Nowcasting Russian GDP using forecast combination approach☆
Introduction
Medium-term forecasting in most central banks largely depends on the estimates of the current macroeconomic situation. The reason for this is that large structural and semi-structural models are not always capable to capture economic dynamics over a short-term forecast horizon, because they are rather aimed at medium-term forecasting. That is why finding an efficient technique for short-term forecasting is the most relevant task for many central banks, including the Bank of Russia.
Since statistics of most macroeconomic indicators have publication lags and a relatively low frequency, economists have to estimate not only current and future economic activity, but also the recent past. In academic literature, this process is called nowcasting. A major challenge in nowcasting is the selection of the best model due to the huge range of various forecast techniques. These models may differ in specifications, assumptions, and data used, while it is hard to rank most of them by forecasting quality because in practice they sometimes produce significantly varying results for different countries and samples. Since macroeconomists need to select the ‘only true’ model, the focus of studies has gradually shifted in recent years towards combining multiple forecasts.
Forecast combination is a relatively simple technique involving the averaging of independent forecasts produced by various models. An apparent advantage of this approach is that it relies on a broad dataset, without compromising the quality of forecasting due to potential problems with the curse of dimensionality. This technique involves combining various indicators instead of incorporating them all at once in a model. As a result, macroeconomists may use the entire data pool available in their forecasting. The second advantage of this method is its adaptability since it implies that the structure of a model is flexible and may be re-evaluated based on incoming data. This makes it possible to automatically adjust weights in the aggregate forecast due to changes in their interconnections caused by structural shocks in the economy. The last but not the least benefit of forecast combination is the possibility to diversify accidental errors in the models.
The aim of this paper is to analyse the accuracy of forecast combination in estimating economic growth in Russia. For this purpose, we suggest a forecast combination technique including the most advanced nowcasting methods. We implement this technique in several stages. At the first stage, we collect a large array of economic indicators. The novelty of our research is that we combine several groups of predictors: key scenario indicators used in the Bank of Russia's forecasts (oil prices, exchange rate, etc.) and a vast array of monthly statistics, which allows us to make preliminary estimates of the economic situation. At the second stage, we build a dynamic factor model that identifies common unobservable factors from the array of collected indicators, divided into 3 groups – real sector indicators, the financial sector indicators and agent's expectations. Finally, we include these factors and scenario indicators into various model combinations, the forecasts from which are aggregated with weights. The weight of each model variation is calculated based on the forecast accuracy in the past. The total number of the models used in our method is approximately 500. Using out-of-sample forecasts in a pseudo real-time sample over the period from January 2003 to December 2020, we demonstrate the advantages of our forecasting technique compared to the standard set of benchmark models. As far as we know, forecast combination based on mixed-frequency factor models, with account of the ragged edge problem, has not been described in the Russian academic literature, which makes our research especially relevant.
The rest of the article is organized as follows. In Section 2, we discuss the development of research on short-term forecasting of economic activity and focus on the various forecasting methods, such as bridge equations, factor and mixed-frequency models, forecast combination approaches. In section 3 Methodology, 4 Data, we provide the research methodology, specifying the models and data used in the article. In section 5, we report the results of our forecast combination accuracy comparing to popular benchmark models. The last section concludes.
