MONETARY POLICY REPORT | WINTER 2021 (box 3)
(authors: Branislav Saxa, Eva Hromádková, Iva Kubicová)
Credit did not escape the adverse effects of the pandemic. Loans to households for consumption were hit the hardest, and growth of loans to corporations also dropped, whereas growth in house purchase loans surged. Credit is an important element of monetary policy transmission. From now on, a forecast of loans by sector will be a regular part of the Chartbook, which is a parallel publication to the Monetary Policy Report. The outlook for the debt financing of economic activity can be used not only to better understand the links between the credit cycle and corporate investment activity, but also to quantify risks connected with household debt. This box describes the new model we use to forecast loans.
The main explained variables in the model are loans to the non-financial sector in their basic breakdown, i.e. loans to non-financial corporations and loans to households for house purchase and consumption. Changes in the stocks of these variables can be explained by the output gap (the deviation of GDP from potential output), the deviation of inflation from the target, financial market interest rates and exogenous factors (e.g. drawdown of EU funds). Outlooks for these variables obtained from the CNB’s macroeconomic forecast using additional satellite models (e.g. for the ten-year government bond yield) enter the credit forecast.
The forecast methodology was selected with two main objectives in mind: first, to maximise its predictive ability, and second, to ensure that the predicted variables respond to changes in input variables in the economically intuitive direction in the Czech economy. Having regard to these objectives, we decided to use a simple Bayesian VAR model.[1]
The charts illustrate the model’s predictive ability using historical data. The simulations of past forecasts use observed data and authentic CNB quarterly macroeconomic forecasts available at the time. All the models show a relatively small mean absolute forecast error, especially by comparison with the other modelling approaches tested (SARIMA, VECM and other specifications of the Bayesian VAR model). The forecast for loans to non-financial corporations has the largest errors (see Chart 1; errors in the note below the chart). This is linked with their high volatility. By contrast, the forecast error for loans for house purchase is very low (see Chart 2). As the chart shows, however, most of the forecasts underestimated growth in house purchase loans in the period under review. This is because demand for house purchase loans in 2015–2020 was affected, besides the macroeconomic situation, by factors not captured by the model.[2]
Chart 1 – Loans to non-financial corporations
Forecasts based on observed data and their error at the one-quarter horizon; growth rates in %
Note: The mean absolute error (MAE) is 1.9 pp for the one-quarter forecast and 3.2 pp for the one-year forecast.
Chart 2 – Loans to households for house purchase
Forecasts based on observed data and their error at the one-quarter horizon; growth rates in %
Note: The mean absolute error (MAE) is 0.4 pp for the one-quarter forecast and 1.2 pp for the one-year forecast.
Finally, the model for loans to households for consumption has an excellent predictive ability, especially at the short end (see Chart 3). This is due to its strong link to GDP and above all consumption, which is well captured by the model.
Chart 3 – Loans to households for consumption
Forecasts based on observed data and their error at the one-quarter horizon; growth rates in %
Note: The mean absolute error (MAE) is 0.3 pp for the one-quarter forecast and 1.2 pp for the one-year forecast.
A comparison of the forecasts with the subsequently observed loan volumes during 2020 illustrates the importance of expert adjustment in the preparation of the forecast. An extreme example is the purely model-based forecast of a 25% drop in loans to non-financial corporations in 2020 Q2. This failed to materialise, due partly to a subsequent decrease in the expected fall in GDP. However, corporations’ reaction played a big role. They cut their investment activity as expected, but they also increased their demand for operating loans for liquidity reasons (including state-guaranteed loans under the COVID 1, 2, 3, Praha, EGAP etc. programmes). A sharper decline in corporate loans was also prevented by the loan moratorium. Knowledge, or mere expectations, of such structural factors must therefore be taken into account in the credit forecast through additional expert adjustments.
[1] A. Dieppe, R. Legrand, B. van Roye (2018): Bayesian Estimation, Analysis and Regression Toolbox, version 4.2.
[2] In particular persistent excess demand on the residential property market.