The evaluation of treatment effects has been recognized for over a century. Among the pioneers were the works of O. Anderson from 1912–1915. Further progress was done, for instance, by Abadie and Imbens (2011), de Chaisemartin and D’Haultfoeuille (2020).

The difference-in-differences (or double difference) principle forms the basis­ for estimating treatment effects. To implement this, the sample is divided into four parts: control (not exposed to treatment, control, counterfactual) and pilot (exposed to treatment); pre-/post-treatment. The change in the indicators of the pilot group relative to the control group is attributed to the treatment effect. For the obtained estimate to be reliable, it is necessary to fulfill the parallel trends (pre-trends) assumption. Possible options for conducting the pre-trend tests are described in Mäkinen (2021), Duong (2021), Behncke (2022). Recent reviews of the application of the method include Brewer et al. (2017), Fredriksson and Magalhães de Oliveira (2019), Roth et al. (2023).

There are modifications in the baseline method. For instance, Goodman-Bacon (2021) considers several periods of treatment. However, it is assumed that the control and pilot groups remain unchanged. In the paper by Olden and Moen (2022), the triple­ difference method is proposed in case external effects can occur. For ­example, when a change in healthcare in one region not only directly benefits the pilot sample repre­sentatives but also indirectly benefits the control group representatives of the same region. Therefore, Olden and Moen (2022) suggest comparing the results with a region where there was no treatment (where no pilot groups were identified).

The estimation of effects is a well-established area within econometrics. However, it’s application to banking regulation has only emerged in recent years.

The Bank for International Settlements (BIS) conducted a comprehensive assessment of macroprudential measures as part of the Central Bank Research Initiative (International Banking Research Network). As a result of its findings, both a detailed BIS (2020) report and a separate article by Gambacorta and Murcia (2020) were published. Under the initiative, the central banks of individual count­ries estimated the policy effects using a unified methodology. Subsequently, all results were consolidated. In particular, a meta-analysis was used, where the results of countries were weighted according to the volume of their banking assets. It is important that the performance of measures was understood as the degree to which tightening measures reduced the growth of lending rates; and easing measures, conversely, facilitated their growth.

Before the BIS report, there were studies by Bruno et al. (2017); Cerutti et al. (2017) that used the aforementioned methodology. Note that these papers contained results that the authors themselves referred to as counter-intuitive. For example, mathematical calculations show that measures restricting credit actually promote its growth. This is applicable for Thailand (BIS, 2020, p. 87), Colombia and Argentina (Gambacorta and Murcia, 2020, Figs 3–4). The authors do not explain this phenomenon­, focusing instead on the conclusions where the expected results were obtained. It can be assumed that the positive coefficient does not reflect the comprehensive effect of the measure. In terms of the difference-in-differences method, there is no ideal control group. In such a group, without tightening measures, the growth of lending could have been much more substantial. This means that the measure had its restraining effect. However, the mathematical model (regression) formally records that there was an increase in lending after the measure was implemented. Moving forward, it is important to remember the limitations of this approach. It does not have an ideal control group, so the estimates obtained may be biased.

We now focus on the substantive conclusions drawn from previous studies. In BCBS (2022), it is noted that Basel III measures are effective, as banks now have more capital and liquidity than before the adoption of Basel III.¹

Thus, there is experience in researching both micro- and macroprudential measures at the level of internationally recognized organizations, including the Basel Committee on Banking Supervision and the Bank for International Settlements. However, these studies have a fundamental limitation. Although in the country-specific articles and in the BIS (2020) report measures were divided into tightening and easing measures, in estimating the effects of bank regulation we implemented both for the first time relative to previous works (see Table 1 for details). We used data for both micro- and macroprudential measures; considered additional measures outside the central bank (government support measures) and used the unique characteristics of banks and their financial indicators as control variables.

Studies similar to the current paper include: Nier and Olafsson (2020), Lewrick et al. (2020), Couaillier et al. (2022), Wong et al. (2022), Dursun-de Neef et al. (2023), Avezum et al. (2021); see Table 2. They also examine supportive measures, but only during the pandemic. This study looks at the easing measures during the COVID-19 pandemic, as well as two additional episodes: in 2014 and 2022. A common feature of these studies is their emphasis on the effectiveness of the easing measures. Regarding the magnitude of the overall impact on lending, the literature indicates that credit growth is about 1–6% per year with 1 pp down in capital adequacy ratio premiums.

3.1.1. Dynamic panel data model

As shown in equation (1), the approach decomposes increases ∆Y_b,t into signals of measures $\sum _{m=1}^M\sum _{j=1}^k$ β_j × MaP_b,m,t–j. It is taken into account that the measure may produce delayed and time-stretched results. Therefore, j lags of measures are considered. It is assumed that a measure can be consolidated (summarizing the effects of all measures) or that m types of measures can be considered. Lending decisions are controlled by the financial and other characteristics of banks X_b,t–1 at the previous moment. Lag is taken to exclude endogeneity by design. This is done in the macroprudential measures study in the BIS (2020) report. In the study of microprudential measures by the Basel Committee in the BCBS (2022) report, banks’ financial indicators are not used. We do more than the Basel Committee did, and as a starting point, we consider the financial indicators of banks.

