FBBABLLR1CBQ_US Commercial Banks: Assets - Bank Credit - Loans and Leases - Residential Real Estate (Bil, $, SA)

Notes on new forecast variables November 2018 Loc Quach Moody s Analytics added 11 new U.S. variables to its global model in November. The variables pertain mostly to bank balance sheets and delinquency rates, but other ad hoc variables were added as well. The purpose of these additions was to enhance our forecast offerings and improve the global model product. The new forecasts leverage the existing Moody s Analytics forecasting suite. FBBABLLR1CBQ_US Commercial Banks: Assets - Bank Credit - Loans and Leases - Residential Real Estate (Bil, $, SA) Moody s Analytics has added a residential real estate bank assets forecast to its global model. The forecast for this variable draws upon the forecast for total real estate loans. This variable is included on the left-hand side of the forecast equation, and the residential percentage of total real estate assets is modeled as the differenced variable with a differenced log functional form. The explanatory variable is the ratio of house prices to commercial real estate prices. An increase in the ratio translates into a larger share of residential real estate assets as a proportion of total real estate assets. The equation fails a Breusch-Godfrey test for serial correlation, but Moody s Analytics permitted serial correlation in this instance because of the persistence of house prices. The equation demonstrated much lower RMSE in back-testing than an AR(2) model. FBBABLLROCCBQ_US Commercial Banks: Assets - Bank Credit - Loans and Leases - Commercial Real Estate (Bil, $, SA) Moody s Analytics has added a commercial real estate bank assets forecast to its global model. The forecast for this variable draws upon the forecasts for total real estate assets and residential real estate assets. The forecast is an identity; its value is equal to the difference between total and residential real estate assets. This identity exists in history, providing Moody s Analytics with justification to maintain the identity in our forecast. FBBLDPCBQ_US Commercial Banks: Liabilities - Deposits (Bil. $, SA) Moody s Analytics has added a commercial bank deposits forecast to its global model. The forecast for this variable leverages Moody s Analytics forecasts on bank deposits included in the M1 and M2 monetary aggregates. Moody s Analytics also utilizes a lagged dependent variable. We tested both a lagged dependent variable version of this equation and a univariate one that depended solely on monetary category components. The one that included a lagged dependent variable produced a better forecast, lower back-testing error, and intuitive shock properties. FBCFDELLCCQ_US Delinquency Rates: Top 100 Commercial banks - Credit Card Consumer Loans (%,SA) Moody s Analytics has added a credit card bank delinquency rate forecast to its global model. We relied heavily on the ABA credit card delinquency rate forecast as the primary explanatory variable for this concept. We modeled in level terms, which was appropriate because the dependent variable is stationary. We included a lagged dependent variable and an AR(1) term as well. We also tested versions of this specification without these two explanatory variables and with a differenced functional form.

However, the chosen specification demonstrated the best combination of shock properties, baseline forecast, and low error during equation back-testing. FBCFDELLIQ_US Delinquency Rates: Top 100 Commercial banks - C&I Loans (%,SA) Moody s Analytics has added a C&I bank loan delinquency rate forecast to its global model. We experimented with a number of specifications and transformations. We considered bank charge-off rates as the primary regressor, and while the specifications of this type resulted in good back-testing performance because of the tight correlation between delinquency and charge-off rates, we could not justify charge-offs as drivers of delinquencies, since in practice a loan becomes delinquent before it is charged off. We settled on using the unemployment rate as our primary cyclical driver. Delinquencies and defaults on all loans rise when the economy falls into recession, and the unemployment rate is a great barometer of the business cycle. We also included lending standards with a lag and moving average to capture the phenomenon of laxer underwriting standards leading to a deterioration in loan performance over time. We also added an AR(1) term to eliminate serial correlation in the equation s error terms. The equation was modeled in differenced functional form. FBCFDELLRCQ_US Delinquency Rates: Top 100 Commercial Banks - Commercial Real Estate Loans (%, SA) Moody s Analytics has added a CRE bank loan delinquency rate forecast to its global model. We wanted the structure of this equation to mimic that of the C&I loan delinquency rate forecast. Therefore, we featured the unemployment rate as our primary cyclical driver and loan-specific lending standards. However, the variables historical fluctuations are far more extreme than for C&I loans. As a result, the estimated elasticities using the full range of historical data were extremely high. This rendered the equation extremely sensitive. Slight fluctuations in the economy were capable of pushing the delinquency rate either skyward or below zero. To combat this challenge, Moody s Analytics estimated in a level functional form with a constant term, a lagged dependent variable, and an AR(1). This dampened the elasticities of the cyclical variables. We also instituted a floor for the variable at 0.1. FBCFDELLRRQ_US Delinquency Rates: Top 100 Commercial Banks - Residential Real Estate Loans (%, SA) Moody s Analytics has added a residential real estate bank delinquency rate forecast to its global model. This equation is far more similar to the credit card delinquency rate equation than the C&I and CRE forecast equations. This is because Moody s Analytics could lean on its forecast for MBA loans past due. We experimented with level and differenced functional form, and with and without a lagged dependent variable. The specification that had the best combination of back-testing error, baseline forecast and shock properties was the one that used a differenced functional form with a one-period lagged dependent variable. FHVACHQ_US Homeowner Vacancy Rate (%, SA) Moody s Analytics has added a homeownership vacancy rate forecast to its global model. This was a difficult variable to model, as few variables are highly correlated with it. Before modeling, Moody s Analytics ran a unit root test to determine whether the variable needed to be transformed. The ERS test failed to reject the null hypothesis of a unit root, so Moody s Analytics modeled this variable in differenced log functional form. Moody s Analytics used the rental vacancy rate as a regressor in this

