A Structural Model for Mortgage Risk

40. The mortgage and real estate markets are a key source of risk for financial stability on the back of rising imbalances and increasing affordability risk. On the one hand, strong growth in real estate prices have resulted in imbalances on the residential real estate market, with an estimated accumulated growth rate of 150 percent in the investment-led segment and 75 percent for owner-occupied property over 2000: Q1–2018: Q3. On the other hand, there are signs of increased affordability risk with the share of new mortgages where imputed costs would exceed one-third of income at an interest rate of 5 percent at about 50 percent.

41. In addition, Swiss banks are heavily exposed to mortgage loans. About three quarters in Swiss banks’ lending portfolio are mortgage loans (excluding amounts due to banks and SFTs). This share increases for DFBs, reaching 95 percent for Raiffeisen and regional and savings banks, and 87 percent for cantonal banks. DFBs in turn account for 64 percent of the CHF 980 billion domestic mortgage market in Switzerland in 2018: Q1. By contrast, the weight of the mortgage portfolio for the Swiss legal entities of the G-SIBs is lower at 62 percent of total loans,²⁰ whereas for other banks, including private banks, mortgage loans account for 40 percent of the loan portfolio.

42. The limited data from actual tail events in Switzerland and structural changes in the mortgage market, makes statistical approaches to estimate mortgage loan loss events unreliable. While Switzerland experienced a surge in loss rates in mortgage claims in the early 1990s this event was triggered by default rates in commercial real estate rather than in the residential segment Also, the structure of the mortgage market has changed markedly since 2010 when borrowers were allowed to use contributions to retirement benefits in the 2^nd Pillar to cover part of the cash down payment to finance the loan.²¹

43. Statistical approaches are also hindered by the current benign cycle. Observed default rates in mortgage portfolios have been about 0.15 percent in the last three years which makes the statistical analysis unsuitable to deal with far tail events. At the same time, low observed default rates in benign times can be consistent with a risky portfolio under stress.

44. The IMF’s modeling approach projects the structure of the loan default process with estimates of behavioral and macroeconomic drivers linked to the scenario. The model builds on TUI, a structural model of the residential housing loan default process and incorporates idiosyncratic features of the Swiss mortgage market²² The model assumes that borrowers will continue to service a loan if they can afford to pay off the scheduled interest and principal payments, even when they have negative equity. The drivers include idiosyncratic demographic events as well as systematic factors generated by the scenario related to changes in house prices, unemployment rate, and mortgage interest rates.

45. The structural approach is calibrated on banks’ total mortgage portfolio. While the calibration is applied jointly to all portfolio segments in banks’ mortgage portfolio, the path for housing prices reflects the composition of the mortgage book across residential (owner-occupied and investment-led) and commercial real estate sectors. On average, the share of residential mortgage loans is about 80 percent (of which 50 percent for owner-occupied property), and 20 percent for commercial real estate.

46. The model assumes that default occurs if two conditions are satisfied:

• The borrower is in distress: he is unable to repay the debt on time due financial difficulties; and

• The net value of the collateral, after disposable costs, is less than the value of the loan: the borrower cannot sell the collateral to service the loan.

47. To capture Swiss regulatory rules, borrowers’ debt service ratio includes the amortization schedule as well as in interest payments. Swiss banks have the option to request a margin call if the value of the collateral (after a housing correction) is insufficient to meet self-regulation rules, tightening the debt servicing ratio. Specifically, the current self-regulation rules stipulate a mandatory amortization (second mortgage) to two-thirds of the value of the collateral at origination over a 15-year period. Before the revision in 2014, self-regulation stipulated that mortgages had to be amortized to two-thirds of the collateral value within 20 years.

48. Affordability risk under stressed conditions depends crucially on the mortgage vintage. Shocks to lending rates (interest payment), and the additional margin that Swiss banks might require if the Point-in-Time (PiT) LTV that exceeds two thirds of the outstanding value of the loan cannot be amortized over the remaining tenor of the mortgage, depend critically on the initial interest rate, housing price, LTV ratio, mortgage tenor, and self-regulation rules at origination.

49. The FSAP team reconstructed outstanding vintages using supervisory data. As vintage information was not available, the FSAP team used data on the average mortgage tenor, binding self-regulation rules, and data on inflow and outflow of mortgages to segment banks’ outstanding portfolio by vintage. The caveats of this approach relate primarily to the third step, as mortgage flows could be related to the refinancing of different tranches of the loan rather than to full amortization. Another caveat of this approach is that these data do not include material “indirect amortization” flows which are collected on a tax privileged Pillar 3a account pledged to the bank, but do not reduce the outstanding of the mortgage.

50. The probability that a borrower gets into distress (PSS) is specified as follows:

Borrower affordability rises with the following factors: (i) a demographic event D which impacts affordability provided the debt service ratio exceeds a given threshold; (ii) the tightening of income gearing ${\mathrm{\Delta }DSR}_t^{\gamma }$ which is linked to the interest path i_t, and the Point-in-time LTV ratio which depends on the outstanding value of the loan L_t and the real estate price P_t; and, (iii) a decrease in the borrower’s income proxied by the starting level of unemployment u_t, and a non-linear function of the shock to unemployment Δu_t, with the sensitivity parameter β₃ increasing with ${DSR}_t^{\gamma }$. The model has been calibrated initially on data from New Zealand supported by an estimation approach that replicates the observed loans loss rates in Switzerland in the early 1990s.

51. Default occurs when the borrower is in distress and the current value of the house net of transaction costs is lower than the value of the loan:

Where the first term PSS_t, denotes the borrower’s capacity to service the loan, and the second term captures the condition that the net value of the collateral ${\tilde{P}}_t-C$, where ${\tilde{P}}_t$ is the market value of the house and Cthe transaction costs, is lower than the value of the loan at time t The value of the loan includes the outstanding principal L (amortized at an annual rate ρ over t periods) and the present value of future interest payments until the maturity of the loan at time T discounted at market rate i_t+j. The default rate is generated using a draw from the stochastic house price index with mean linked to the scenario and repeated 10,000 times to generate an array of iterations.

52. The conditional LGD is driven by the discounted sale price of the house. The model assumes that the sale occurs at time t+s (where s denotes the time to sell the collateral currently calibrated at 2 years) and the sale proceeds are net of transaction costs discounted at a rate reflecting the risk premium of the foreclosed asset (linked to the spreads projected in the scenario). For each draw of the house prices used to determine the default rate (systematic risk), the model generates 2,000 draws to generate individual house prices (idiosyncratic risk). For each of these house prices the model determines whether the value of the loan is greater than the net proceeds of selling the house. The average of the 2,000 losses is the LGD for the risk bucket.

53. Bank-specific forecasts for stressed PDs and LGDs are built by risk bucket over the horizon period. The model calculates PDs and LGDs for specified risk buckets defined by LTV tranche and vintage. Outputs for a given portfolio are calculated by combining separately estimated outputs of all the portfolio’s risk buckets and weighting them by their share in the bank’s portfolio. The data for the initial distribution of outstanding mortgages across risk buckets are sourced from the BBA templates for large banks are applied to the remaining banks of the sample. The model acknowledges that loan losses take time to realize and projects cumulative forecasts of PDs and LGDs by bank over 2018–20 and 2020–2023. It generates annualized rates by dividing accumulated losses over the number of years.

54. A key component to the implementation of IMF’s model-based credit risk assessment is model validation. The aim is to ensure that the model chosen is accurate and perform consistently. This is performed in two stages. First, the model is subject to stress testing under baseline and adverse conditions, examining the model’s expected outcome under normal and stress conditions. Second, the model is subject to backtesting. Observed outcomes in a relevant historical period are compared against the model’s expected outcomes. This process enables forecast evaluation of the model.

55. The model was simulated to a replication of ‘benign’ conditions, ‘good time’ conditions, and ‘bad time’ conditions (Figure 7). The ‘benign’ conditions’ simulation is captured by the baseline scenario with a gradual improvement of macroeconomic conditions; the ‘good time’ scenario replicates the improvement in loan loss rates observed in Switzerland in 1999; and the ‘bad time’ scenario reproduces the Swiss housing mortgage default experience of the early 1990s. A caveat of this approach is that there is no data on the historical structure of banks’ mortgage portfolio by LTV/LTI bucket. The working assumption is that it replicates the current distribution.

Switzerland: Credit Risk for Housing Loans

Citation: IMF Staff Country Reports 2019, 189; 10.5089/9781498321808.002.A001

Source: IMF staff estimates.