$\begin{array}{rl}\mathrm{\Delta }{Y}_{b,t}=a_0& +\sum _{j=1}^J{\gamma }_j\times \mathrm{\Delta }{Y}_{b,t-1}+\sum _{m=1}^M\sum _{j=0}^J{\beta }_{mj}\times Ma{P}_{b,m,t-j}+\\ & +\sum _{q=1}^Q{\vartheta }_q\times X_{q,b,t-1}+\sum _{p=1}^P{\theta }_p\times {\text{macro}}_{p,t-1}+{\epsilon }_{b,t}\end{array}$(1)

where ∆Y_b,t is the dependent variable — the increase per time period of the loan ­category of interest (taken as the difference of logarithms from the volume of loans on the balance sheet); MaP_b,m,t–j is a variable characterizing the intensity of the impact of the applied measure on bank b at the moment t − j. The baseline estimation option is expressed in percentage points of the capital adequacy ratio (% of risk-weighted assets, RWA). X_b,t–1 — control variables: financial indicators of the bank’s activities (SIZE, LIQ, CAP, DEP) and characteristics of banks (SIFI, GOV, FOR, IRB, SANC, noSWIFT, LicUniv, retail); macro_t — control macroeconomic variables (oil price on the world market; the Bank of Russia key rate); b — bank (when aggregating to quarters­); t — time step (we consider months and quarters); m — the type of the selected measure (for a consolidated measure (0) there is one variable, that is, M = 1; for all banks there is data on 9 measures, then M = 9); q — counter of control variables; p — counter of macro­economic indicators; J = 4 — the number of lags of the independent variable measure. The number was chosen to yield an annual estimate from quarterly data, in line with the studies of BIS (2020); Kozlovtceva et al. (2022); ϵ_b,t ~ iid represents an independent identically distributed component (random noise).

We analyze the following types of loans:

• • Lc — corporate loans (to legal entities, LE). The category includes loans to SMEs, individual entrepreneurs, and financial institutions (non-banks). The impact of changes in the ruble value of loans denominated in foreign currency (currency revaluation effect) has been excluded, as approximately 30% of loans to LEs are in foreign currency;
• • Lr — retail loans (to natural persons, FL); the exchange rate adjustment is not applied, as the proportion of retail loans in foreign currency is nearly zero. Additional consideration is given to the effects on two subcategories of retail loans: Lrc — unsecured consumer loans; Lrm — mortgage loans.

In the BIS approach, equation (1) employs the lagged dependent variable­ $\sum _{m=1}^M\sum _{j=1}^M$γ_j ∆Y_{i,b,t – j} among the explanatory variables. This requires us to use the Arellano–Bond approach to dynamic panel models. Estimates in such a model­ are significantly divergent. This is a typical indication of multicollinearity. Hence, the primary estimation is conducted using the fixed effects model from equation (2), without incorporating the lagged dependent variable.

$\begin{array}{rl}\mathrm{\Delta }{Y}_{b,t}=a_0& +\sum _{m=1}^M\sum _{j=0}^J{\beta }_{mj}\times Ma{P}_{b,m,t-j}+\sum _{q=1}^Q{\vartheta }_q\times X_{q,b,t-1}+\\ & +\sum _{q=1}^Q\sum _{m=1}^M\sum _{j=0}^J{\delta }_{qmj}\times Ma{P}_{b,m,tj}\times X_{q,b,t-1}+\\ & +\sum _{p=1}^P{\theta }_p\times {\text{macro}}_{p,t-1}+\sum _{b=1}^B{u}_b+{\epsilon }_{b,t}\end{array}$(2)

where u_b is a fixed effect (dummy variable) on bank b.

Interpreting the coefficients. To make conclusions about the effects of measures, we are primarily interested in the coefficients B_mi. The focus is on the sum of estimates of these coefficients SUM_MAP_m = $\sum _{j=1}^k\gamma j∆Y_{i,b,t\−j}$ β_mj per measure m. This sum reflects the measure effect over five cycles (five months or five quarters), as we take the measure for the current time step and for the four preceding lags (monthly or quarterly, depending on the data frequency).

A positive sum of coefficients is interpreted as a beneficial effect of anti-crisis measures. The larger the positive value of the measure, the greater the growth of lending in the examined segment per unit of the measure. Therefore, let the sum of coefficients be equal to +60, on average per step, loans will grow by 60/5 = 12 pp per 1 pp of the measure effect. If the measure effect on N1.0 is half and equals 0.5 pp, the growth of lending on the balance sheet will be on average 12 × 0.5 = 6 pp per step.

Periods under review. We estimate the models at four intervals (the designation used in the column headers of the coefficient estimate tables is given at the start of the row):

• • 14: January 2014–December 2016. For this period we have data on the application of three measures by individual systemically important financial institutions (SIFIs);
• • 20: March 2019–May 2021. The pandemic period; for this period we have data on 9 measures for all banks;
• • 22: June 2021–April 2023. This period also has data for all measures for all banks;
• • P (pooled) — combined dataset: January 2014–April 2023. During this period, tightening measures may have a greater impact (having been in effect for a longer number of cycles).

All intervals were selected to have approximately one year before the application of the first support measure for the episode under review. The boundary between the 2020 and 2022 intervals was chosen to be in the middle (June 2021). This choice allows us to consider information on banks’ actions before and after the implementation of measures, without overlapping intervals.

To eliminate changes in FX, we excluded the effect of currency revaluation in corporate loans.

3.2.1. Consideration of interaction terms

In the BIS (2020) report they introduce additional interaction terms between measures and control variables $\sum _{q=1}^Q\sum _{m=1}^M\sum _{j=0}^J$ δ_qmj × MaP_b,m,t–j × X_b,t–1.

At the first glance, the inclusion of interactions terms appears promising, as it can enhance the interpretability of the results. For example, this approach allows for conclusions such as: “when using measure M in a larger bank, the effect differs from the effect in a smaller bank.” Or like “the comprehensive effect on lending is larger in banks with a smaller capital buffer, as found in…” In their terms, the coefficient at CAP_MAP is negative. The lower the capital buffer CAP is, the greater the effect is. However, in this study, such an approach loses its interpretability. Banks had 9 measures at their disposal. BIS uses four basic financial characteristics of banks. Thus, it results in 9 × 4 = 36 variable interactions (interaction terms, cross-effects). Taking interaction terms into account is considered one of the robustness checks. For this, equation­ (1) is added with component $\sum _{q=1}^Q\sum _{m=1}^M\sum _{j=0}^J$ δ_qmj × MaP_b,m,t–j × X_b,t–1, and we estimate the parameters of Model 3.

