Nowcasting the Indian economy

Policymakers, economists and financial market analysts are heavily reliant on economic activity data for the purpose of adequate decision-making. Gross domestic product (GDP) is the conventional measure to gauge economic activity, but the problem with GDP data is that it is not available real time and is published by national statistical offices with a certain time lag. In order to circumvent this problem, economists increasingly have been focussing on nowcasting tools to make a proper assessment of the state of economy, now and in the near future. The basic principle of nowcasting (a contraction of the words now and forecasting) is the exploitation of data at higher frequencies to get an early estimate of the target variable before the actual print is released. For instance, in nowcasting GDP, ‘hard’ data (such as industrial production) is often combined with ‘soft’ data from surveys, such as the purchasing managers’ index (PMI).

Nowcasting has especially become popular in the aftermath of the global financial crisis and currently many central banks in especially developed countries are using these tools for monetary policy purposes (see, e.g. NY Fed and ECB, 2010). Nowcasting as a methodology has recently also been picked up in emerging markets, such as Mexico, Indonesia, Turkey and India. In emerging markets, however, proper nowcasting is perhaps even more complex than in developed countries, due to often inferior quality of statistics and unconventional data revisions resulting in severe breaks in data series. In this Special we will develop our own nowcasting models for the Indian economy, which ultimately combines a simple OLS (ordinary least square) model with a Bayesian Vector Autoregressive (BVAR) model.

Rabobank started forecasting the Indian economy using an integral model-based approach in 2017 (see Figure 1 for an overview).

Over calendar 2017Q1 to 2018Q2[1], the average prediction error of our GDP forecasts was 3.1% (Figure 2)[2], which was more than twice as low as the average prediction error (7.1%) of Bloomberg’s Consensus[3]. However, our average prediction error rose significantly (to 4.3%) after India’s most recent GDP release for calendar Q3. We had pencilled in a GDP growth forecast of 8% (y-o-y), which was much higher than the actual print of 7.1% (see figure 2).

The deterioration of our prediction accuracy underlines that, going forward, our forecasts should ideally require cross-validation using alternative approaches. The most appropriate approach in this case would be the development of robust Indian nowcasting models.

[1] All references to dates (be it quarters or years) refer to calendar dates instead of fiscal dates.

[2] To calculate the prediction accuracy, we use the mean absolute percentage error (M), which is defined as , where Yt is actual GDP, is the forecast value and n represents the number of observations.

[3] Bloomberg’s Consensus is the median forecasts of a large number of economists in their economic survey. Somewhere between 35 to 45 economists around the globe usually participate in the India GDP survey.

Before Giannone et al. (2008) published their seminal article on nowcasting in 2008, the academic literature developed so-called small-scaled ‘bridge equations’, which focussed on bridging the gap between quarterly GDP growth and information on economic activity embodied in monthly indicators (e.g. Baffigi et al., 2004). The problem with bridge equations is that they are incapable to deal with large sets of information, as single-equation models have limited degrees of freedom. Stock and Watson (2002) used principal component analyses to reduce a large set of indicators in a small number of indices, which provided the groundworks of dynamic factor models (DFM). The DFM approach was put to use by Giannone et al. (2008) for nowcasting purposes. Basically the framework by Giannone et al. (2008) should be regarded as a large bridge model which: 1) uses 200 time series for the US economy, 2) ‘bridges’ monthly data released with quarterly GDP data, 3) updates the model and integrates new information into the forecasts.

The DFM model basically exploits collinearity of the underlying series to derive a number of common factors. This way, the information of a vast amount of series can be exploited, without running into the problem of overfitting, which results in poor (out-of-sample) forecast results. Another approach to circumvent overfitting are the use of so-called Bayesian Vector Autogressive (BVAR) models (see Litterman, 1986). According to Bańbura et al. (2010), BVAR models are a valid alternative to factor models or for the analysis of large dynamic systems. For instance, the CPB Netherlands Bureau of Policy Analysis shows that BVAR models outperform more conventional methods (i.e. Okun’s Law) in forecasting Dutch unemployment (CPB, 2018). In practice, BVAR models are complex VAR models which use a large set of variables, but also adopts so-called Bayesian shrinkage to prevent the problem of overfitting. For more information on the technical aspects of our BVAR model, see the Annex.