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Switzerland: Credit Risk for Housing Loans

Citation: IMF Staff Country Reports 2019, 189; 10.5089/9781498321808.002.A001

Source: IMF staff estimates.

• Download Figure
• Download figure as PowerPoint slide

Switzerland: Credit Risk for Housing Loans

Citation: IMF Staff Country Reports 2019, 189; 10.5089/9781498321808.002.A001

Source: IMF staff estimates.

• Download Figure
• Download figure as PowerPoint slide

56. Under current ‘benign’ conditions, the expected loss rate of the portfolio reaches 0.13 percent. The model is simulated under the baseline scenario whereby unemployment rate improves by 10 basis points to 2.6 percent, mortgage rates increase gradually from 1.5 percent to 2.8 percent over two years, and house prices remain stable. The 0.13 percent expected loss rate for the average bank mortgage portfolio compares against the latest observed default rate 0.014 percent which is supported by the extraordinary low interest rate environment.

57. For the ‘good time’ conditions observed during the late 1990s’ housing market recovery, the model replicates accurately the observed default rate. After the housing crash and sharp increase in mortgage rates in the early 1990s, the observed loan loss rate in Switzerland halved over 1997–99 from 1.32 percent to 0.61 percent. The model simulates the macroeconomic conditions prevailing in 1997 and produces a forecast for 1999 whereby unemployment rate decreased by 250 basis points to 2.7 percent, interest rates contracted by 50 basis points to 3.9 percent, and housing prices remained stable at about -0.3 percent. Applying the LGD projected by the model on observed loan losses in 1999 yields a default rate of 1.30 percent against the default rate simulated by the model of 1.28 percent.

58. The model is able to reproduce the Switzerland mortgage default event of the early 1990s which experienced the largest relative increase in default rates. Loss rates on mortgage claims doubled in 1991 from 0.53 percent in 1989 to 1.03 percent. The model is fed with the inputs as of 1989 and simulates the macroeconomic recession of 1991 including a 50 basis points increase in the unemployment rate to 1.1 percent, the 180 basis points increase in mortgage rates to 7.8 percent, and the sharp fall in real estate of 20 percent.²³ Using the LGD estimated by the model on observed loan losses in 1991 delivers a observed default rate of 1.90 percent with matches the 1.91 percent default rate projected by the model.

59. The back-testing exercise suggests that the model is a useful diagnostic tool, however, results are subject to caveats. The caveats of this approach relate primarily to assumptions over time to repricing of outstanding mortgages, the construction of vintages using inflow/outflow data, and assumptions on the LTV/LTI distribution of historical vintages. At the same time, this model has clear strengths related to the transparency, precision, and flexibility of the tool which outperforms statistical analysis which depend critically on the estimation period, are limited to deal with far tail events, and apply judgments related to the application of past experiences to the future despite structural changes in the mortgage market.

60. Results suggest that default rates are particularly sensitive to lending rates, margin calls, and the vintage distribution. The adverse scenario includes rises to mortgage lending rates from 1.5 to 5 percent on the back of a widening of credit risk premia.²⁴ The vintage of the portfolio is constructed using the self-regulation rules which after 2014 stipulate a maximum amortization period of 15 years for the second mortgage. Default rates increase to 4.9 percent if the portfolio is more heavily weighted to more recent originations under a 5-year amortization rule. Similarly, default rates increase to 6.2 percent under a 470 basis points increase in mortgage rates.

An Econometric Approach for Other Credit Exposures

61. Default risk in retail unsecured and corporate exposures by geography and asset class was proxied by the average one-year EDF extracted from Moody’s Analytics. For each geography, the following portfolios were used: the corporate group, the construction and real estate development group, the financials group, and the consumer nondurables and services group. These categories were mapped to Basel credit portfolios, i.e., corporate and small- and medium-sized enterprise (SME) exposures, specialized lending exposures, exposures to financial institutions, and retail unsecured, respectively. The econometric analysis was conducting using quarterly data over 2002: Q2 through 2018: Q4. Bank specific PDs were generated by attaching top-down parameter paths to bank starting points with a locational and portfolio perspective.²⁵

62. To address the truncated distribution nature of default rates, a logit transformation was applied before conducting the analysis. The logit transformation addressed biases generated by the truncated distribution of default rates and ensures that the projected rate is contained within the 0–1 bound once the logit forward path is applied on the forecast.

63. Credit risk was linked to a set of local drivers, regional variables, and global factors. Arguably, credit risk, particularly for wholesale clients, is not only a country-specific type of risk but it is also driven by regional macroeconomic forces external to the country as well as by global funding conditions. To estimate credit risk, a statistical-type model was developed by country and portfolio level using the following set of drivers:

• Local variables: there are a number of macroeconomic and financial variables that can impact the risk profile of a credit portfolio including real GDP growth, a GDP-based recession indicator (to capture non-linear effects from negative rates), inflation, unemployment, output gap, credit growth, equity prices, housing prices, FX rates, 3-month T-bill, 10-year yield, 3-month libor rate, the slope of the yield curve, real long-term rate, the swap rate, and the spread of the 10-year yield over the swap rate. Local drivers were forecasted for the twelve geographies included in the geographical breakdown of the lending portfolio;

• Regional variables to capture information about developments in the euro area, and emerging markets. Regional factors include GDP growth and inflation in the euro area and emerging economies, developments in the euro swap curve, 3-month euro repo rates, 3-month euro Libor, European corporate spreads, Emerging market bond index spreads, Eurostoxx 50, and structural challenges from debt sustainability concerns (measured by sovereign debt spreads of the Italian 10-year yield over the Bund);

• Global financial market variables: the ability of wholesale clients to repay also depends on the state of the global economy; in addition, shifts to global liquidity are likely to trigger default risk shocks. Global variables include world GDP growth, developments in commodity prices (fuel, non-fuel), changes in 3-month U.S. dollar Libor, TED spread, changes in U.S. investment-grade (IG) and high-yield (HY) corporate bond spreads, volatility index (VIX), and U.S. equity prices.

64. Given the uncertainty associated to tail risk predictions, a robust econometric framework was implemented using a variety of approaches:

• A time series econometric technique using Newey-West HAC-robust standard errors to obtain consistent estimators when the error term is heteroskedastic, autocorrelated, or both, once the regressors are proved stationary;

• A quantile regression approach to address the concern that the drivers of the conditional mean might be different in the higher tail of the credit risk distribution. The distribution was divided in quartiles (four segments). Robustness of the results was checked using a distribution based on deciles (ten segments) to explore cliff-effects far in the tail;

65. The quantile approach was used to benchmark the econometric model against severe credit risk in the upper quantiles of the default rate distribution. The empirical specification of the PD level of portfolio j in geography i is as follows:

where $Z_{t-1}^i\cdot$ is the set of lagged determinants including the country, regional, and global factors specified above, and λ⋅ denotes the quantile of the conditional distribution.

66. The Newey-West HAC time series approach suggests that credit risk in domestic non-financial corporate exposures increases with deflationary pressures, a widening in the output gap, and the tightening of financial conditions. Table 3 shows that low inflation, negative output gap, a dummy variable capturing negative economic growth, contribute to a rise in default rates. A tightening of financial conditions reflecting low equity prices, low fuel prices, rises in the T-bill, a steepening of the yield curve, and sovereign stress in the Euro Area also contribute to a worsening in the outlook of Swiss corporates.

Table 3.

Switzerland: EDF Projections for Non-Financial Corporates

Standard errors in parenthesesSources: Moody’s KMV, WEO, Bloomberg, and IMF estimates.Note: The estimation period is 2002: Q2 through 2018: Q4. The dependent variable is Moody’s 1-year average EDF for the corporate portfolio (i.e,. “1-Yr EDF 9 Average”) in Switzerland. Regressors include: inf (inflation rate), g_nom (nominal GDP rate), gap (output gap), tb (3-month T-bill), ltr (10-year real interest rate), slope (10-year over 3-month sovereign yield), credit (credit growth to the private sector), equity (equity price growth), g_w (global GDP growth), g_oil (spot fuel price growth), spread_IT (10-year Italy sovereign minus 10-year Bund yield), and stress_g#cL.g (a dummy variable taking the lagged GDP rate when it’s negative). Three specifications are shown for each corporate portfolio for the full sample period (66 observations) and a truncated sample (59 observations) to conduct out-of-sample forecasting. L denotes 1 lag, L2 denotes 2 lags and L3 denotes 3 lags. The model is used for forecasting purposes rather than for the structural identification of shocks. Therefore, the forecasting ability of the specification is the relevant metric for model performance rather than the economic interpretation and magnitude of the drivers.

Table 3.