Section snippets
Literature review
The use of bridge equations is one of the most popular methods implemented by many central banks for short-term forecasting. The key idea of these equations is to bridge a target indicator with one or several key variables released without significant time lags, with the frequencies of these predictors converted into a single one. According to the majority of studies, models of this type very often produce more accurate forecasts than simple models. For instance, Baffigi et al. (2004) try to
Methodology
Our approach embraces most advanced techniques for short-term GDP forecasting described in the academic literature. The GDP forecasting scheme in its general form is presented in Fig. 1. The headline GDP forecast is aggregated from the forecasts of every GDP by expenditure component. Since the dynamics of these components may be described through various processes, the model designed solely for aggregate GDP may be misspecified. However, the weighted sum of GDP by expenditure components may not
Data
In this study we use the chain indices of GDP by expenditure components in fixed prices and of real-time indicators (as % QoQ SA for quarterly predictors and as % MoM SA for monthly predictors). We seasonally adjust all series through the X-13 ARIMA-SEATS method. The series of GDP by expenditure and scenario indicators are used at a quarterly frequency for the period from 2003 Q2 through 2020 Q4, with the observations in the sample numbering 70. Real-time indicators are specified at a monthly
Results
We compare the accuracy of the pseudo-real-time nowcasting of GDP yielded by the forecast combination technique with similar forecasts generated by the benchmark models. All the models are evaluated on data for the period from 2003 Q2 through 2014 Q4; and their out-of-sample mean squared forecast error (MSFE) for a rolling period is estimated over the period from 2015 Q1 through 2020 Q4.
Fig. 4, Fig. 5, Fig. 6 demonstrate the unobserved common factors extracted in our model from high-frequency
Conclusions
Our approach yields the best accuracy of out-of-sample GDP forecasts in Russia for the period from 2011 to 2020 in comparison with the standard benchmark models, but only over a short-term horizon. The difference in the prediction accuracy over 1-2-quarter forecast horizon is significant; nevertheless, as the sample expands in the future, it may be necessary to reassess the quality of models. An important result is that our technique does not generate a systematic forecast error, which may
References (37)
- et al.
Regression models with mixed sampling frequencies
J. Econom.
(2010) - et al.
Bridge models to forecast the euro area GDP
Int. J. Forecast.
(2004) - et al.
Are more data always better for factor analysis?//
J. Econom.
(2006) - et al.
Nowcasting and forecasting GDP in emerging markets using global financial and macroeconomic diffusion indexes
Int. J. Forecast.
(2019) - et al.
A two-step estimator for large approximate dynamic factor models based on Kalman filtering
J. Econom.
(2011) - et al.
Predicting volatility: getting the most out of return data sampled at different frequencies
J. Econom.
(2006) - et al.
Small D. Nowcasting: the real-time informational content of macroeconomic data
J. Monetary Econ.
(2008) Forecasting euro area inflation: does aggregating forecasts by HICP component improve forecast accuracy?
Int. J. Forecast.
(2005)- et al.
Mining big data using parsimonious factor, machine learning, variable selection and shrinkage methods
Int. J. Forecast.
(2018) - et al.
UK macroeconomic forecasting with many predictors: which models forecast best and when do they do so?
Econ. Modell.
(2011)
MIDAS vs. mixed-frequency VAR: nowcasting GDP in the euro area
Int. J. Forecast.
Forecast combination through dimension reduction techniques
Int. J. Forecast.
Real-time forecasting of German GDP based on a large factor model with monthly and quarterly data
Int. J. Forecast.
Forecast combinations
Handb. Econ. Forecast.
Nowcasting of the Russian GDP Using Current Statistics: Approach Modification
Measuring the effects of monetary policy: a factor-augmented vector autoregressive (FAVAR) approach//The Quarterly
J. Econ.
Tail Forecasting with Multivariate Bayesian Additive Regression Trees
Cited by (8)
The economic impact of conflict-related and policy uncertainty shocks: The case of Russia
2023, International EconomicsA MACHINE LEARNING APPROACH TO GDP NOWCASTING: AN EMERGING MARKET EXPERIENCE
2023, Buletin Ekonomi Moneter dan PerbankanApplication of Markov-Switching MIDAS models to nowcasting of GDP and its components
2023, Applied EconometricsNowcasting of the Components of Russian GDP
2022, HSE Economic Journal
- ☆
The views expressed in the paper are solely those of the author and do not necessarily reflect the official position of the Bank of Russia. The author expresses deep gratitude to the specialists of the Bank of Russia's Monetary Policy Department, Research and Forecasting Department and anonymous referees for their valuable advice and helpful comments.