forecast equation. This is because we believe a priori that the two should move in tandem, since they are both measures of housing vacancy. We also added a lagged dependent variable. We experimented with an AR(1) term, but it was significantly disturbing the co-movement of the homeowner and rental vacancy rate forecasts. The arrived-upon specification demonstrates no serial correlation, beats the AR(2) benchmark in back-testing, produces a forecast that is in sync with the rental vacancy rate, and has good shock properties. FSP500EQ_US S&P 500 Composite: Price Index - End of period, (Index 1941-43=10, NSA) Moody s Analytics has added an S&P 500 end-of-period forecast to its global model. The dependent variable had to be influenced by our forecast for the S&P 500, since the only thing separating them was that one was a quarterly average whereas the other was the value of the stock market benchmark at the end of a quarterly period. We experimented with lags, moving averages and different transformation, but it was important to us that the coefficient of the S&P 500 regressor was near 1, implying a 1-for-1 movement in the broader index and the dependent variable. This was accomplished by a simple univariate regression. We did not add a constant term because we didn t want to introduce a time trend in the long-term forecast that would break the co-movement of these two variables. The arrived-upon specification beats the AR(2) benchmark in unconditional back-testing and trounces it in conditional back-testing. In scenarios, it moves in line with the S&P 500 index. FTWDMJRQ_US Weighted Average Exchange Value of U.S. Dollar: Major Currencies Index Nominal Moody s Analytics has added a weighted average exchange value of the U.S. dollar major currencies index forecast to its global model. This variable differs from the nominal broad trade-weighted dollar index only in that it considers just seven currencies rather than a broader basket. The currencies it contains are those issued by the countries of Japan, Canada, Switzerland, Sweden, the euro zone, the United Kingdom, and Australia. Moody s Analytics attempted to put all seven bilateral exchange rates in this regression, but the euro zone, U.K. and Australian currencies had the wrong signs. Therefore, this architecture had to be abandoned. We instead use the nominal broad trade-weighted dollar index. We transformed the regressor and the dependent variable by taking a differenced log since the dependent variable failed an ERS unit root test. The specification outperformed the AR(2) benchmark in backtesting and moves alongside the nominal broad trade-weighted dollar in scenarios. FCPIUEHC_US CPI: Urban Consumer - Owners' equivalent rent of residences, (Index Dec1982=100, SA) Moody s Analytics has added a CPI of owners equivalent rent forecast to its global model. We modeled in differenced log functional form to avoid a spurious regression. We tested different housing-related CPI components and even measures of population. However, the most effective regressor was the shelter CPI for urban consumers. This variable explains the vast majority of the movement in the dependent variable. The equation does not display serial correlation. We did not add a constant term so as not to add an unwanted time trend. The specification performs significantly better than the AR(2) benchmark in both conditional and unconditional back-testing.