Switzerland: EDF Projections for Non-Financial Corporates

	(1)	(2)	(3)	(4)	(5)	(6)
VARIABLES	1	2	3	1	2	3
L.inf	0.010 (0.149)	0.157 (0.130)	-0.073 (0.216)	-0.019 (0.144)	0.044 (0.161)	-0.090 (0.170)
L2.inf		0.150 (0.169)	0.146 (0.237)		0.184 (0.176)	0.064 (0.358)
L.g_nom	-0.136 (0.083)	-0.026 (0.122)	-0.265* (0.130)	-0.200** (0.086)	-0.127 (0.183)	-0.401** (0.183)
L2.g_nom		-0.318* (0.183)	0.203 (0.175)		-0.381 (0.236)	0.106 (0.248)
L.gap	-0.038 (0.068)	-1.878** (0.821)	-0.719 (1.183)	-0.008 (0.072)	-2.764*** (0.838)	0.216 (1.583)
L.tb	0.016 (0.124)	0.563* (0.324)	0.574 (0.349)	0.044 (0.119)	0.899* (0.453)	0.823 (0.467)
L.ltr	0.040 (0.152)	-0.251 (0.374)	-0.980 (0.594)	0.062 (0.176)	-0.284 (0.358)	-0.813 (0.497)
L2.ltr		0.030 (0.251)	0.033 (0.363)		-0.474 (0.368)	-0.356 (0.539)
L.slope	0.408** (0.154)	0.592 (0.371)	0.784* (0.449)	0.393** (0.159)	0.594 (0.444)	0.853 (0.494)
L.credit	-0.015 (0.040)	0.004 (0.045)	0.027 (0.044)	-0.018 (0.045)	0.004 (0.051)	0.036 (0.057)
L.equity	-0.011** (0.004)	-0.012 (0.007)	0.005 (0.013)	-0.008 (0.006)	-0.006 (0.007)	0.008 (0.009)
L2.equity		-0.001 (0.009)	-0.008 (0.006)		0.014 (0.011)	0.005 (0.013)
L.g_w	-0.015 (0.077)	0.200 (0.130)	-0.316* (0.168)	0.011 (0.097)	0.440** (0.162)	-0.162 (0.226)
L.g_oil	0.003 (0.002)	0.001 (0.003)	0.003 (0.005)	0.004 (0.003)	0.008* (0.004)	0.010* (0.005)
L2.g_oil		-0.010** (0.004)	-0.013*** (0.004)		-0.013** (0.005)	-0.011*** (0.004)
L.spread_IT	0.173*** (0.045)	0.500*** (0.152)	0.546* (0.271)	0.159*** (0.050)	0.567** (0.211)	0.561* (0.306)
l.stress_g#cL.g	0.095 (0.074)	0.171 (0.123)	0.229 (0.152)	0.122* (0.072)	0.239* (0.125)	0.173 (0.145)
Constant	-4.877*** (0.423)	-6.867*** (0.837)	-8.411*** (1.395)	-4.974*** (0.413)	-6.633*** (0.871)	-7.165*** (1.796)
Observations	66	66	66	59	59	59

Standard errors in parenthesesSources: Moody’s KMV, WEO, Bloomberg, and IMF estimates.Note: The estimation period is 2002: Q2 through 2018: Q4. The dependent variable is Moody’s 1-year average EDF for the corporate portfolio (i.e,. “1-Yr EDF 9 Average”) in Switzerland. Regressors include: inf (inflation rate), g_nom (nominal GDP rate), gap (output gap), tb (3-month T-bill), ltr (10-year real interest rate), slope (10-year over 3-month sovereign yield), credit (credit growth to the private sector), equity (equity price growth), g_w (global GDP growth), g_oil (spot fuel price growth), spread_IT (10-year Italy sovereign minus 10-year Bund yield), and stress_g#cL.g (a dummy variable taking the lagged GDP rate when it’s negative). Three specifications are shown for each corporate portfolio for the full sample period (66 observations) and a truncated sample (59 observations) to conduct out-of-sample forecasting. L denotes 1 lag, L2 denotes 2 lags and L3 denotes 3 lags. The model is used for forecasting purposes rather than for the structural identification of shocks. Therefore, the forecasting ability of the specification is the relevant metric for model performance rather than the economic interpretation and magnitude of the drivers.

View Table

Table 3.

Switzerland: EDF Projections for Non-Financial Corporates

	(1)	(2)	(3)	(4)	(5)	(6)
VARIABLES	1	2	3	1	2	3
L.inf	0.010 (0.149)	0.157 (0.130)	-0.073 (0.216)	-0.019 (0.144)	0.044 (0.161)	-0.090 (0.170)
L2.inf		0.150 (0.169)	0.146 (0.237)		0.184 (0.176)	0.064 (0.358)
L.g_nom	-0.136 (0.083)	-0.026 (0.122)	-0.265* (0.130)	-0.200** (0.086)	-0.127 (0.183)	-0.401** (0.183)
L2.g_nom		-0.318* (0.183)	0.203 (0.175)		-0.381 (0.236)	0.106 (0.248)
L.gap	-0.038 (0.068)	-1.878** (0.821)	-0.719 (1.183)	-0.008 (0.072)	-2.764*** (0.838)	0.216 (1.583)
L.tb	0.016 (0.124)	0.563* (0.324)	0.574 (0.349)	0.044 (0.119)	0.899* (0.453)	0.823 (0.467)
L.ltr	0.040 (0.152)	-0.251 (0.374)	-0.980 (0.594)	0.062 (0.176)	-0.284 (0.358)	-0.813 (0.497)
L2.ltr		0.030 (0.251)	0.033 (0.363)		-0.474 (0.368)	-0.356 (0.539)
L.slope	0.408** (0.154)	0.592 (0.371)	0.784* (0.449)	0.393** (0.159)	0.594 (0.444)	0.853 (0.494)
L.credit	-0.015 (0.040)	0.004 (0.045)	0.027 (0.044)	-0.018 (0.045)	0.004 (0.051)	0.036 (0.057)
L.equity	-0.011** (0.004)	-0.012 (0.007)	0.005 (0.013)	-0.008 (0.006)	-0.006 (0.007)	0.008 (0.009)
L2.equity		-0.001 (0.009)	-0.008 (0.006)		0.014 (0.011)	0.005 (0.013)
L.g_w	-0.015 (0.077)	0.200 (0.130)	-0.316* (0.168)	0.011 (0.097)	0.440** (0.162)	-0.162 (0.226)
L.g_oil	0.003 (0.002)	0.001 (0.003)	0.003 (0.005)	0.004 (0.003)	0.008* (0.004)	0.010* (0.005)
L2.g_oil		-0.010** (0.004)	-0.013*** (0.004)		-0.013** (0.005)	-0.011*** (0.004)
L.spread_IT	0.173*** (0.045)	0.500*** (0.152)	0.546* (0.271)	0.159*** (0.050)	0.567** (0.211)	0.561* (0.306)
l.stress_g#cL.g	0.095 (0.074)	0.171 (0.123)	0.229 (0.152)	0.122* (0.072)	0.239* (0.125)	0.173 (0.145)
Constant	-4.877*** (0.423)	-6.867*** (0.837)	-8.411*** (1.395)	-4.974*** (0.413)	-6.633*** (0.871)	-7.165*** (1.796)
Observations	66	66	66	59	59	59

Standard errors in parenthesesSources: Moody’s KMV, WEO, Bloomberg, and IMF estimates.Note: The estimation period is 2002: Q2 through 2018: Q4. The dependent variable is Moody’s 1-year average EDF for the corporate portfolio (i.e,. “1-Yr EDF 9 Average”) in Switzerland. Regressors include: inf (inflation rate), g_nom (nominal GDP rate), gap (output gap), tb (3-month T-bill), ltr (10-year real interest rate), slope (10-year over 3-month sovereign yield), credit (credit growth to the private sector), equity (equity price growth), g_w (global GDP growth), g_oil (spot fuel price growth), spread_IT (10-year Italy sovereign minus 10-year Bund yield), and stress_g#cL.g (a dummy variable taking the lagged GDP rate when it’s negative). Three specifications are shown for each corporate portfolio for the full sample period (66 observations) and a truncated sample (59 observations) to conduct out-of-sample forecasting. L denotes 1 lag, L2 denotes 2 lags and L3 denotes 3 lags. The model is used for forecasting purposes rather than for the structural identification of shocks. Therefore, the forecasting ability of the specification is the relevant metric for model performance rather than the economic interpretation and magnitude of the drivers.