$\begin{array}{rl}\mathrm{\Delta }{Y}_{b,t}=a_0& +\sum _{j=1}^J{\gamma }_j\times \mathrm{\Delta }{Y}_{b,t-j}+\sum _{m=1}^M\sum _{j=0}^J{\beta }_{mj}\times Ma{P}_{b,m,t-j}+\\ & +\sum _{q=1}^Q{\vartheta }_q\times X_{q,b,t-1}+\sum _{q=1}^Q\sum _{m=1}^M\sum _{j=0}^J{\delta }_{qmj}\times Ma{P}_{b,m,t-j}\times X_{q,b,t-1}\\ & +\sum _{p=1}^P{\theta }_p\times \text{macrop,}1-1+{\epsilon }_{b,t}\end{array}$(3)

When assessing the effects of bank regulation, it is common to consider the asymmetric effects of measures. It is expected that banks may respond more positively to easing than to tightening of regulation. For example, Levieuge and Sahuc (2020) found that GDP increased in a smaller proportion after a key rate cut than it decreased when the key rate is raised. To account for this disproportionate, or asymmetric response, measure values are divided into positive and negative. For this estimation, we use equation (4).

$\begin{array}{rl}\mathrm{\Delta }{Y}_{b,t}={\alpha }_0& +\sum _{m=1}^M\sum _{j=0}^J{\beta }_{mj}\times {\text{MaP}}_{b,m,t-j}^{+}+\sum _{m=1}^M\sum _{j=0}^J{\gamma }_{mj}\times {\text{MaP}}_{b,m,t-j}^{-}+\\ & +\sum _{p=1}^P{\theta }_p\times {\text{macro}}_{p,t-1}+\sum _{b=1}^B{u}_b+{ϵ}_{b,t}\end{array}$ (4)

where MaP⁺_b,m,t–j = max(MaP_b,m,t–j; 0) — positive change in the measure (in support of lending); MaP–_b,m,t–j = min(MaP_b,m,t–j; 0) — negative change in the measure (towards restricting lending).

In an asymmetric reaction, a positive effect can occur in two cases:

• • when a significant positive coefficient is obtained for a measure with a “plus” sign;
• • when a significant negative coefficient value is obtained for a measure with a “minus” sign.

Authors in BIS (2020), Gambacorta and Murcia (2020), and Kozlovtceva et al. (2022) referred to quarterly data for macroprudential measures, while Cerutti et al. (2017) — to annual data. To ensure comparability with the first set of studies, we also perform estimates of the baseline specification from equation (2) using quarterly data.

To carry out such an estimation, we adopt the approach from BIS (2020), ­specifically accounting for bank mergers. For this, data on over 100 bank mergers since 2014 have been compiled. It includes the license numbers of the banks (license BEFORE), which were consolidated with other banks (license AFTER). Given that mergers could have occurred in multiple phases, we also provide the license number of the bank (­license FINAL), to which the assets of the initial credit institution were ultimately transferred by 2023.

For banks, the capital adequacy ratio serves as a regulatory limit on lending volumes­. However, comparing two banks only based on the actual value of the ratio is not very informative. Banks can belong to different supervisory ­categories (for example, SIFI and others). Therefore, they may be subject to varying minimum capital adequacy requirements.

To account for the significance of the buffer (excess over the minimum) rather than just a requirement, alternative specifications are implemented with the ­replacement of the adequacy ratio by the amount of capital buffer over the known minimum requirement. The results will be comparable with Dursun‑de Neef et al. (2023), who also examined the effects of bank regulation in conjunction with the capital buffer.

If the assumption of independence of the application of treatment, of the randomness of selection in the experiment is violated, the treatment becomes endo­genous. Inverse probability weighting (IPW) and propensity score matching (PSM) methods are used to adjust for it.

Stata program has the commands to estimate the related models: Teffects; ­Eteffects; Etregress. The comparison of the commands is summarized in Table 3.

Endogeneity of the treatment selection exists in all three commands, but only two (Eteffects; Etregress) include dependent variables. In Eteffects, the difference in the dependent variable values between the two groups (pilot and control) is examined; whereas in Etregress, being part of the pilot group is one of the factors in the multivariate model for the dependent variable. Therefore, the Eteffects specification when calculating the Average Treatment Effect on Treated (ATET) estimate may encompass other factors that may not be considered by the pilot group sample. For our study, a potential factor could be the repercussions of implementing the Bank of Russia mandatory supportive measures.

Looking at the above-mentioned toolkit, we understand that the selection of measures by banks is a well-thought-out decision. Furthermore, the measures are implemented by those banks that are expected to incur substantial losses. For ­instance, a bank has foreign currency assets and liabilities in rubles. It can provide loans to exporters. Its borrowers find it advantageous to have a FX loan if their income is primarily in FX from overseas. In this case, the bank has a positive open currency position (OCP). If the ruble exchange rate strengthens, the bank’s assets decrease in ruble equivalent, resulting in losses for the bank. A bank with such an OCP would benefit from implementing a measure to suspend­ currency revaluation for the previous quarter. If the bank had a negative currency position, it would be unlikely to resort to using the measure under the same conditions.

The use of endogenous effects estimation models appears to be the most appropriate­ approach for evaluating the Bank of Russia supportive measures. However, there are two limitations. First, not all support measures are voluntarily adopted by banks. For instance, the dissolution of macroprudential buffers (measures 3–6) became mandatory for all. Formally, all banks became pilot banks, and the control group ceased to exist. Second, specifications with endogenous effects estimates do not allow for the separation of the effects of multiple measures applied at the same time. Moreover, valuable information about the heterogeneous effect of even the same measures on different banks remains neglected. Such a difference already suggests that decisions to change lending in these banks could vary. Therefore, the use of methods to estimate endogenous effects has been implemented as an alternative. The findings are discussed in Subsection 5.2. Banks are categorized into the pilot group if they have utilized at least one optional measure. If they have not used any measures or only mandatory ones, they are classified into the control group.

To obtain a more reliable in-sample forecast of the effects of measures on the banking system, we consider the results from multiple model specifications. Specifically, we use estimates on three episodes (2014, 2020, and 2022) and the combined array (P); for all banks and for systemically important ones. This approach aligns with the principle of constructing harmonic regressions defined in Ershov (2008). If the coefficients in different specifications have the same sign, the estimate is considered harmonic. Thus, it is important to consider a range of specifications, not limiting ourselves to the best one by a single criterion.