New equation specifications Dependent Variable: DLOG(FBBABLLR1CBQ_US/FBBABLLRCBQ_US) Date: 10/24/18 Time: 16:24 Sample (adjusted): 2004Q4 2018Q2 Included observations: 55 after adjustments DLOG(@MOVAV(FHOFHOPIPOQ_US/FZFL 075035503Q_US(-3),6)) 0.136656 0.044597 3.064198 0.0034 R-squared 0.071356 Mean dependent var -0.002340 Adjusted R-squared 0.071356 S.D. dependent var 0.007866 S.E. of regression 0.007580 Akaike info criterion -6.908603 Sum squared resid 0.003103 Schwarz criterion -6.872106 Log likelihood 190.9866 Hannan-Quinn criter. -6.894489 Durbin-Watson stat 1.086090 @IDENTITY FBBABLLROCCBQ_US = FBBABLLRCBQ_US - FBBABLLR1CBQ_US Dependent Variable: DLOG(FBBLDPCBQ_US) Date: 10/25/18 Time: 15:22 Sample (adjusted): 1973Q3 2018Q2 Included observations: 180 after adjustments DLOG(FDDQ_US+FOCDQ_US+FSAVCQ_U S+FSTDCQ_US+FMMMFRQ_US) 0.374831 0.046286 8.098180 0.0000 DLOG(FBBLDPCBQ_US(-1)) 0.525337 0.052282 10.04810 0.0000 R-squared 0.133852 Mean dependent var 0.016558 Adjusted R-squared 0.128986 S.D. dependent var 0.008644 S.E. of regression 0.008067 Akaike info criterion -6.791040 Sum squared resid 0.011583 Schwarz criterion -6.755562 Log likelihood 613.1936 Hannan-Quinn criter. -6.776655 Durbin-Watson stat 2.525589

Dependent Variable: FBCFDELLCCQ_US Method: ARMA Generalized Least Squares (Gauss-Newton) Date: 10/26/18 Time: 12:51 Sample: 1991Q2 2018Q2 Included observations: 109 Convergence achieved after 6 iterations Coefficient covariance computed using outer product of gradients d.f. adjustment for standard errors & covariance FABABC_US 0.127361 0.047264 2.694653 0.0082 FBCFDELLCCQ_US(-1) 0.887233 0.039325 22.56173 0.0000 AR(1) 0.561637 0.091123 6.163492 0.0000 R-squared 0.973630 Mean dependent var 4.047064 Adjusted R-squared 0.973132 S.D. dependent var 1.175774 S.E. of regression 0.192725 Akaike info criterion -0.424490 Sum squared resid 3.937156 Schwarz criterion -0.350416 Log likelihood 26.13468 Hannan-Quinn criter. -0.394450 Durbin-Watson stat 2.072632 Inverted AR Roots.56 Dependent Variable: D(FBCFDELLIQ_US) Method: ARMA Generalized Least Squares (Gauss-Newton) Date: 11/01/18 Time: 17:21 Sample: 1993Q1 2018Q2 Included observations: 102 Convergence achieved after 3 iterations Coefficient covariance computed using outer product of gradients d.f. adjustment for standard errors & covariance D(@MOVAV(FLBR_US,2)) 0.562029 0.075608 7.433492 0.0000 @MOVAV(FXSLASCILQ_US(-8),4) -0.005809 0.000954-6.089409 0.0000 AR(1) 0.320450 0.096818 3.309822 0.0013 R-squared 0.692082 Mean dependent var -0.035490 Adjusted R-squared 0.685861 S.D. dependent var 0.260170 S.E. of regression 0.145820 Akaike info criterion -0.982852 Sum squared resid 2.105090 Schwarz criterion -0.905647 Log likelihood 53.12544 Hannan-Quinn criter. -0.951589 Durbin-Watson stat 2.131062 Inverted AR Roots.32

Dependent Variable: FBCFDELLRCQ_I_US Method: ARMA Generalized Least Squares (Gauss-Newton) Date: 11/02/18 Time: 14:29 Sample: 1994Q1 2018Q2 Included observations: 98 Convergence achieved after 30 iterations Coefficient covariance computed using outer product of gradients d.f. adjustment for standard errors & covariance C -1.332920 0.688319-1.936486 0.0558 @MOVAV(FLBR_US,2) 0.417489 0.120478 3.465284 0.0008 @MOVAV(FXSLASREQ_US(- 11),4) -0.004429 0.004421-1.001761 0.3191 FBCFDELLRCQ_I_US(-1) 0.677195 0.071029 9.534128 0.0000 AR(1) 0.936934 0.050830 18.43258 0.0000 R-squared 0.992970 Mean dependent var 3.137755 Adjusted R-squared 0.992667 S.D. dependent var 2.632258 S.E. of regression 0.225403 Akaike info criterion -0.070726 Sum squared resid 4.725013 Schwarz criterion 0.061160 Log likelihood 8.465555 Hannan-Quinn criter. -0.017380 F-statistic 3283.859 Durbin-Watson stat 2.157754 Prob(F-statistic) 0.000000 Inverted AR Roots.94