67. The determinants of credit risk in foreign exposures differ across geographies. Appendices II and III show key drivers for corporate default rates in the United Kingdom and the United States. By contrast to the determinants of corporate default risk in Switzerland, a rise in inflation, lower credit growth, a slowdown in global growth, and a widening on money market spreads in the U.S. are key drivers of corporate credit risk for U.K. firms (Appendix II). In the United States, a flattening of the yield curve, a rise in long-term real interest rates and a slowdown in emerging markets also contribute to corporate distress (Appendix III).

68. Quantile results suggest that credit risk projections are more severe using estimated coefficients from the upper quantiles of the EDF distribution. Figure 8 shows quantile regression projections for default risk in corporate exposures in Switzerland. Under the baseline scenario, credit risk increases from the 0.9 percent value observed in 2018 to 1.6 percent (50th percentile) and jumps to 2.6 percent (90th percentile). Under the adverse scenario, EDF increases to 2.8 percent (30th percentile) and to a peak of 7.4 percent in 2020 under the 80th percentile.²⁶ The chart also reports the estimates across quantiles benchmarked against a linear ordinary least squares-based specification. The impact is non-linear across quantiles, albeit with wide confidence intervals, suggesting cliff effects from a worsening of underlying risk factors.

Switzerland: EDF Corporate Exposures, Quantile Approach

Citation: IMF Staff Country Reports 2019, 189; 10.5089/9781498321808.002.A001

Source: IMF estimates. The panel shows the estimated coefficients across quantiles for key drivers. Estimated coefficients using quantile regressions are shown against linear-based coefficients including confidence intervals. Regressors include: g (growth rate), inf (inflation rate), g_nom (nominal GDP rate), gap (output gap), tb (3-month T-bill), ltr (10-year real interest rate), slope (10-year over 3-month sovereign yield), spread_swap (10-year sovereign yield minus 10-year swap rate), credit (credit growth to the private sector), equity (equity price growth), g_w (global GDP growth), g_em (GDP growth of emerging markets), g_oil (spot fuel price growth), libor_usd_tbill (3m libor USD minus 3m U.S. T-bill), spread_IT (10-year Italy sovereign minus 10-year Bund yield), and stress_g_l_g (a dummy variable taking the lagged GDP rate when it’s negative). Lagged variables are preceded by “l”.

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Switzerland: EDF Corporate Exposures, Quantile Approach

Citation: IMF Staff Country Reports 2019, 189; 10.5089/9781498321808.002.A001

Source: IMF estimates. The panel shows the estimated coefficients across quantiles for key drivers. Estimated coefficients using quantile regressions are shown against linear-based coefficients including confidence intervals. Regressors include: g (growth rate), inf (inflation rate), g_nom (nominal GDP rate), gap (output gap), tb (3-month T-bill), ltr (10-year real interest rate), slope (10-year over 3-month sovereign yield), spread_swap (10-year sovereign yield minus 10-year swap rate), credit (credit growth to the private sector), equity (equity price growth), g_w (global GDP growth), g_em (GDP growth of emerging markets), g_oil (spot fuel price growth), libor_usd_tbill (3m libor USD minus 3m U.S. T-bill), spread_IT (10-year Italy sovereign minus 10-year Bund yield), and stress_g_l_g (a dummy variable taking the lagged GDP rate when it’s negative). Lagged variables are preceded by “l”.

• Download Figure
• Download figure as PowerPoint slide

Switzerland: EDF Corporate Exposures, Quantile Approach

Citation: IMF Staff Country Reports 2019, 189; 10.5089/9781498321808.002.A001

Source: IMF estimates. The panel shows the estimated coefficients across quantiles for key drivers. Estimated coefficients using quantile regressions are shown against linear-based coefficients including confidence intervals. Regressors include: g (growth rate), inf (inflation rate), g_nom (nominal GDP rate), gap (output gap), tb (3-month T-bill), ltr (10-year real interest rate), slope (10-year over 3-month sovereign yield), spread_swap (10-year sovereign yield minus 10-year swap rate), credit (credit growth to the private sector), equity (equity price growth), g_w (global GDP growth), g_em (GDP growth of emerging markets), g_oil (spot fuel price growth), libor_usd_tbill (3m libor USD minus 3m U.S. T-bill), spread_IT (10-year Italy sovereign minus 10-year Bund yield), and stress_g_l_g (a dummy variable taking the lagged GDP rate when it’s negative). Lagged variables are preceded by “l”.

• Download Figure
• Download figure as PowerPoint slide

69. The final model selection was based on a combination of criteria. This included goodness of fit of the regression, forecasting performance in-sample (2002: Q–2018: Q4) and out-of-sample (2016: Q4–2018: Q4), and expert judgment applied over the forecasted paths under baseline and adverse conditions based on the behavior observed during the 2008–09 global financial crisis. Estimated projections were applied over the forecast window 2019: Q1–2023: Q4 under each scenario.

A Merton-based Approach for Sovereign Exposures

70. The FSAP team constructed a database of sovereign exposures by counterparty mapping SNB surveys, Basel data, and banks’ annual reports. The SNB ‘Debt Securities’ template was used to estimate the country shares of fixed income securities. The country breakdown was applied to the amount of IRB and STA sovereign and non-financial institutions’ exposures reported in Basel templates.²⁷ As the ‘Debt Securities’ breakdown for foreign exposures does not include the breakdown by counterparty (counterparties are only available for domestic debt securities), the FSAP team extracted counterparty data from banks’ annual reports and Pillar 3 to identify foreign sovereign and corporate debt securities.²⁸ Under the adverse scenario, the slope of the yield curve increase for most bond curves due to term premium (Table 4).

Table 4.

Switzerland: Slope Yield Curve Projections

While short-term yields decline in the adverse scenario as monetary policy reacts to the global recession, long-term yields increase due to the widening of credit spreads, leading to the steepening of the yield curve.

Source: Bloomberg, WEO, IMF Staff estimates. The slope is measured as the 10-year sovereign yield over the 3-month Treasury bill.

Table 4.

Switzerland: Slope Yield Curve Projections

While short-term yields decline in the adverse scenario as monetary policy reacts to the global recession, long-term yields increase due to the widening of credit spreads, leading to the steepening of the yield curve.

	Observed	Baseline					Adverse
	2018	2019	2020	2021	2022	2023	2019	2020	2021	2022	2023
Switzerland	96	105	108	116	123	130	320	388	352	280	234
United States	49	24	‐5	17	39	44	248	286	264	205	156
United Kingdom	84	68	59	55	50	44	290	349	301	215	154
China	69	69	69	69	69	69	217	277	236	169	130
Germany	130	150	159	163	163	168	324	410	378	304	258
Japan	33	40	46	68	77	87	131	89	83	79	84
Australia	106	91	86	81	64	64	319	380	331	234	181
Spain	225	256	276	292	314	319	580	654	606	531	469
Korea	118	118	118	118	118	118	348	416	363	279	224
Hong Kong	69	69	69	69	69	69	217	277	236	169	130
Italy	248	286	309	327	320	317	611	688	642	538	467
Netherlands	113	113	113	113	113	113	287	364	328	253	203
Ireland	168	182	197	210	221	229	357	449	426	361	319
Luxembourg	113	113	113	113	113	113	287	364	328	253	203
Canada	102	85	54	83	89	89	307	346	333	257	201
Sweden	160	154	146	135	71	107	378	440	384	236	215
Brazil	593	515	437	424	424	424	661	643	592	528	489
Thailand	148	128	138	148	158	158	273	348	319	261	220
Singapore	118	118	118	118	118	118	348	416	363	279	224
Malaysia	93	93	93	93	93	93	240	305	266	198	156
Belgium	199	224	238	247	253	263	399	490	462	393	353

Source: Bloomberg, WEO, IMF Staff estimates. The slope is measured as the 10-year sovereign yield over the 3-month Treasury bill.

View Table

Table 4.

Switzerland: Slope Yield Curve Projections

While short-term yields decline in the adverse scenario as monetary policy reacts to the global recession, long-term yields increase due to the widening of credit spreads, leading to the steepening of the yield curve.