To obtain the forecast, we estimate the final models, calculate the sums of ­coefficients for each measure, accounting for relevant lags. Only statistically significant coefficients are retained. Results presented in the tables use significance levels of up to 10%. The recalculation of models can slightly alter the point estimates and the likelihood of accepting the null hypothesis of their (collective) equality to zero for individual coefficients. As a result, a coefficient might be significant at 9% at one point and 11% at another. Substantively, the coefficient estimate in both instances is essentially significant at 10%. However, if a strict significance criterion of 10% is applied, coefficients with p-values above this level would be excluded. Therefore, a 15% significance criterion was adopted to account for such a variability.

In the next stage, a threshold was imposed to the obtained estimates: only sums of coefficients with absolute value less than 1500 percentage points were retained for further analysis. This restriction was introduced to exclude estimates that, while statistically significant, are implausible in magnitude.

The significant sum of coefficients estimated for each measure was multiplied by the value of that measure for a given bank and its previous loan portfolio volume. For corporate loans, the calculations were based on values adjusted for currency revaluation using method 3. For each bank, we assessed the change in lending volume attributable to each significant measure. Notably, when a significantly negative coefficient is paired with a positive measure value (or vice versa), the resulting contribution to the loan growth could be negative. Thus, the model allows for both increases and decreases in a bank’s loan portfolio­. Using final model with monthly data for corporate and retail loans, we obtained estimates of the change in lending volume (in rubles) across four periods, covering nine measures for all banks.

After each final model was calculated, we summed up the portfolio changes for all banks on each date, that is, across the entire banking system. To obtain a reliable (less volatile, more conservative) estimate of system-level portfolio change, we averaged the estimates (calculated the simple arithmetic mean) of the total effects across the system between the model estimate on the combined array (P) for all banks, on the combined array for SIFIs, and the estimate for all banks in a specific subperiod. We added those changes together to get a calculation for all measures. We called the ratio of such a change to the portfolio size the contribution of the sum of all measures L_prop0.

Additionally, we calculated the growth rate of the total portfolio in the banking system, excluding currency revaluation. We denoted it as the natural growth rate of the portfolio d_L_prop. We set the start of the period as 1 January 2015, when the effect without support measures was zero. At that point, the portfolio with the banks’ application of measures was equal to the portfolio without the appli­cation of measures. Then we obtained a portfolio without applying measures $L_t^{NO\_M}$ for all subsequent dates t as the portfolio of the past step $L_{t-1}^{NO\_M}$, adjusted for the natural rate of portfolio growth and the effect of measures. To do this, we applied equation (5):

$L_t^{NO\_M}=L_{t-1^{-}}^{NO\_M}\times \left(1+d\_L\_\text{prop }\right)\times \left(1+L\_\text{prop }0\right).$ (5)

To construct a confidence interval (CI) of the forecast, we took into account the degree of variability in the growth rate of the loan portfolio d_L_prop after excluding currency revaluation. We calculated the increments of the value over the entire data range. We took the increment values corresponding to the extreme 5% of the distribution on both the left and right. In other words, we consider a 10% two-sided confidence interval. These increases apply to the portfolio size predicted within the sample without taking into account the effect of measures on $L_t^{NO\_M}$.

The forecast for the banking system was obtained as the sum of values for loans to businesses and households. The confidence interval of such a forecast was constructed according to the procedure described above, applied to the ­dynamics of the sum of loans to businesses and households.

Over the past 10 years, since 2014, the loan portfolio to both businesses and households has approximately doubled (see Fig. 1). However, its ratio to GDP remains at a comparable level of over 50% with a slight decreasing trend (see row 8 in Table 5). The total amount of loans to businesses and households in 2023 is approximately RUB 85 trillion in nominal terms. Loans to businesses make up about 40% of the assets of the entire Russian banking system; loans to households make up about 20%. During periods of instability, loans to businesses were reduced more significantly. There is a decrease in the share of business loans in foreign currency. After 2014, the households practically stopped borrowing, and banks stopped offering loans in currencies other than rubles.

Fig. 1.

Dynamics of dependent variables (loans on balance sheet) across the entire banking system.

Note: RW — risk-weight; CAR — capital adequacy ratio; FX — the change in the exchange rate; CPI — inflation (year-over-year); RHS — right-hand scale. Source: Authors’ calculations.

In Fig. 1, a steady decrease in the average estimated risk-weight (RW) can be observed. Notably, its increase (tightening) is evident after the RCAP programme of the Basel Committee’s review of domestic banking regulation in 2016, BCBS (2016), and after the pandemic in 2021. During periods of instability­, risk‑weights declined, especially in 2022–2023. This is one of the clearly ­observed direct easing effects of the Bank of Russia measures.

Over the period under review, there were two significant increases in the key rate to 17 and 20% in 2014 and 2022, and one significant decrease to 4.25% in 2020–2021.

For the most comprehensive control over the state of the macroeconomy, we additionally take into account the following factors in the form of a lag in the growth of the variables: FX — the change in the exchange rate; CPI — inflation (year-over-year).

Amid changes in the macroeconomic variables described above, the institutional environment for enterprises and banks also evolved. Due to the lack of detailed data on the direct effects of institutional shifts, we account for them indirectly through the introduction of the following dummy variables:

• • DIA is a generalized indicator that equals one from January 2015 and zero before that. Primarily, it takes into account the increase in the insurance coverage limit from RUB 700 thousand to RUB 1.4 million. Since 2015, the Bank of Russia has shifted to an inflation targeting (IT) and floating exchange rate policy. As these events (CER, IT, rate) occurred simultaneously, the variable serves to distinguish the post-2015 institutional framework from the pre-2015 regime, although their individual effects cannot be separately identified.
• • BudRule captures changes in Russia’s fiscal rule. Initially introduced in 2004, the budget rule underwent four revisions, with changes to the oil price threshold above which excess revenues were taxed—this includes periodic indexation of the cut-off price. Over the study period (starting in 2014), three versions of the rule were in effect, two of which were later suspended (Tsibanov, 2023, timestamp 17:13). These changes are represented using three separate dummy variables under the collective label BudRule.