Dependent Variable: FBCFDELLRCQ_I_US Method: ARMA Generalized Least Squares (Gauss-Newton) Date: 11/02/18 Time: 14:29 Sample: 1994Q1 2018Q2 Included observations: 98 Convergence achieved after 30 iterations Coefficient covariance computed using outer product of gradients d.f. adjustment for standard errors & covariance C -1.332920 0.688319-1.936486 0.0558 @MOVAV(FLBR_US,2) 0.417489 0.120478 3.465284 0.0008 @MOVAV(FXSLASREQ_US(- 11),4) -0.004429 0.004421-1.001761 0.3191 FBCFDELLRCQ_I_US(-1) 0.677195 0.071029 9.534128 0.0000 AR(1) 0.936934 0.050830 18.43258 0.0000 R-squared 0.992970 Mean dependent var 3.137755 Adjusted R-squared 0.992667 S.D. dependent var 2.632258 S.E. of regression 0.225403 Akaike info criterion -0.070726 Sum squared resid 4.725013 Schwarz criterion 0.061160 Log likelihood 8.465555 Hannan-Quinn criter. -0.017380 F-statistic 3283.859 Durbin-Watson stat 2.157754 Prob(F-statistic) 0.000000 Inverted AR Roots.94 Dependent Variable: D(FBCFDELLRRQ_US) Date: 10/29/18 Time: 15:04 Sample (adjusted): 1991Q3 2018Q2 Included observations: 108 after adjustments D(FMBAD_US) 0.577335 0.094242 6.126070 0.0000 D(FBCFDELLRRQ_US(-1)) 0.478675 0.069707 6.866956 0.0000 R-squared 0.610194 Mean dependent var 0.005463 Adjusted R-squared 0.606516 S.D. dependent var 0.410556 S.E. of regression 0.257534 Akaike info criterion 0.143018 Sum squared resid 7.030344 Schwarz criterion 0.192688 Log likelihood -5.722996 Hannan-Quinn criter. 0.163157 Durbin-Watson stat 2.293152

Dependent Variable: DLOG(FCPIUEHC_US) Date: 10/26/18 Time: 08:51 Sample (adjusted): 1983Q2 2018Q2 Included observations: 141 after adjustments DLOG(FCPIUAH1_US) 0.984187 0.010205 96.44126 0.0000 R-squared 0.899204 Mean dependent var 0.008058 Adjusted R-squared 0.899204 S.D. dependent var 0.003359 S.E. of regression 0.001066 Akaike info criterion -10.84220 Sum squared resid 0.000159 Schwarz criterion -10.82129 Log likelihood 765.3750 Hannan-Quinn criter. -10.83370 Durbin-Watson stat 1.839412 @IDENTITY FGGDEBTGDP_US = FGTSOTQ_US / FGDP_US * 100 Dependent Variable: DLOG(FHVACHQ_US) Date: 11/01/18 Time: 20:20 Sample (adjusted): 1990Q1 2018Q2 Included observations: 114 after adjustments DLOG(@MOVAV(FHVACRQ_US,20)) 1.128636 0.579712 1.946890 0.0541 DLOG(FHOWNRQ_US(-1)) -0.773770 1.410498-0.548579 0.5844 R-squared 0.032160 Mean dependent var -0.001207 Adjusted R-squared 0.023519 S.D. dependent var 0.048793 S.E. of regression 0.048216 Akaike info criterion -3.208867 Sum squared resid 0.260375 Schwarz criterion -3.160863 Log likelihood 184.9054 Hannan-Quinn criter. -3.189385 Durbin-Watson stat 2.077522

Dependent Variable: DLOG(FSP500EQ_US) Date: 10/25/18 Time: 12:55 Sample (adjusted): 1960Q1 2018Q3 Included observations: 235 after adjustments DLOG(FSP500Q_US) 0.994451 0.053165 18.70505 0.0000 R-squared 0.581740 Mean dependent var 0.016531 Adjusted R-squared 0.581740 S.D. dependent var 0.079295 S.E. of regression 0.051282 Akaike info criterion -3.098698 Sum squared resid 0.615389 Schwarz criterion -3.083976 Log likelihood 365.0970 Hannan-Quinn criter. -3.092763 Durbin-Watson stat 2.795057 Dependent Variable: DLOG(FTWDMJRQ_US) Date: 10/30/18 Time: 23:38 Sample (adjusted): 1973Q2 2018Q3 Included observations: 182 after adjustments DLOG(FTWDBRD_US) 1.033357 0.044301 23.32560 0.0000 R-squared 0.750215 Mean dependent var -0.000783 Adjusted R-squared 0.750215 S.D. dependent var 0.031126 S.E. of regression 0.015557 Akaike info criterion -5.483193 Sum squared resid 0.043803 Schwarz criterion -5.465588 Log likelihood 499.9705 Hannan-Quinn criter. -5.476056 Durbin-Watson stat 0.743111