Forecasting the implications of foreign exchange reserve accumulation with a microsimulation model

ABSTRACT

We develop a stock-flow-consistent microsimulation model that comprises all relevant mechanisms of money creation and parametrise it to fit actual data. The model is used to make out-of-sample projections of broad money and credit developments under the commencement/termination of foreign reserve accumulation by the Bank of Russia. We use direct forecasts from the miscrosimulation model as well as the two-step approach, which implies the use of artificial data to pre-train the Bayesian vector autoregression model. We conclude that the suggested approach is competitive in forecasting and yields promising results.

KEYWORDS:

• Money supply
• foreign exchange reserves
• forecasting
• microsimulation
• Russia

Acknowledgments

The views expressed in this paper are those of the authors and do not necessarily represent the position of the Bank of Russia. We are grateful to the anonymous referees, Valery Charnavoki, Vadim Grishchenko and the participants in the 24^th International Conference on Computing in Economics and Finance, WEHIA 2018 and the seminar at the Bank of Russia for their helpful comments and suggestions.

Disclosure statement

No potential conflict of interest was reported by the authors.

Data availability statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Notes

1. See International Relations Committee Task Force (2006), Mohanty and Turner (2006), Cook and Yetman (2012), Filardo and Grenville (2012), Filardo and Yetman (2012), Gadanecz et al. (2014), and Blanchard et al. (2016).

2. See Aizenman and Glick (2009), Ouyang and Rajan (2011), and Cavoli and Rajan (2015).

3. See Bindseil (2004), Borio and Disyatat (2010), Carpenter and Demiralp (2012), and Bundesbank (2017).

4. In fact, to the best of our knowledge, this is the first example to use an MSM for forecasting.

5. When the reserve accumulation policy implies that the central bank sells domestic assets (i.e. claims on the banking sector) to the foreign sector in exchange for foreign reserves, an increase in foreign assets of the central bank will be offset by an increase in commercial banks’ liabilities that they now owe to foreigners. Accordingly, the NFA of the banking system will not change. In the (less common) case in which the central bank sells claims on the non-banking domestic sector to foreigners in exchange for foreign reserves, the NFA of the banking system will increase but will be offset by a decrease in credit to the non-banking sector (see equation (2)), reflecting that money was not created.

6. The accumulation of net foreign assets/liabilities is usually associated with a widening of currency mismatches, which are undesirable (and in many cases forbidden by banking regulations). See Luca and Petrova (2008) for a discussion of the relationship between banks’ net foreign assets and the currency mismatch in domestic assets/liabilities. Another, more general, determinant of flexibility in commercial banks’ NFA may be related to capital mobility. Gagnon (2012, 2013)) and Bayoumi and Saborowski (2014) point out that, in the presence of capital controls, a balance of payments adjustment to FXIs is more likely to happen through the current account.

7. In theory (Lavoie, 1999; Tobin, 1963), the “excess” money created by external transactions can subsequently be destroyed by the adjustment of credit. This, however, would require the newly created money to be transferred immediately to the indebted agents, which would use these funds to repay their loans instead of increasing their consumption. The agent-based framework allows us to model this process more realistically.

8. Defined in terms of such indicators as the loans to deposits ratio, net stable funding ratio or liquidity creation (Berger & Bouwman, 2009).

9. We exclude the observations for 2009 from the analysis due to dramatic exchange rate fluctuations and ensuing changes in foreign currency nominated items on the banking system’s balance sheet (which occurred due to both portfolio shifts and the re-evaluation effect). For the same reason, we do not include the observations for 2014–2015. As stated in Section 3, modelling financial dollarisation developments is beyond the scope of this paper.

10. Although DSGE models that address the issue of foreign exchange reserve accumulation are far from unprecedented, they do not fully accommodate the credit and money creation mechanisms and the evolution of banks’ balance sheets (see Jakab and Kumhof (2015) for a notable exception). Instead, these models are based on a set of simplifying assumptions. For example, García-Cicco (2011) assumes that the central bank may directly control broad money. Benes et al. (2015) assume that credit spreads are affected by changes in the stock of the central bank’s FX reserves.

11. The most notable examples include the EURACE model (Cincotti et al., 2010; Dawid et al., 2014, 2018; Teglio et al., 2019), the JAMEL model (Seppecher, 2012; Seppecher & Salle, 2015; Seppecher et al., 2019), the K + S model (Dosi et al., 2013, 2015, 2010, 2017, 2019), the CATS model (Assenza et al., 2015; Caiani et al., 2018, 2019; Gatti et al., 2011, 2010).

12. For brevity, we only show notations for domestic producers when the equations are exactly analogous for both types of agents.

13. The debt service ratio is calculated as the sum of the principal and interest payments on all of an agent’s loans expected this month as the ratio to permanent income.

14. If the central bank sells foreign reserves, this item is nominated in foreign currency and is moved to the nominator in the equation above.

15. We deem this sufficient as we found that the fit of the best performing set of parameters does not increase sufficiently if the number of sets is increased beyond 2000.

16. These priors are selected arbitrary although before launching the fitting algorithm we ensured that the actual money and credit developments are within the range generated by models with random parameter values drawn from the prior distribution. The set of priors is the same for both sub-samples. The only difference in initial conditions for the two sub-samples is the starting values of deposits and cash, which are roughly calibrated to match the loans to money ratio observed at the start of the sub-samples.

17. We start collecting money and credit series after a 10-period-long burn-in period during which the exogenous variables are fixed at the sample means.

18. Over the time samples under analysis, this variable was almost entirely driven by the changes in the Bank of Russia’s net foreign assets and therefore may be regarded as a policy variable. Admittedly, when this is not the case (e.g., in times of evolution of domestic financial dollarisation), the MSM should be augmented to model the endogenous developments of commercial banks’ net foreign assets.

19. Namely, we produce a simulation using each of 2000 parameter sets and calculate the weighted average as outlined by Tran et al. (2016). In turn each of 2000 simulations is the median of 100 independent model runs under the same set of parameters. This algorithm is also used for forecasting in Section 5.

20. Arguably, this procedure is methodologically close to imposing the DSGE-based priors on VAR models (Del Negro & Schorfheide, 2004) or utilising synthetic data to train neural networks (Gupta et al., 2016; Jaderberg et al., 2014, 2016).

21. We compile the artificial data set by generating alternate 10-month-long “tranquil” and “active” periods of observations. In tranquil periods, SWF_t and FXI_t variables are distributed $\text{~}N\left(0,0.015\right)$ and in “active” periods FXI_t$\text{~}N\left(0.02,0.015\right)$ and SWF_t $\text{~}N\left(0.01,0.015\right)$. Other exogenous variables are fixed at mean values. Endogenous variables (money and credit) are generated by running the MSM conditional on these sets of exogenous variables. The total number of observations in the artificial data set is equal to half of the number of observations in the respective training sample.