As control variables, we consider standard financial indicators of banks, also used in BIS (2020); Gambacorta and Murcia (2020): bank size (SIZE), capital adequacy ratio (CAP), shares of liquid assets (LIQ) and deposits (DEP) relative to total assets. In general, the distributions of banks’ financial indicators in the sample are close to normal.

The capital reserve (buffer) is calculated by subtracting the known applicable minimum from the actual value of the total capital adequacy ratio (equity). Three regulatory thresholds are considered:

• • 8% — for basic license-holder banks;
• • 10.5% — for banks with a universal license that are not SIFIs. In comparison to banks with a basic license, they are required to maintain a capital conservation buffer (an additional charge for maintaining capital adequacy, CCB; conservation buffer) of 2.5% of RWA;
• • 11.5% — for SIFIs. In relation to banks with a universal license, there is an ­additional requirement to maintain a systemic importance surcharge of an extra 1% of the RWA.

It is important to note that the capital adequacy ratio (CAP) and the margin over it (K_buf) are highly correlated for all banks, as well as for SIFIs. Subtracting of the regulatory minimum (CAP _min) from the actual value (CAP) constitutes a level-shift and does not convey additional information. Therefore, in estimating effects and controlling for capital reserves, it does not matter whether we use the actual value of the standard or the margin over it in the control variables.

First we look at the easing measures that were implemented by the Bank of Russia. In 2020 and 2022, the data on the use of measures is comprehensive and detailed. In 2014, the information was only partially available, so it was reconstructed using the Heckman model, as detailed in Subsection 4.3.1.

When considering the measures and their relationship to banks’ decisions to change their lending strategy, it became clear that not only the absolute scale of the measures was important, but also their relative importance to the bank. To account for this, two alternative approaches to measuring relative importance are discussed in Subsection 4.3.2.

In this study, we consider the following groups of measures implemented by the Bank of Russia:

• • 9 easing measures from the Bank of Russia applicable to all banks. The greatest­ variation is observed in relaxations in credit risk assessment (Measure 8). It is important to note that measures can also have negative values. For example, actual stock and bond prices during a market recovery may exceed the values established by the measure.
• • 50 stages of tightening measures from the Bank of Russia, which were implemented between the periods of support measures. The multi­plicity of stages is apparent. Many of the stages refer to a single step. For example, after the Basel Committee’s review according to BCBS (2016), Russia began to actively implement Basel III regulations. In particular, they included a multi-stage implementation of capital surcharges (conservative, for SIFI).

Tightening measures include both micro- and macro-prudential components. The microprudential category includes changes in the levels of capital adequacy ratios and uniform risk weights. The macroprudential category pertains to consumer loans, mortgages, FX loans, and macro-additions for internal ratings-based (IRB) banks. Tightening measures vary by types of banking licenses and types of banks (SIFI, IRB) as shown in Fig. 2. Therefore, the value of tightening measures applicable to each bank was calculated. Since the effect of each individual measure was not specified, as with support measures, the fact (sign) of the tightening measure was taken into account. The sum of all tightening measures results in a conditional index. For clarity, it is displayed in Fig. 3.

Fig. 2.

Overview of the scale of the Bank of Russia anti-crisis measures (pp CAR (N1.0)).

Note: CAR — capital adequacy ratio; SIFI — systemically important financial institutions. Source: Authors’ calculations.

Fig. 3.

Indices of tightening measures of the Bank of Russia and easing measures of the Russian Government.

Source: Authors’ calculations.

This study departs from earlier works by Bruno et al. (2017); Gambacorta and Murcia (2020); Kim and Oh (2020), where regulatory measures were considered as the sum of the number of measures introduced, as an index of macro­prudential regulation (similar to Fig. 3). The prevalence of measure estimation in the form of an index (as a fact counter of introduction of different instruments) is due to the availability of the relevant IMF database for countries worldwide, proposed in Cerutti et al. (2017). However, as emphasized by a representative of the European Central Bank in Budnik (2020), in the early stages of assessing the effectiveness of bank regulatory measures, it is essential to consider not only the number of measures implemented, but — where possible and data permits — their intensity. This study represents the first attempt to conduct an analysis at the scale of the entire Russian banking system.

The primary focus of this research is on the Bank of Russia easing measures. Therefore, we use the fact of the application of the Bank of Russia easing measures for preliminary visual analysis. Final models are built considering the intensity of these supportive measures.

The dataset is limited by incomplete information on banks’ use of support measures in 2014 particularly for SIFIs. However, there are banks that either know exactly which measures they used and what effect they had or know which measures they used and which they definitely did not use. Along with this, it is known how these same banks used the measures in subsequent episodes of availability of the Bank of Russia easing measures. This situation gives rise to a classical formulation of the Heckman selection problem: for some banks, it is known that they participated in support programs, but the specific measures used — and their direct impact on capital — remain unobserved.

To address this, for each of the three measures available in 2014, the Heckman model is constructed, where the dependent variable is the direct effect of the measure in percentage points of the capital adequacy ratio. It takes on particular values for banks, where the information is available; or it remains empty — for banks with missing information for 2014. Model estimation is constructed for SIFI. Based on the resulting model, a forecast is made, which includes the retrospective decision to use the measure or not. When deciding to use, a forecast of the measure’s scale is given.

For each model, the correlation coefficient of the residuals from the selection and response equations is significant. This means that the Heckman model is preferable to estimating the two equations separately.

We use the resulting in-sample predictions for the three measures for SIFI for 2014 to supplement the available evidence in estimating the full effects of the measures.

The raw data for the measures are expressed in percentage points of the capital adequacy ratio. However, banks can possibly have identical direct effects of measures on capital, but with varying responses to them. It could result from the different significance of the same measure for each bank. Let us look at the example in Fig. 4.