	Observed	Baseline					Adverse
	2018	2019	2020	2021	2022	2023	2019	2020	2021	2022	2023
Switzerland	96	105	108	116	123	130	320	388	352	280	234
United States	49	24	‐5	17	39	44	248	286	264	205	156
United Kingdom	84	68	59	55	50	44	290	349	301	215	154
China	69	69	69	69	69	69	217	277	236	169	130
Germany	130	150	159	163	163	168	324	410	378	304	258
Japan	33	40	46	68	77	87	131	89	83	79	84
Australia	106	91	86	81	64	64	319	380	331	234	181
Spain	225	256	276	292	314	319	580	654	606	531	469
Korea	118	118	118	118	118	118	348	416	363	279	224
Hong Kong	69	69	69	69	69	69	217	277	236	169	130
Italy	248	286	309	327	320	317	611	688	642	538	467
Netherlands	113	113	113	113	113	113	287	364	328	253	203
Ireland	168	182	197	210	221	229	357	449	426	361	319
Luxembourg	113	113	113	113	113	113	287	364	328	253	203
Canada	102	85	54	83	89	89	307	346	333	257	201
Sweden	160	154	146	135	71	107	378	440	384	236	215
Brazil	593	515	437	424	424	424	661	643	592	528	489
Thailand	148	128	138	148	158	158	273	348	319	261	220
Singapore	118	118	118	118	118	118	348	416	363	279	224
Malaysia	93	93	93	93	93	93	240	305	266	198	156
Belgium	199	224	238	247	253	263	399	490	462	393	353

Source: Bloomberg, WEO, IMF Staff estimates. The slope is measured as the 10-year sovereign yield over the 3-month Treasury bill.

71. A structural Merton-based model was used to extract PD estimates from the sovereign spreads projected under each scenario. This approach assumes that the difference between a risk-free security and a risky security is the put option on the value of the assets defined as the expected loss under stressed conditions (i.e., PD*LGD). Using the credit spreads projected by the scenario for sovereign i, i.e., $S_{t,T}^i$ , the residual maturity of security i (T-t), and assuming LGD=45 percent, the implied risk-neutral default rate for security i is computed as:

72. Impairments on sovereign exposures were calculated on claims not measured at fair value through profit and loss. Newly created adjustments to provisions for default risk were projected for sovereign exposures booked in available-for sale, loans and receivables, and held-to-maturity investments. This contrast with the fair value assessment of sovereign positions booked in the held for trading category which were subject to market shocks (see market risk modelling section).

Stressing Complex Credit Risk Products

73. Default risk for leverage loan exposures were estimated using G-SIBs’ BBA templates. Leveraged loan exposures include temporary exposures from syndicated underwriting activities, including syndicated loans, bridge loans, and equity bridges. These exposures cover facility type positions including term loans, revolving credit facilities, and loan commitments. The IMF team applied stressed PD and LGD parameters consistent with the credit risk paths for high-yield corporate indices in the scenario on banks’ reported leverage loan exposures to estimate credit losses from these positions.

74. Credit risk from banks’ structured credit products was covered in the scope of credit risk loss estimates. This category includes securitization products (e.g., asset backed securities (ABS), mortgage backed securities (MBS)) of financial assets including mortgages, credit card receivables, or auto loans, as well as structured credit derivatives. For the computation of the profit and loss (P&L) impact, the amount of impairments was calculated on the carrying value for accrual positions and on net exposures for fair value positions. For each security, the underlying pool’s credit and prepayment models was stressed under the IMF scenario to produce consistent impairment estimates, drawing on banks’ estimates on stylized price shocks reported in BBA templates.

75. In addition, default risk from Lombard lending positions on uncovered exposures was also included in the analysis. The analysis drew on BBA reporting templates which include the exposures and losses of the entire Lombard portfolios.²⁹ To determine the size of uncovered exposures, the IMF scenario on market shocks was applied. The stress loss was computed by applying stressed PDs from retail exposures to the net exposures after applying the scenario shocks to market risk factors.

Impact of Credit Risk on Regulatory Capital

76. The impact of credit risk on banks’ capital ratios depends on the regulatory approach used by banks to book credit exposures. For exposures booked under the IRB approach, credit risk evolves with the exposure at default (EaD), the PD, and the LGD. For exposures under the STA approach, credit risk is reflected in higher provisions for default risk and higher risk weights from a deterioration of the creditworthiness of underlying exposures. This is driven by higher LTVs in mortgage loans, credit risk migration to substandard categories for performing loans, and a rise in past due loans for non-performing loans.

77. At end-2018, there are four main IRB banks in the Swiss banking system. Main IRB banks include the two G-SIBs, and two of the largest cantonal banks. While the G-SIBs are A-IRB banks that book most credit exposures, including sovereign claims, under IRB (about 95 percent), the cantonal banks are F-IRB banks with very conservative F-IRB implementation, and RWA levels close to standardized values. For these banks a considerable part of the sovereign exposures are cash holdings at the SNB.

Credit Risk Model for IRB Exposures

78. The impact of macrofinancial conditions on provisions was calculated by estimating expected losses: ${EL}_{i,t}^j={PD}_{i,t}^j*{LGD}_{i,t}^j*{EAD}_{i,t}^j$ where i denotes the bank, j denotes the asset class, and t is the time dimension. Credit risk parameters were estimated by asset class, geography, and scenario. The EaD evolves with the growth of the balance sheet and the migration of off-balance sheet exposures to on-balance sheet using credit conversion factors (CCF). Even if under current Swiss accounting standards, banks do not apply the expected credit loss approach, banks using IRB approaches need to deduct any “shortfall of provisions” from CET1 capital.

79. The regulatory credit risk parameters (PD, LGD, EaD) for Swiss banks’ approved IRB models tend to be Through-the-Cycle (TTC). Switzerland adopted Basel Standards that required a calibration based on long-run averages and additionally a downturn calibration for CCF and LGD. Although credit risk parameters used to compute capital requirements are mostly TTC, there are some PiT components across portfolios.

80. To derive PD projections, a multiplier approach was applied on the paths generated by IMF credit risk models. The parameter paths, derived from the satellite models, form the basis for projecting bank loan losses conditional on the macro-financial scenario. To quantify the build-up of loan loss reserves, the estimated paths were applied on bank-level starting regulatory parameters using corresponding country-based default rate multipliers estimated by IMF models. These were applied on an exposure-weighted basis across the geographical breakdown of exposures.

81. For the LGD parameters, the structural model for mortgage risk was used for housing-related portfolio segments.³⁰ For other portfolios, the FSAP team used the multipliers calibrated by European Central Bank (ECB) staff for the Euro Area FSAP as well as on insights from the global financial crisis. Shifts to LGDs were calibrated by geography and portfolio. LGD projections were derived by bank, using estimated non-defaulted LGD ratios and shifts to LGD by portfolio as:

where ${LGD}_{i,t}^j$ is bank i post-credit risk mitigation LGD for portfolio j at time t, and ${\mathrm{\Delta }LGD}_t^j$ is the aggregate LGD shift for portfolio j at time t.

82. The projection of EAD was driven by balance sheet assumptions, structural FX risk in foreign geographies, and triggered credit lines and guarantees. Specifically, changes to EAD in the IRB portfolio were governed by:

where i denotes the bank, j denotes the asset class, and t is time, g_i,t is the growth rate of the IRB portfolio, $f_{i,FX}^j$ is the fraction of foreign currency loans. ΔFX_t is the shock to foreign currency under the macro scenario, $\left({1-PD}_{i,t-1}^j\right)$ represents the non-defaulted portfolio, ${\mathrm{\Delta }L}_{i,t}^j$ is the shock to triggered credit lines and guarantee s, and ${UCL}_{i,t-1}^j$ is the amount of undrawn guarantees.

83. To compute capital requirements, regulatory risk parameters conditional on the scenario, were projected using the Basel III formula for IRB exposures. The derivation of RWAs is dependent on estimates of PD, LGD, EAD, correlation assumptions, and effective maturity for each exposure. According to the Basel III framework, RWAs were computed after applying the scaling factor of 1.06 to credit RWAs. Also, a multiplier of 1.25 was applied to the correlation parameter of all exposures to large regulated financial institutions and to all unregulated financial institutions.

84. An adjustment factor was applied on projected RWAs to account for idiosyncratic elements in banks’ IRB modeling approach. While IMF projected paths are derived using credit risk proxies at the portfolio level, banks tend to use credit scoring models at the loan level in their estimation of credit risk. Although banks are expected to apply risk-weight functions provided by the Basel III regulatory framework, which maps the risk parameters to risk-weighted assets, the risk-weight function is concave on default risk which yields lower capital requirements when credit risk is estimated at the loan level. To address this bias, an adjustment factor calculated as the ratio between the projected RWA at the starting point and banks’ reported RWA is applied over the estimated period.

Credit Risk Model for STA Exposures

85. To estimate loans loss provisions for STA exposures the FSAP team considered a range of approaches to compute the starting value of the risk parameters. Four approaches were used to construct the starting point of default rates:

• Compute the ratio of the stock of exposures in default over the amount of total exposures drawing on Basel templates.