India’s GDP data is published by the Central Statistical Office (CSO) every quarter, with a lag of approximately two months. To nowcast Indian GDP, we use data published on a monthly frequency (see Table 1). The indicators all cover different aspects of the Indian economy, such as private consumption, industrial sector activity and services sector activity. In Annex 2, we have included a heatmap illustrating the dynamics on these high-frequency indicators. However, the direction of these indicators do not necessarily provide information about their impact on the Indian economy. For that purpose we need economic modelling. We distinguish between two types of models: BVAR models and simple OLS models.

BVAR model

Our BVAR model cover the period 2004Q1 to 2017Q1, We have experimented with a different set of explanatory variables, dependent on data availability over time and lags in the interdependency matrix. We use the Minnesota prior (see Annex 1) to estimate the BVAR. An important precondition is that all variable are stationary. We use ADF tests to examine stationarity for each individual variable. In case of non-stationarity, we use differencing until the series is stationary.

OLS models

A significant drawback of BVAR models is that we are unable to trace back the dynamic underpinnings of the GDP forecast. For instance, the BVAR might predict that the Indian economy will grow by X% in Q4, but is not able to tell whether this figure is due to favourable private consumption growth (indicated by, for example, higher vehicle sales) or services sector activity. As the BVAR model basically estimates a large interdependency matrix containing cross lags, it is very difficult to interpret the calculated coefficients and derive the contributions of the individual underlying high-frequency indicators to the GDP growth forecast. To circumvent this problem, we also estimate simple OLS models for two sub samples: a full sample (FS) OLS covering the period 2003Q3-2017Q1 and a small sample (SS) OLS for 2007Q3-2017Q1. This model takes the general form of:

where Y is GDP, X is a vector of high-frequency leading indicators, t is a time index and Δ denotes first differencing.

Out-of-sample testing

In order to test the accuracy of our models, we use out-of-sample testing. This means that we do not fit the model for the entire data sample, but exclude GDP realisations from the data sample to test the prediction accuracy of our models. Figure 3 illustrates how out-of-sample testing works in practice. First, we estimate our model up to 2017Q1 and we cross validate our models predictions with actual GDP realisations for the period 2017Q2 up to 2018Q3. For nowcasting purposes, we first optimise our model for the complete sample period and use this model to forecast GDP for the next quarter.

We have tested various models using different combinations of the leading indicators described in Table 1. The optimal BVAR model adopts three lags in the interdependency matrix in combination with following leading indicators: vehicle sales, car registrations, oil consumption, personal loans, industrial production, the services sector PMI, vehicle exports, the monetary base and the Sensex equity index. Table 2 shows the estimation results of our OLS models, where D denotes first differencing (y-o-y), log is the natural log and t indicates quarter.

Most leading indicators show a statistically significant impact on GDP. For instance, an increase in industrial production by 10ppts results in an increase of GDP growth by roughly 1.5ppts to 2.5ppts. What is striking in the results for the regressions using different samples (which dependents on data availability for individual leading indicators) is that we do observe more or less a transition of important growth drivers. For instance, in the full sample, inflation has a statistically significant negative impact on Indian economic growth, whereas the negative impact becomes less robust when we adopt a more narrow estimation sample.

Out-of-sample testing

With our optimal models in place, we apply out-of-sample testing to our three models. The out-of-sample test of BVAR model is shown in Figure 4. Although the in-sample fit is quite decent, the model tends to undershoot the peaks and troughs, especially in the period 2012-2017. One upside of the model is that it does pick up the important slump in the aftermath of demonetisation in 2017Q2.