Fig. 4.

Comparison of methods (options) or gauging measures.

Note: CAR (local N1.0) — capital adequacy ratio; buffer — capital ratio value above the minimum; for illustration assume the minimum equal to 8 pp. of RWA; d_buf — buffer change between periods; MaP — measure impact in terms of the CAR (N1.0). Our intent is to interpret the larger measure value as a more favorable contribution to a bank’s performance. Source: Authors’ calculations.

Banks A, B, C used the same measure. Its direct effect amounts to 3 pp. Let it be Method 1 of gauging the measure (MaP_t^I) . According to it, the measure is identical for all three banks. For bank B, the measure was more crucial than for A, as bank B was experiencing a capital deficiency (it did not meet the minimum capital adequacy requirements; for simplicity, we disregard other regulatory add‑ons­). Furthermore, after the measure was implemented, bank B came to meet the minimum requirement. Bank A consistently met the minimum requirements with a substantial margin and continued to do so. In reality, a Type A bank would be less inclined to resort to easing measures.

To account for the measure’s significance, we introduce a method for gauging the measure (MaP_t^II) . For this, we calculate the effect using equation (6):

$Ma{P}_t^{II}=\frac{Ma{P}_t^I}{K{B}_{t-1}},$ (6)

where MaP_t^I and MaP_t^II — the direct effect of the measure at time t, calculated using Methods 1 and 2 respectively; KB_t–1 = $CA{P}_{t-1}^{ACTUAL}-CA{P}_{t-1}^{MIN}$ — capital buffer (difference between the actual value of the ratio $CA{P}_{t-1}^{ACTUAL}$ and the minimum $CA{P}_{t-1}^{MIN}$) at time t – 1.

The negative sign in equation (6) is added before the fraction to preserve the interpretability of the measure coefficient’s sign. Specifically, we aim for a positive coefficient to indicate a supportive effect of the measure — that is, a positive contribution to credit growth. Therefore, the measure’s values should be positive when the measure is significant (for bank B). For bank A, where the measure is less significant (superfluous), the value of the measure by Method 2 will be negative. However, it should be noted that Method 2 doesn’t differentiate between banks B and C as shown in Fig. 4.

For the difference, we introduce Method 3 in equation (7):

$Ma{P}_t^{III}=\frac{Ma{P}_t^I}{d·K{B}_t},$ (7)

where MaP_t^I and MaP_t^III — direct effect of the measure at time t, gauged by Methods 1 and 3; d.KB_t = KB_t – KB_t–1 — the change in the reserve (buffer) of capital for one time step at moment t.

In Method 3, a larger value of the measure MaP_t^III reflects its greater significance, or the presence of additional limiting circumstances, due to which the bank was not able to increase the buffer to the same extent as the effect of the measure; or was able to use the effect of the measure to increase lending (for which, among other things, easing measures were developed by the Bank of Russia).

The proposed alternative methods of accounting for measures may have limitations. For instance, when there is a low capital reserve in the previous step or when there is minimal fluctuation in it. In such hypothetical scenarios, the fraction’s denominator will be nearly zero. In that case, the values of alternative benchmarks will approach infinity. Let us examine what we see in practice.

The measure values obtained through Method 1 (direct), after removing outliers, range from –12 to + 17% of risk-weighted assets. Method 2 yields values from –17 to +26% of the previous cycle’s capital reserve, while Method 3 results in values from –828 to + 1369% of the ratio change over the cycle. From this comparison, it becomes clear that Method 3 of measure gauging can generate substantial values, so it is advisable to avoid it if possible. Method 2 is on par with the first, both producing moderate values (within 25%, but from different bases). Therefore, in the study, Method 2 is preferable to use for measure gauging rather than Method 3.

Unlike the detailed information about the effects of the Bank of Russia easing measures, the data on the Russian Government measures are less comprehensive. Hence, a two-step procedure was implemented for their analysis: direct and indirect accounting of government programs.

First, models with two types of dependent variables were evaluated: the change in lending volumes as is and their change after excluding volumes issued under subsidized loan government programs. Each model was used to predict what the lending volume would have been without the Bank of Russia easing measures. The difference in predictions was attributed specifically to the effect of the Russian Government support measures.

To verify the stability of the results, two estimates of the volumes of the Russian Government subsidized loan program were considered:

• • Formal (average) estimate (1A in Fig. 5): loans are considered subsidized if they have a subsidized status in the bank as of the reporting date (the letter “T” in the loan description field in Form 0409303). By April 2023, subsidized loans to businesses of this character totaled RUB 6 trillion; to households — RUB 4.5 trillion.
• • Upper estimate (1B in Fig. 5): loans are considered subsidized if they were subsidized at least once. A loan might not have been subsidized at the start of the term, while the Federal Executive Branch (FEB) had not yet confirmed the borrower’s eligibility for the subsidy to the bank. A loan could lose its subsidized status if the borrower failed to meet one of the requirements (for example, maintaining employment numbers during the COVID-19 pandemic). However, all borrowers expected benefits when they applied for such loans. Therefore, this upper estimate of subsidized loan amounts is of interest. As of the most recent available date, the volume of such loans was over RUB 1 trillion­ beyond the average estimate.

The dynamics of total lending to businesses and households are depicted by solid lines in Fig. 5. Dotted lines represent lending estimates after deducting the average estimate of subsidized loan volume (1A for corporate loans) and the upper estimate (1B for corporate loans). It should be noted that data on subsidized loans to businesses is more reliable from May 2019 onward. Nevertheless, this allows us to incorporate such information when assessing the impact of the Russian Government’s measures on corporate lending in 2020 and 2022. Data on subsidized mortgages is available from 2021, so the effect of Government measures can only be reliably assessed for the 2022 period.

Fig. 5.

Lending dynamics: all loans and volumes of non-subsidized loans (RUB trillion).

Source: Authors’ calculations.