• Calculate the share of impaired loans and non-performing loans over total receivables in the credit portfolio;

• Estimate the ratio of newly created adjustments to provisions for credit risk charged to income statement over gross loans, to exclude the impact of write-offs and recoveries;

• Use the 1-year expected default rate reported for IRB exposures for IRB banks and that of peer banks for STA banks.

86. Given the benign cyclical conditions of the Swiss economy, the FSAP team with decided to use the forward-looking IRB default rate in peer banks’ portfolios as a starting point for STA exposures. The average NPL ratio in 2017 for the sample banks reached 0.5 percent for customer loans and 0.3 percent for mortgage loans. The amount of newly created provisions on the gross carrying amount of loans was even lower. Also, the available time series did not pick up sufficiently severe stress events making the statistical approach on these credit measures unsuitable. While the IRB default rate is a forward-looking measure rather than an incurred loss measure, it has the advantage of reporting a 1-year flow and therefore does not require additional assumptions on write-off rates and recoveries over the stress test horizon.

87. Currently, the Basel III regulatory treatment of credit losses for STA exposures allows a forward-looking concept in general provisions (GP) albeit this is added to Tier II capital. Under Basel regulatory framework, GP are defined as “provisions or loan-loss reserves held against future, presently unidentified losses which subsequently materialize” and while they are subtracted from CET1 capital, they are added to Tier II capital. By contrast, specific provisions (SP) include “provisions ascribed to identified deterioration of particular assets” and they are excluded from both regulatory capital and from RWAs for credit risk.

88. At the same time, the Basel Committee is considering creating a level-playing field in the capital regulatory treatment of provisions by adopting the expected loss approach for STA provisions. One of the options being considered is to fundamentally change the current regulatory treatment of provisions – remove the GP/SP distinction and introduce regulatory expected losses under SA. This is in line with the approach taken by the IMF team to project credit impairments for standardized exposures.

89. The coverage ratio for defaulted exposures was floored at the LGD rate reported for the same portfolio under the IRB approach.³¹ Banks report the general and specific provisions computed on defaulted exposures in Basel templates. This measure captures the eligible collateral and guarantees on defaulted claims in line with credit risk mitigation techniques. Under the adverse scenario, a 65 percent estimated coverage ratio was applied subject to the following constraint:³²

90. To forecast the flow of exposures in default, the default rate was multiplied by the amount of performing exposures projected each period. General and specific credit risk adjustments were projected by multiplying the flow of new impairments by the stressed LGD by geography and portfolio. The flow of new impairments was projected using the same econometric approach applied to IRB exposures. In line with balance sheet assumptions for the IRB portfolio, the stock of performing exposures was driven by the growth of the loan book, structural FX risk in foreign exposures, triggered credit lines and guarantees, and the flow of new impairments.

91. Regulatory capital requirements were computed using three steps:

• For the defaulted portfolio, the risk weight of exposures in default, excluding exposures at 0 percent risk weight, was computed for each portfolio drawing on Basel templates;

• The risk weight of exposures in default was set at an average 100 percent, following Basel III’s regulatory framework which specified that a 100 percent risk weight should be applied on past-due loans when specific provisions are no less than 20 percent of the outstanding amount of the loan. The risk weight on the new flow of exposures in default was computed as:

  ${RW}_t^{c,i}|default=\max \left\{{RW}_{2017}^{c,i}\right|default,100\%\}$

• For the non-defaulted portfolio, one-notch downgrade of the underlying exposure was assumed under the adverse scenario.³³ The initial risk-weight density reported by banks for non-defaulted exposures in Basel templates was mapped to the external rating by portfolio laid out by Basel, and one-notch downgrade was added to calculate the projected risk-weight density.³⁴

92. Capital requirements for STA exposures were also driven by changes in provisioning rates, growth of EAD, structural FX risk, triggered credit lines and guarantees, and migration effects. There are four main components driving shifts to RWAs in the STA portfolio. The first component reflects the motion of RWAs generated by the flow of provisions, the growth rate of the portfolio, and FX effects. The second component shows the increase in risk weights resulting from triggered off-balance sheet credit lines and guarantees. The third component reflects the increase in risk density from the transition of loans from the performing to nonperforming category. Finally, the fourth component denotes the change in risk density from the transition matrix estimated for performing exposures.

93. To avoid a structural break in the loan loss impairment projections, an adjustment factor was applied based on the reported newly created provisions at the starting point. To address the IMF upward bias of a rise in provisions from the choice of the forward-looking approach relative to Swiss GAAP incurred loss accounting standards, an adjustment factor was computed as the ratio between the projected loan loss impairment charges at the starting point and banks’ reported new inflows of provisions. This limited the contribution of credit losses to banks’ capital shortfall under stress.

Data Granularity

101. The FSAP team extracted data from SNB BS surveys, G-SIBs’ BBA reporting, and banks’ annual reports and Pillar 3 disclosures to build a granular database. Fair value positions were identified by issuer, regulatory book, and counterparty. SNB balance sheet surveys were used to identify the volume of trading portfolio, other financial investments at fair value, and financial investments. The latter groups AFS and held to maturity (HTM) positions. The breakdown by accounting category was achieved by combining information from banks’ annual reports.

102. While SNB BS positions include fine breakdowns for some categories, supplementary information was required to identify the portfolio sensitivities to market shocks. SNB BS data includes breakdowns for domestic and foreign positions, material currencies, major counterparties for domestic portfolios, and geographies for foreign positions. However, additional data was extracted from banks’ annual report identify the sensitivity of positions to market shocks, including reference equity indices, sovereign benchmark curves, and country corporate indices.

103. The analysis shows that sovereign exposures are booked mainly in HFT and HTM portfolios (Figure 10). On an asset-weighted basis, about 50 percent of sovereign exposures are booked under HTM portfolios, 40 percent in trading and fair value portfolios, and the remaining 10 percent in AFS portfolios. However, there is a large dispersion across banks with G-SIBs booking primarily in HFT, DFBs in HTM, and private banks either in AFS or HTM. In terms of issuers, sovereign exposures are predominantly to the United States, United Kingdom, Switzerland, and other foreign issuers given the insufficient supply of HQLA to meet the LCR requirements in CHF (excluding the current high levels of banks’ sight deposits with the SNB). In Switzerland, banks are allowed to include additional HQLA in foreign currencies when calculating the LCR under the Alternative Liquidity Approaches options.

Switzerland: Sovereign Exposures by Accounting Category

Most exposures are booked under the HTM and HFT books with a large dispersion across banks.

(In percent)

Citation: IMF Staff Country Reports 2019, 189; 10.5089/9781498321808.002.A001

Source: IMF Staff Estimates. The sample of banks included the twelve major Swiss banks. Boxplots include the mean (yellow dot), the 25th and 75th percentiles (boxes) and the 15th and 85th percentiles (whiskers).

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Switzerland: Sovereign Exposures by Accounting Category

Most exposures are booked under the HTM and HFT books with a large dispersion across banks.

(In percent)

Citation: IMF Staff Country Reports 2019, 189; 10.5089/9781498321808.002.A001

Source: IMF Staff Estimates. The sample of banks included the twelve major Swiss banks. Boxplots include the mean (yellow dot), the 25th and 75th percentiles (boxes) and the 15th and 85th percentiles (whiskers).

• Download Figure
• Download figure as PowerPoint slide

Switzerland: Sovereign Exposures by Accounting Category

Most exposures are booked under the HTM and HFT books with a large dispersion across banks.

(In percent)

Citation: IMF Staff Country Reports 2019, 189; 10.5089/9781498321808.002.A001

Source: IMF Staff Estimates. The sample of banks included the twelve major Swiss banks. Boxplots include the mean (yellow dot), the 25th and 75th percentiles (boxes) and the 15th and 85th percentiles (whiskers).

• Download Figure
• Download figure as PowerPoint slide

Market Risk Approach for G-SIBs

104. To compute the P&L impact from shocks to market risk factors on G-SIBs, the FSAP team used data on banks’ risk profile from BBA templates. The BBA templates group exposures in modules that represent the key drivers of banks’ risk profile. The templates also collect delta sensitivities by major index/counterparty, with a breakdown of long vs. short positions, and cash vs. derivative positions. To account for non-linear valuation effects from large shocks, the report also collects fair value impact to severe shocks prescribed at two severity levels, i.e., ‘severe event’, and ‘extreme event’.