The short sample (SS) OLS model (column 2 in Table 2) significantly outperforms the full sample (SS) OLS (column 1 in Table 2) in terms out-of-sample prediction accuracy (compare Figure 5 and Figure 6). The FS OLS model especially seems to have a good fit with actual Indian growth in the period before and during the Global Financial Crisis, but seems less accurate in the period 2012-2017. This could possibly be explained that the fact the FS model lacks proper services sector indicators and external sector indicators, whereas we know that the economic structure of emerging markets is transforming at a rapid pace. This transition is perhaps picked up more accurately by the SS OLS model which does incorporate service and external sector indicators. The SS OLS model has a good fit with Indian growth over the entire sample period. The out-of-sample test also accurately follows the pattern of the actual realisations, but also tends to overshoot growth by a margin.

Nowcasting

As the BVAR model tends to underestimate Indian economic growth and the SS OLS model tends to marginally overshoot growth, we will combine information of both models to optimise our prediction accuracy. This is done by estimating:

where is the fit of our BVAR model and the fit of our SS OLS model. Estimating equation (2) shows that the combined model attributes a weight of 60% to the BVAR model and 40% to the SS OLS.

In a second step, we use ARIMA models to extrapolate the high-frequency indicators (see Table A.2 in Annex 2) for the full quarter. This information in combination with the monthly realisation for the high-frequency indicators is subsequently used to produce our nowcast. Our combined nowcasting model shows that growth for the fourth quarter is expected to arrive at 6.3% (Figure 7).

As shown in the figure, the bandwidth of our nowcast is far from narrow. Our SS OLS predicts GDP growth for Q4 of 6.9% and the BVAR shows a prediction of 6.0%. Moreover, keep in mind that these individual model outcomes are also subject to a substantial degree of uncertainty.

Given the outcome of our nowcasting model of 6.3% for calendar 2018Q4, this raises the question: should India brace for a new round of economic weakness, after experiencing adverse demonetisation effects on the economy only six quarters ago? Given the broad weakness in economic data in numerous countries (see reports covering the Euro Area, the US, China and recessionary risks in general), a rapid slowdown in India would certainly fit the global pattern. Weakness in the external environment also seems to have repercussions for India’s external sector (see variable vehicle export in Table A.2 of Annex 2). But also on the internal front, we see increasing weakness in private consumption and industrial production, whereas the services sector seems to be holding up quite well (again see Table A.2 of Annex 2).

Despite weakness in the fourth quarter, we expect growth to rebound in 2019Q1 and 2019Q2 on the back of favourable fiscal and monetary policies. As expected, the FY2019/20 Union Budget, presented on 1 February to Parliament, contains a number income support programs. One of the most noteworthy measures is a farmer income support program of 75bn INR (USD 10.5bn), which is broken down in trances of 20bn for the fiscal 2019/20 and 55bn for fiscal 2020/21. See here a link to the Union Budget and our India 2019 Economic Outlook. Ultimately, we revise downward our expected GDP growth rate for fiscal 2018/19 from 7.4% to 7.1% (Figure 8).

In this Special we present our new nowcasting methodology. Nowcasting is a contraction of the words now and forecasting and is based on the principle of exploiting high-frequency leading data to gauge economic activity. Nowcasting has been increasingly popular as a tool to estimate growth of gross domestic product (GDP), as this indicator of economic activity is published by national statistical offices with a certain time lag.

The motivation to develop a new nowcasting methodology lies in the fact that Rabobank’s average prediction error for Indian GDP dropped significantly after India’s latest GDP release, from 3.1% to 4.3% (over the period 2017Q1-2018Q3). Although this error is still lower than Bloomberg’s Consensus of 7.1%, the deterioration in accuracy underlines that, going forward, our GDP forecasts should ideally require cross validation using nowcasting models.

After experimenting with several subsets of high-frequency indicators and performing out-of-sample testing, our preferred model combines predictions from a so-called Bayesian Vector Autoregressive (BVAR) model and an Ordinary Least Square (OLS) model. These two models use information on vehicle sales, car registrations, electricity production, oil consumption, personal loans, industrial production, the services sector PMI, vehicle exports, inflation, the monetary base and the Sensex equity index.