Second, even after directly excluding the volumes of loans recorded under the Russian Government’s subsidized programs, the specifications incorporate two variables representing the generalized indices of announced program volumes for businesses and households. The inclusion of variables is justified for two reasons. First, it enables the assessment of potential secondary (indirect, cross) supportive effects between the segments. For example, when business lending programs stimulated lending to households. Second, it serves as a robustness check: if direct program accounting is adequate, the coefficients of the introduced variables should prove to be insignificant. Otherwise, the presence of significant coefficients would indicate the existence of cross-supportive effects.

To construct additional variables, 17 episodes of government supportive programs were considered. These measures are categorized into two: those targeting businesses (PRAV 1) and those aimed at the public (PRAV 2). The measures are accounted for as an exponent of the volume of active programs at the date under consideration. It is important to note that declared (allocated) program volumes and actual selected program limits are not equivalent. In the absence of data on selected limits by bank — data which is available for Bank of Russia easing measures — the current approach represents the most feasible method of accounting. Based on these two measures, we can calculate the amounts, which are referred to as the Russian Government supportive measures (see Government measures index on Fig. 3).

For robustness checks, other methods of accounting for the tightening measures­ of the Bank of Russia and easing measures of the Russian Government were considered. They do not significantly alter the results.

It can be anticipated that the concurrent implementation of a set of anti-crisis measures by the Bank of Russia and the Russian Government yields a greater effect than implementing such measures individually. The conventional method to account for such Synergy (cross-, joint) effects is to include the interactions of the two variables responsible for each program individually in the specification. The rationale behind this approach was previously described in Subsection 3.2.1.

We attempt to assess the Synergy effect of the measures of the Bank of Russia and the Russian Government indirectly, comparing the estimates of the Bank of Russia measures obtained considering both aggregated data on the measures of the Russian Government and more detailed ones. The first (less detailed) data were only presented in the form of the index discussed in Subsection 4.4. The values varied over time but were uniform for all banks on each date. The result of their consideration was reflected in the draft report on the effectiveness of the Bank of Russia measures (Lymar and Penikas, 2023). The second (more detailed) data represents the volume of loans issued under subsidized loan programs. They vary not only over time but also across banks. More detailed data allows for a direct way of accounting for the Russian Government’s support measures, less detailed one — an indirect way.

Thus, the difference in the estimates of the effect of the Bank of Russia measures­ between the stage of drafting the report and this work can be considered a Synergy effect.

Atypical observations were excluded during data processing. Both univariate distributions of indicators and their relation to the dynamics of the dependent variable were considered.

Three categories of outliers were considered:

• • Dependent variables. In some banks, loans increased ten times from month to month. Therefore, all observations with credit growth exceeding 75% in absolute value were excluded as outliers.
• • Control variables. As a rule, observations exceeding the share values of 100% (loans, deposits, liquid assets) were considered to be outliers. One credit institution has been identified as an outlier in terms of the proportion of business and household deposits (criterion 6). One bank experienced unusual deposit proportions in the final time units, despite the deposit proportion typically ranging between 0 and 100% most of the time. Hence, all observations related to this bank were classified as outliers.
• • Values of the measures. The first lag values of the measures­ were analyzed in relation to the growth of corporate loans. Observations that significantly deviated from the majority of observations were highlighted with gray shading. However, the model from equation (2) incorporates not only the first lags but also the current measure values and lags up to the fourth time unit. Therefore, the current value and all lags with identified atypical values were also classified as outliers. According to these criteria, certain retail banks were excluded (some of which had substantial buffer values). To verify the robustness of the results, we have additionally calculated specifications for a subset of data from excluded observations, specifically, for individual retail banks.

In total, 3.5% of observations were selected based on outlier criteria, which is approximately 1.6 out of 45,000 bank-month data points since 2014.

Initially, we assess the potential effects of the Bank of Russia easing measures using equivalent to the conventional difference-in-differences method. The limitations in application of the conventional tools herein is that macro­prudential measures were applied to all banks. This eliminates the control sample but does not negate the fact that banks deliberately — not randomly, endogenously — chose whether or not to use the measure. This may have depended on both their initial capital and their credit growth appetite. We will discuss the attempt to account for this endogenous nature of the treatment application (use of measures) in Subsection 5.2, bearing in mind the described methodological limitations of such application in the absence of an actual control sample.

We then identify a quasi-control (control) group based on the measures that banks could select, that is, measures 1, 2, 7–9 (excluding measures 3–6). We refer to banks that used the measures as quasi-experimental (treated). We assess the trends in credit growth before and after the implementation of the measures. We also assess what all three episodes (one year before and after 2014, 2020, 2022) have in common. We defer all observations regarding the moment the first measure was used in the episode. Such a moment is to be numbered zero. Then, Fig. 6 displays an equivalent of the difference-in-differences method.

Fig. 6.

Preliminary effects for banks that used the Bank of Russia easing measures (treated) and others (control) by lending areas: corporate (Lc) and retail (Lr).

Source: Authors’ calculations.

Fig. 6 shows that in corporate lending, the lines (pre-trends) were similar for the two groups, but after the implementation of the measures, the growth in banks that used the measures was significantly higher, even though there was a decrease in growth on average in both groups at the start of the episodes. The situation is more illustrative in retail lending. The levels of trends changed diametrically when the slope angles were comparable before and after the start of the episodes. In the group of banks that used the Bank of Russia easing measures, the average lending growth became twice as high as in the control group. This comes as the first evidence that the Bank of Russia easing measures have significantly benefited domestic banks in supporting lending.

It is likely that banks selected which measures to implement based on the expected benefits, primarily the desire to free up capital to offset increased risks or to support and expand lending. However, such expectations cannot account for all potential economic scenarios under complex conditions. For example, the temporary suspension of currency revaluation in 2022 could have benefited banks with a short open currency position (OСP) in anticipation of further depreciation of the national currency. Conversely, this measure would be less attractive to banks with a long OCP and similar expectations regarding currency depreciation. Subsequently, the strengthening of the exchange rate prevented banks from profiting, regardless of their OCP position or the application of the revaluation suspension.

