Israel); Allon Hammer (Tel-Aviv University); Noam Koenigstein (Tel-Aviv University)
    Abstract: We present a hierarchical architecture based on Recurrent Neural Networks (RNNs) for predicting disaggregated inflation components of the Consumer Price Index (CPI). While the majority of existing research is focused on predicting headline inflation, many economic and financial institutions are interested in its partial disaggregated components. To this end, we developed the novel Hierarchical Recurrent Neural Network (HRNN) model, which utilizes information from higher levels in the CPI hierarchy to improve predictions at the more volatile lower levels. Based on a large dataset from the US CPI-U index, our evaluations indicate that the HRNN model significantly outperforms a vast array of well-known inflation prediction baselines. Our methodology and results provide additional forecasting measures and possibilities to policy and market makers on sectoral and component-specific prices.
    Keywords: Inflation forecasting, Disaggregated inflation, Consumer Price Index, Machine learning, Gated Recurrent Unit, Recurrent Neural Networks
    JEL: C45 C53 E31 E37
    Date: 2021–03
  6. By: Ewa Batyra (Max Planck Institute for Demographic Research, Rostock, Germany); Tiziana Leone; Mikko Myrskylä (Max Planck Institute for Demographic Research, Rostock, Germany)
    Abstract: The Brazilian period fertility rate (PTFR) dropped from six to 1.8 between 1950 and 2010. Due to the shifts in the timing of fertility, the PTFR might be providing a misleading picture of fertility levels. Moreover, the national average hides important educational differences, as in 2010, the PTFR was 2.3 among the lower educated, whereas it had fallen to 1.5 among the higher educated. The consequences of these changes for the cohort total fertility rate (CTFR) – a measure that is free from tempo distortions – and for the educational differences in completed fertility have not been previously studied. Due to the scarcity of time series of fertility rates, the application of CTFR forecasting methods outside of high-income countries (HICs) has been rare, and has been largely limited to population-level analysis. We use four Brazilian censuses to forecast the CTFR for the total population and by educational level using rates reconstructed with indirect techniques. The results of four forecasting methods indicate that the CTFR is likely to decline to 2.1 for the 1980 cohort, and to 1.9 for the 1984 cohort. Educational differences in the CTFR are likely to remain stark – at between 0.7 and 0.9 depending on the cohort and the method – and to be larger than they are in HICs with comparable CTFRs. We show how the CTFR can be forecasted, including by educational level, in settings with limited data. Finally, we call for more research on the educational differences in completed fertility in low- and middle-income countries.
    Keywords: Brazil, census data, cohort fertility, education, forecasts
    JEL: J1 Z0
    Date: 2021
  7. By: Marcelo A. T. Aragão
    Abstract: An event like Covid-19 pandemic brings about a deadly human toll and mayhem to the economy. With such a great exogeneous shock, policy makers and forecasters alike face a set of challenges to keep on contributing to the economic response. Many distinguishing researchers came forward with their own assessments of the lasting macroeconomic impacts of the Covid-19 pandemic. Modestly, this paper attempts a different instance: investigating how a practitioner can cope with some pressing forecasting challenges while avoiding naive pitfalls. Without claiming any quantification, it experiments with usual US economy data sources and macroeconomic models to exemplify these challenges and their possible overcoming. Finally, it summarizes some empirical, pragmatical conclusions.
    Date: 2021–05
  8. By: Mateusz Buczyński (Interdisciplinary Doctoral School, University of Warsaw); Marcin Chlebus (Faculty of Economic Sciences, University of Warsaw)
    Abstract: This study proposes a new GARCH specification, adapting a long short-term memory (LSTM) neural network's architecture. Classical GARCH models have been proven to give substantially good results in the case of financial modeling, where high volatility can be observed. In particular, their high value is often praised in the case of Value-at-Risk. However, the lack of nonlinear structure in most of the approaches entails that the conditional variance is not represented in the model well enough. On the contrary, recent rapid advancement of deep learning methods is said to be capable of describing any nonlinear relationships prominently. We suggest GARCHNet - a nonlinear approach to conditional variance that combines LSTM neural networks with maximum likelihood estimators of probability in GARCH. The distributions of the innovations considered in the paper are: normal, t and skewed t, however the approach does enable extensions to other distributions as well. To evaluate our model, we have executed an empirical study on the log returns of WIG 20 (Warsaw Stock Exchange Index) in four different time periods throughout 2005 and 2021 with varying levels of observed volatility. Our findings confirm the validity of the solution, however we present several directions to develop it further.
    Keywords: Value-at-Risk, GARCH, neural networks, LSTM
    JEL: G32 C52 C53 C58
    Date: 2021

This nep-for issue is ©2021 by Rob J Hyndman. It is provided as is without any express or implied warranty. It may be freely redistributed in whole or in part for any purpose. If distributed in part, please include this notice.
General information on the NEP project can be found at For comments please write to the director of NEP, Marco Novarese at <>. Put “NEP” in the subject, otherwise your mail may be rejected.
NEP’s infrastructure is sponsored by the School of Economics and Finance of Massey University in New Zealand.