105. FINMA reporting templates suggest that G-SIBs rely on delta hedging strategies to manage market risk. Delta hedging is an option-based strategy that aims to reduce, or hedge, the risk associated with price movements in the underlying asset, by offsetting long and short positions. The small net delta, with netting of long and short positions in cash and derivative positions, reported by banks for a 1-bp move to the price of the underlying, suggests that banks are, on aggregate, well hedged under normal market conditions.

106. Banks are, however, exposed to basis risk if correlations between cash and derivative positions break down under stress. Basis risk is the most important risk when trading in the derivative market. Since derivatives essentially are traded on the basis of the value of the underlying asset, any disproportionate fall in the value of the underlying asset would cause a crash in the derivatives designed for that purpose. If there is non-convergence of spot price and relative price on the offset date of trade, banks are exposed to adverse price shocks to the underlying.

107. To assess basis risk, the IMF team used differentiated shocks across the cash and derivative positions. Relative shocks to cash and derivative positions were informed by:

• The financial crisis. The implicit assumption is that to hedge sovereign bond curve risk in fair valued government securities, banks use derivatives priced against the swap curve. Figure 8 suggests that basis risk in fixed income markets widened during the global financial crisis, exposing banks to basis risk. While the daily spread between the 5-year sovereign yield and the swap rate hovered about -31 bps in 2000–2018, it tripled to over -100 bps in October 2008;

• Market liquidity. Differential impact in the fall of the underlying asset and its derivative can be triggered by the different degree of market liquidity that characterizes the trade risk scenario. This is associated to the liquidity horizon under which banks can exit positions in view of the likely market trade volumes under the stress scenario. For corporate bond curve risk, the IMF team assumed a 30-day holding period while for swap-referenced positions, the holding period is assumed to be 10-days (Figure 8).

• The 2018 Comprehensive Capital Analysis and Review (CCAR) severely adverse market scenario. For BBB corporate bonds, the relative shock to the bond position and the underlying derivative indexed by the Credit Default Swap Index-Investment Grade (on the run) was estimated at a scaling factor of 2.

108. The P&L fair valuation impact resulted from applying the corresponding shock to the cash and derivative position separately. BBA data includes the net delta on cash and derivative positions. For instance, for credit spread risk, net delta is reported by credit rating (IG, HY), counterparty (sovereign, public entities, financial, corporate), and region (Switzerland, North America, Asia...). This enabled the application of differentiated shocks calibrated in the IMF scenario on the fair value of the position to compute P&L impact. Information related to the valuation impact reported by banks on prescribed shocks was also used to account for non-linear effects in the size of the shock.

Market Risk Approach for DFBs and Private Banks

109. The valuation impact on debt securities was disaggregated into repricing risk and credit spread risk. The impact on P&L and capital is recognized over the stress test horizon. Although the notional value of exposures evolves with the balance sheet, banks are assumed to maintain the same structure of the portfolio throughout the stress testing horizon.

110. The impact of repricing risk depends on the shock to the risk-free curve and bond duration. Interest rate sensitivity is measured by the effective maturity of the bond proxied by duration. Duration is approximated by the residual maturity of the bond at the cut-off date. Under the full valuation approach, the change in the risk-free bond price is the result of multiplying modified duration by the shock to the risk-free rate:

where Dⁱ denotes average duration of bond i,B_t denotes the carrying value of the debt security at time t, $r_t^i$ captures the risk-free rate level and ${\mathrm{\Delta }r}_t^i$ the shock to the risk-free rate.

111. The risk-free curve is proxied by the Swiss bond curve. Term premium decompression in fixed income markets together with the absence of flight-to-quality effects contribute to steepen the Swiss bond curve with the 10-year rate rising by 100 bps in 2020: Q4 relative to 2018: Q2 despite the decrease in policy rates. A polynomial interpolation method is applied to span the term structure between the 3-month and the 10-year rate generated in the macroeconomic scenario.

112. Credit spread risk in other geographies depend on the re-emergence of sovereign stress and monetary policy space. For countries with sovereign distress (e.g., peripheral Europe, some emerging markets) and at the zero-lower bound, credit spreads widen significantly relative to the Swiss bond curve. For countries with monetary policy space, the expansionary policy response assumed in the adverse scenario makes the yield curve pivot downwards mitigating the pricing impact at the short-end (e.g., China, United States).

113. The credit spread impact on fair value depends on shocks to credit spreads and bond duration. Using the modified duration approach, the FSAP team calculated the haircut for each fixed income instrument under each scenario by multiplying modified duration by the shocks to credit spreads:

where Dⁱ denotes average duration of bond j, $B_t^j$ denotes the carrying value of security j at time t, $r_t^i$ shows the risk-free rate, ${cs}_t^j$ denotes the credit spread, and ${\mathrm{\Delta }cs}_t^j$ the shock to credit spreads.

114. Although duration (proxied by residual maturity) is typically low in the HFT portfolio, is larger in the AFS portfolio, exacerbating valuation impact in a few cases. Duration affects valuation through two channels. First, as the curve pivots downwards and steepens under stress, long-duration assets are hit by larger shocks. Second, the sensitivity of the portfolio to shocks increases with the duration of the portfolio. Relatively less monetary policy space in some jurisdictions (e.g., euro area) limits the downward shift of the yield curve increasing the relative impact of term premium decompression and sovereign stress.

115. For equities held with a trading intent, the fair value impact was floored using the 2018 EBA methodology. The market impact from full revaluation of equity holdings was floored using the following constraint:

where the VaR scaling factor has been set to the upper bound of 1.5, and the trading position includes the fair value of equity instruments (assets) and the short positions in equity instruments (liabilities).

116. Hedges are assumed to be ineffective for DFBs and Private banks. Given data limitations, the FSAP team did not have access to market risk sensitives for DFBs and Private banks. Although DFBs typically carry small fair value positions, some private banks have larger sensitivity to market risk factors, particularly interest rate risk, credit spread risk, and equity risk. Under the assumption that Private banks hedge their position effectively (i.e., they do not take directional trades), the IMF approach was rather conservative.

Earnings Approach

123. Swiss banks’ net interest income has been compressed by the low interest rate environment (Figure 11). The average net interest margin (nim) on outstanding loans hovered about 150 bps in 2017. It reached 121 bps for outstanding claims, including financial claims, due to the exceptionally low interest rate environment with 3-month libor rates at -75 basis points and 5-year sovereign benchmark rates at -52 basis points. Banks have, however, managed to keep the nim broadly stable since the introduction of negative rates in Dec 2014 partly by widening asset margins on new mortgages. Going forward, the concern is that average lending rates will continue to fall as loans taken out in the past are renewed at lower rates. Also, there is recent evidence that asset margins on newly extended mortgage loans have decreased adding pressures to profitability. In addition, liability margins have been negative since 2015 given the zero lower bound threshold applied by bank on most retail deposits.

Switzerland: Interest Rate Environment

Citation: IMF Staff Country Reports 2019, 189; 10.5089/9781498321808.002.A001

Source: SNB, Bloomberg, and IMF calculations.

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Switzerland: Interest Rate Environment

Citation: IMF Staff Country Reports 2019, 189; 10.5089/9781498321808.002.A001

Source: SNB, Bloomberg, and IMF calculations.

• Download Figure
• Download figure as PowerPoint slide

Switzerland: Interest Rate Environment

Citation: IMF Staff Country Reports 2019, 189; 10.5089/9781498321808.002.A001

Source: SNB, Bloomberg, and IMF calculations.

• Download Figure
• Download figure as PowerPoint slide

124. Interest rate risk is a major risk for domestic banks. DFBs rely mainly on interest income to boost profitability. NII represents about 70 percent of total earnings. Banks are exposed to maturity transformation risk as they lock in rates on assets for longer periods that rates on liabilities which are mainly sourced from retail deposits. This risk has increased with the replacement of variable rate mortgages, which were repriced at the discretion of the bank, by long term fixed-rate mortgages as borrowers are increasingly inclined to secure fixed-rate mortgages in the current low interest rate environment. Also, the government benchmark curve has flattened since 2008 from 150 bps in the 5-year to 3-month Treasury spread to the current 50 bps, constraining long-term rates further.

125. The focus of the IMF test on IRRBB is on the EA rather than on the value of securities held at amortized cost. The key risk from sudden moves of interest rates is their potential adverse effect on the bank’ capital and earnings prospects. While the 2016 Basel standards on IRRBB include the monitoring of the impact of adverse shocks to interest rates on both the economic value of the bank’s banking book positions weakening the value of equity (EVE) as well as on the bank’s EA, the IMF IRRBB module investigated EA sensitivity as the economic impact of interest rate moves on banking book securities (AFS) was examined in the market risk module.