Therefore, it should be recognized that the bank’s choices of measure were not random, but were largely shaped by their expectations. However, non-random selection does not ensure that banks utilizing those measures were in a more advantageous position than others.

Accordingly, we conducted robustness checks using models that account for the endogenous selection of measures (i.e., models with endogenous treatment effects), as presented in Table 4. However, we acknowledge certain limitations: these models do not account for mandatory support measures implemented for all banks, which may bias the estimated effects presented in this section. Furthermore, current specifications do not permit the disentanglement of the effects of individual measures.

To determine whether banks were indeed selecting measures non-randomly, we examined the correlation coefficients between the response and selection equations. In two out of the three specifications for retail and corporate loans (see models 3, 5, 6 in column 4 of Table 4), these correlation coefficients are statistically significant for the entire sample. This finding suggests that non-random selection is primarily associated with specific episodes — namely, with measures supporting corporate lending in 2020 and retail lending in 2022. Therefore, it is accurate to conclude that, during these periods, banks sought to utilize available measures to the fullest extent.

In episodes characterized by persistently non-random selection, the factors influencing banks’ use of support measures were relatively consistent across different model specifications. Larger banks (as measured by SIZE), those with lower capital adequacy ratios (CAP), higher liquidity (LIQ), and greater deposit volumes (DEP) were more likely to adopt these measures.

Teffects models for retail and corporate loans do not include a condition for the dependent variable. Therefore, Eteffects and Etregress models are more preferable­ as they include this condition, allowing for a more accurate estimation of the treatment effect. The treatment effect estimation in the Etregress model differs from the corresponding estimation in the Eteffects model as it does not consider measures that are mandatory for all banks (measures 3, 4, 5, and 6). Hence, the difference in the treatment effect estimations of the two mentioned models can be interpreted as an indirect effect of mandatory measures.

For instance, as shown in Table 4, the coefficient estimate of –10.2 in row (3) for corporate loans suggests that banks’ voluntary measures reduced their credit growth by 10.2 pp. The estimate of –27.2 in row (2) for corporate loans is not significant. This suggests that considering mandatory measures eliminates the significant negative effect of those measures that banks choose on their own.

5.3.1. Assessments of the Bank of Russia measures effect on the banking system

Comprehensive effect of the measures on the system. This study revealed that the Bank of Russia easing measures yielded positive results during the periods of 2014, 2020, and 2022, as shown in Table 5.

The substantial magnitude of the effect is attributable to the supportive measures implemented by the Bank of Russia and the Russian Government, which helped to mitigate the impact of sanctions. Over time, however, the effectiveness of the Bank of Russia easing measures diminished. This outcome is consistent with the transition from using market-based benchmarks (mark-to-market) to model-based, or lagged, values (mark-to-model). Hence, the Bank of Russia announcement2 in May 2023 regarding the planned termination of easing measures from June 2023 appears justified. Notably, the actual lending volumes exceed the confidence interval of the predicted volumes in the absence of support measures. This suggests that the positive impact of the Bank of Russia easing­ measures is statistically significant across the entire ­banking system.

A comparison of the direct effect of the Bank of Russia supportive measures in rubles of released capital — with the comprehensive effect of the measures — measured in rubles of additional loans across the banking system — reveals a doubling of the multiplicative effect over the past decade. For example, in 2014, for RUB 0.2 trillion of the Bank of Russia measures effect on capital (see row 14 in Table 5), overall lending grew by RUB 0.3 trillion (see row 3), i.e. by 1.7 times (see row 15); in 2020, for approximately the same RUB 0.2 trillion of the system-wide capital effect, the increase in lending due to the Bank of Russia measures amounted to RUB 0.2 trillion, i.e. it was comparable to the direct effect on capital. However, in 2022, for RUB 1.2 trillion of the direct effect on capital, the final effect on the credit growth amounted to RUB 6.0 trillion, i.e. five times the direct effect on capital.

Therefore, the effect of 1 ruble of a bank’s freed capital due to the measures began to yield a three times greater effect in credit growth (from 1.7 to 5.0 times) from 2014 to 2022. This indicates that the effectiveness of the Bank of Russia easing measures has significantly improved over the past 10 years.

There is available data on the volumes of loans issued under Russian Government subsidized programs for both corporate and mortgage loans. Accordingly, the effects of the Russian Government measures were assessed for these two categories. The results are presented in Table 6. It should be noted that due to the limited historical data on the volume of subsidized loans (see Fig. 5) we can discuss the effect of the Russian Government measures only for the following episodes: corporate loans — for 2020, 2022; households loans — for 2022.

Table 6 presents six estimates of the effects of the measures of the Bank of Russia and the Government of Russia. They are derived from three options of excluding currency revaluation and two estimates of the volume of subsidized loans. Estimates of the effects of the Bank of Russia easing measures for 2022 range from RUB 4.15 trillion to RUB 9.00 trillion of additional loans granted due to these actions. For the sake of conservatism, we take the smallest estimate of the effect for further discussion, i.e. RUB 4.15 trillion. It is obtained by applying a lagged exchange rate (Method 1 of excluding currency revaluation) and by excluding reported volumes of subsidized loans (not through overhead estimation). Under these assumptions, the effect of the Russian Government measures on corporate loans is estimated nearly RUB 1 trillion.

For mortgage loans, only one estimate is available as currency revaluation is not excluded and there is a single volume estimate. Under these conditions, the effect of the Russian Government measures on mortgage loans is estimated at RUB 0.9 trillion for 2022.

Table 5 provides information on the effects of the Bank of Russia measures as a whole (row 3), which are based on this study and account for detailed data on the programs of the Russian Government, and on the preliminary effect of the Bank of Russia measures with undetailed data on government programs (row 4). The latter estimate was presented in a draft report (Lymar and Penikas, 2023, p. 6). The difference between these two values (row 5) can be regarded as a synergy effect of the support measures of the Bank of Russia and the Government of Russia. Notably, the combined and coordinated efforts, particularly in 2022, resulted in a significant positive impact on the financial resilience of domestic credit institutions, providing an additional RUB 1.7 trillion of lending to the economy.