Our combined nowcast model shows that economic growth for the fourth quarter of 2018 is expected to arrive at 6.3%, which would imply a significant slowdown of economic activity after the already disappointing GDP print of 7.1% for the third quarter. We have to stress though that the bandwidth of our nowcast is far from narrow. Our upper bound is 6.9% and the lower bound is 6.0%. So there is still a large degree of uncertainty surrounding predictions of Indian economic growth. All in all, we are pretty confident that India’s next GDP print will put the Indian economy on a downward-sloping growth trajectory. Nevertheless we also expect growth to rebound in 2019Q1 and Q2 on the back of favourable fiscal and monetary policies.

The expected slowdown is not surprising given the disappointing global economic data, which also has repercussions for India’s external sector. On the internal front, we see increasing weakness in private consumption and industrial production, whereas the services sector seems to be holding up quite well. Ultimately, we revise downward our expected GDP growth rate for fiscal 2018/19 from 7.4% to 7.1%.

Baffigi, A., Golinelli, R. and G. Parigi (2004). Bridge models to forecast the euro area GDP. International Journal of Forecasting, 20 (3), 447–460.

Bańbura, M., Giannone, D. and L. Reichlin (2010). Large Bayesian vector auto regressions. Journal of Applied Econometrics, 25(1), 71-92.

ECB (2010). Nowcasting. ECB Working Paper Series no. 1275, Frankfurt.

Delajara, M., Álvarez, F. H. and A.R. Tirado (2016). Nowcasting Mexico's short-term GDP growth in real-time: A factor model versus professional forecasters. Economía, 17(1), 167-182.

Giannone, D., Reichlin, L. and D. Small (2008). Nowcasting: the real time informational content of macroeconomic data. Journal of Monetary Economics, 55, 665-676.

Giannone, D., Lenza, M. and G.E. Primiceri (2015). Prior selection for vector autoregressions. Review of Economics and Statistics, 97(2), 436-451.

Kitchen, J. and R.M. Monaco (2003). Real-time forecasting in practice: the U.S. treasury staff’s real-time GDP forecast system. Business Economics, 10–19.

Modugno, M., Soybilgen, B. and E. Yazgan (2016). Nowcasting Turkish GDP and news decomposition. International Journal of Forecasting, 32(4), 1369-1384.

Roy, I., Sanyal, A. and A.K. Ghosh (2016). Nowcasting Indian GVA growth in a mixed frequency setup. Reserve Bank of India Occasional Papers, 37(1&2).

Stock, J.H. and M.W. Watson (2002). Forecasting using principal components from a large number of predictors. Journal of the American statistical association, 97(460), 1167-1179.

Woźniak, T. (2016). Bayesian vector autoregressions. Australian Economic Review, 49(3), 365-380.

To shed more light on how our BVAR model work, let us consider the following VAR model:

where Y t is a large vector with observed variables. In our case, , where X is Indian GDP growth and Z is a vector of leading indicators. Furthermore, in equation (A.1) c is a vector of constants, Ai is a n x n matrix of autoregressive parameters which are dependent on the number of lags i. Finally, is an idiosyncratic error term. The parameters of the VAR model are estimated using the likelihood function (see for a technical representation Wozniák, 2016). The problem with VAR models is that they require many parameters, even with a limited amount of explanatory variables. To overcome the danger of overfitting, BVAR models are often used which imposes priors to the parameters.

One of the most common priors used is the Minnesota prior (Litterman, 1986), which is also the one that we adopt for our Indian BVAR model. The basic principle of the Minnesota prior is that all equation are ‘centered’ around a random walk, which is a reasonable approximation of the behaviour of an economic variable. The prior adopts the following distribution of the coefficients and covariance V:

where coefficients A1, …, Ai are assumed to be normally distributed, σi is the diagonal element i of . The hyperparameters λ1, λ2 and λ3 control the tightness of the prior distribution and the scale of the variance and covariances. More specifically, λ1 determines the relative tightness (in our case 0.1), λ2 captures the cross-variable weight (in our case: 0.99) and λ3 captures the lag decay (in our case 1). Ultimately, given our chosen parameters, the Minnesota prior assumes that the variance is lower for the coefficients associated with more distant lags. For more information on aspects of Bayesian estimation techniques and applications for forecasting, we refer to Litterman (1986), Banburá et al. (2010) and Giannone et al.(2015).