126. Results from the SNB’s IRRBB analysis suggest that banks would benefit from a moderate increase in interest rates but would be strongly negative affected by large upward shocks.³⁹ Banks’ net interest income would increase, on average, in the event of a small to moderate upward interest rate shock, despite their exposure to maturity transformation risk, as they would be able to restore the liability margin that has remained negative since the introduction of negative rates. The SNB estimates that following a 200-bps interest rate shift, banks’ net interest margin would increase by about 20 bps. By contrast, a large upwards shock in the interest rate curve would negatively affect earnings as the decrease in structural margin and the behavioral decrease in asset margin would offset the beneficial impact of the restoration of the liability margin. In the event of a 400-bps shock, banks’ interest rate margin would fall by about 20 bps from the current low levels.

127. The FSAP assessment of IRRBB followed a four-pronged approach which shares some similarities with the SNB approach:

• The ’economic value approach’, which measures the valuation effect of an interest rate shock on the fair value of assets, and impacts EVE due to long duration (e.g., long-dated fixed rate bonds), was evaluated using a market risk approach (see section on market risk). By contrast to the Basel EVE measure, only marked-to-market assets in the banking book (through other comprehensive income) were re-valued rather than all balance sheet positions;

• The ’earnings approach’, which measures the impact of an interest rate scenario on banks’ future net interest income over a 5-year horizon was assessed against a range of interest rate shocks calibrated at the bank/product level. While the SNB approach has developed a granular approach in terms of products and time bands, and has incorporated dynamic elements from bank and borrowers’ behavior, rates are projected at the banking system level;

• The scenario includes a commercial margin and an intermediation margin. In particular, two components are included in the calibration of the interest rate shocks: a systematic risk factor (broadly aligned with the scenario) and an idiosyncratic component (linked to banks rates’ historical behavior); and

• As in the SNB test, shocks were applied to the repricing structure of assets and liabilities using the Interest Rate Risk report as of June 2018. Impact is measured separately for the stock of loans/liabilities and for the flow of new lending/funding instruments.

Structure of Assets and Liabilities

128. The exposure of banks’ earnings to interest rate risk depends on changes in market rates as well as on banks’ structure of assets and liabilities. The impact of a change in market rates on funding costs depends on the liability margin which varies across funding instruments (e.g., sight deposits vs debt securities). The impact of a rate shock on interest income depends on the commercial margin which differs across loan products (e.g., mortgage vs consumer loans). Margins are a function of banks’ business model as well as the level of market competition in product markets.

129. For the sample of banks included in the stress test, mortgage loans represent about 40 percent of total loans.⁴⁰ About 18 percent are amounts due from banks, whereas Lombard loans represent 16 percent of outstanding claims. The share of ‘other customer loans’ and securities financing transactions is similar at 13 percent of total loans. This hides, however, important differences in the composition of the loan portfolio across banks, in line with banks’ business models. Whereas G-SIBs have relatively larger exposures to other banks (24 percent of loans), DFBs are heavily exposed to mortgage loans (82 percent of loans), and private banks have larger shares of Lombard loans (62 percent).

130. Swiss banks fund their balance sheet mainly with retail deposits. In line with banks domiciled in other advanced economies, large Swiss banks have re-oriented their business away from trading and more complex activities, towards less capital-intensive activities, including a retrenchment from foreign wholesale funding, supported by the low interest rate environment, and FINMA implementation of Basel LCR regulation. This pattern is reflected in banks’ increased reliance on customer deposit funding.

131. In the stress testing sample, retail funding represents over 70 percent of interest-bearing liabilities. By contrast, reliance on less stable sources of funding (’wholesale funding’) is low. Amounts due to banks and repos contribute to about 10 percent while debt securities issued account for 20 percent. All banks rely heavily on customer deposits regardless of their business model, with private banks showing greater reliance on retail deposits at about 85 percent of interest-bearing liabilities.

132. While market sensitivity of liquidity reserves at the SNB (asset side) and of retail funding (liability side) to money market stress is low, positions are subject to considerable behavioral uncertainty. This is because cash balances at the SNB are motivated by the ultra-low interest rate environment, and therefore, are subject to behavioral shocks if rates start to raise.⁴¹ Moreover, the decoupling of rates paid to depositors from market rates has increased the uncertainty of the pricing of these deposits if rates were to rise. Another element of uncertainty relates to the shift in retail deposits from sight deposits to term deposits should interest rates start to rise.

133. Some relevant dynamic elements were captured by the FSAP’s IRRB assessment. Given that banks’ capacity to generate substantial interest income depends significantly on dynamic effects, the FSAP team attempted to model explicitly changes in commercial margins, liability margins, and pass-through effects. Non-maturing deposits (e.g., current accounts) where modeled as overnight deposits. On the other hand, the team did not change the maturity profile of maturity positions at renewal, or customer behavior related to shifts between different types of deposits, in line with the constant portfolio nature of banks’ balance sheets. This contrast with SNB approach that incorporates dynamic elements in customer shifting patterns across deposits, and banks’ implied duration targets.

Modelling Strategy

134. The FSAP team calculated interest income and expense at the portfolio level using a twofold approach:

• For competitive loan products and funding instruments, rates were projected at the aggregate level using shocks to benchmark rates and assuming no changes in margins.

• For non-competitive loan products and credit sensitive liabilities, rates were projected at the bank level using shocks to systematic and idiosyncratic risk factors, with bank margins determined endogenously.

135. This approach addresses two major developments since the global financial crisis. First, funding market conditions have become heterogeneous across banks reversing a pre-crisis trend of lower and more similar rates. Second, there have been concerns over the policy rate pass-through to lending rates. Possible sources in the inefficient transmission of monetary policy have been attributed to the misalignment of policy rates and funding costs and to banks’ internal transfer policies to set lending rates. Both aspects are examined in this section.

136. The modeling strategy was enabled by the granular data shared by the Swiss authorities collected in FINMA’s interest rate risk (IRR) report. This report includes all banking book positions according to their repricing maturities. Time bands are highly granular with eighteen time-buckets, ranging from up to 1 day to over 15 years. Positions are classified by reporting category including: Category I: positions with determined repricing maturities (e.g., fixed-rate loans, or medium-term notes); and, Category II: positions with undetermined repricing maturities (i.e., sight claims, claims against customers, variable mortgage loans, sight liabilities, and customer deposits callable).⁴²

137. The IRR report was mapped to SNB’s maturity structure survey to deepen the portfolio composition and enable margin projection. The IRR does not report separately positions with remaining contractual maturity for instruments at fixed rate and instruments at floating rate.⁴³ This distinction is, however, required to project margins when fixed rate instruments are re-issued over the stress test horizon. Another limitation of the IRR report is that is based on cash-flows which include the nominal value (principal) and interest payments, whereas the IMF analysis is based on carrying amounts. The FSAP team mapped the IRR report to the SNB’s maturity structure survey which also has a finer breakdown of portfolios (e.g., Lombard loans, amounts due to banks, SFTs, ‘other customer loans’…) and splits carrying amounts by time to maturity including on sight, cancellable, and remaining maturity buckets (<1m, 1–3m, 3m-1y, 1–5y, and >5y).

138. To project individual bank specific margins, position data was combined with income statement information from banks’ annual reports and Dealogic. To estimate empirically banks’ pricing behavior, interest income on loans was split into main product categories. To project funding cost, information from banks’ annual reports on cost on deposits and other liabilities was mapped to interest spread over benchmark at issuance for debt instruments drawing on Dealogic data.

Funding Costs Modeling

139. A key contribution of this approach is to carefully construct an effective measure of funding costs. Interest rates paid on liabilities vary across banks depending on their funding patterns and liquidity management practices.⁴⁴ Given that core funding markets vary across the sample of banks, we expect different rates paid on deposits across banks. Another important determinant of interest expense is banks’ liability margin which depends, among other factors, on the proximity to the zero lower bound.

140. The effective funding cost measure is built by matching bank’ liability structure with the interest rate path by funding instrument conditional on the scenario:

• Amounts due to banks and funding from repo transactions are linked to 3-month libor rates by currency. The breakdown of liabilities by instrument and currency (CHF, USD, EUR, other) is available in SNB balance sheet surveys;

• The rate on customer deposits is a function of the type of deposit (sight deposits, saving deposits, time deposits) and evolves with benchmark rates (libor rates, government yields);

• Debt securities are repriced according to the swap rate (at time to repricing bands) and the spread over benchmark at issuance (at re-issuance intervals) which is bank and instrument specific. Debt securities include cash bonds, bond issues, and central mortgage institution loans;