We answer this question using data from the Japanese government bond (“JGB”) market dating back to 1974.
In February, we discussed “ Essential Model Validation for Interest Rate Risk and Asset and Liability Management ” using U.S. Treasury data for the model validation exercise. In another worked example, we focused on Government of Canada yields and addressed this question in the validation process: “Is there any version of a one factor term structure model which is sufficiently accurate for general risk management purposes and for asset and liability management?”
The short answer in Canada is no, and the short answer in Japan is also no.
We explain the model validation process in this note and shown the reasons for this strong conclusion, fortified with more than two decades of Japanese experience with extremely low and (sometimes) negative rates. Readers who want to see the difference between a best-practice Heath, Jarrow and Morton model and a common practice one factor model in a U.S. context are referred to this June 24 simulation analysis for the U.S. Treasury curve .
Defining “How Good is Good Enough?” for Interest Rate Risk Modeling
In our March 5, 2014 note “ Stress Testing and Interest Rate Risk Models: How Many Risk Factors are Necessary? ” we showed that nine interest rate risk factors were necessary for a best practice model of the U.S. Treasury curve. In a companion piece on March 18 titled “ Stress Testing and Interest Rate Risk Models: A Multi-Factor Stress Testing Example ,” we outlined the process for determining risk factors and the parameters used in a multi-factor interest rate model, again using U.S. Treasury data. In the Canadian case, we found that 12 factors were statistically significant in explaining movements in the Government of Canada yield curve.
In this note, we address the same three questions that we raised in the March 2015 piece, but this time the answers are derived for the Japanese government bond yield curve:
- How do you measure the accuracy of an interest rate risk simulation technique?
- Given that measure of accuracy, how many risk factors are necessary?
- How does accuracy change as the number of factors increases?
In answering the question “how good is good enough” for interest rate risk modeling, we follow the procedures that Bharath and Shumway (2008) used in testing the accuracy of the Merton model of risky debt versus the reduced form approach to credit risk modeling. We test these two hypotheses about one factor term structure models:
Strong form of hypothesis: One factor term structure models are so accurate that there are no other variables than the first factor that have statistically significant explanatory power.
Weaker form of hypothesis: There are other factors beyond the first factor that are statistically significant, but their impact is very modest and the benefits of using more than one factor are very minor.
Non-Parametric Tests of One Factor Term Structure Models
Jarrow, van Deventer and Wang (2003) (“JvDW”) provide another testing procedure that we address first. In examining the Merton model of risky debt, JvDW provide a very intuitive testing procedure that is independent of the parameters fitted to the model structure. They asked this question: “Are the implications of the model true or false?” Since no model is perfect, they answer this question with a probability.
We again address two classes of one factor term structure models in this section:
One factor models with rate-dependent interest rate volatility;
Cox, Ingersoll and Ross (1985)
Black, Derman and Toy (1990)
Black and Karasinski (1991)
One factor models with constant interest rate volatility (affine models)
Vasicek (1977)
Ho and Lee (1986)
Extended Vasicek or Hull and White Model (1990, 1993)
Non-parametric test 1: The one factor models with rate-dependent interest rate volatility make it impossible for interest rates to be negative. Is this implication true or false? It is false. According to the Japan Ministry of Finance, there have been negative rates in Japanese government bill auctions at 2 months (once), 3 months (10 times), and 6 months (6 times) between April 7, 1999 and July 9, 2015. The Japan Ministry of Finance also reports on secondary market yields for maturities of 1 year or more on a daily basis. Negative yields have been reported for maturities of 1 year (49 days), 2 years (60 days), 3 years (32 days), and 4 years (15 days) through July 13, 2015. For this reason alone, we advise analysts to reject the one factor rate-dependent volatility models as inconsistent with historical facts.
Non-parametric test 2: The Vasicek, Ho and Lee, and Extended Vasicek/Hull and White models assume that interest rate volatility is a constant, independent of the level of interest rates. This assumption implies that both the level and the changes in interest rates are normally distributed over time. We use quarterly data on Japanese government bond yields provided by the Japan Ministry of Finance from September 24, 1974 through July 13, 2015. We extract zero coupon bond yields from this data using Kamakura Risk Manager version 8.1 and maximum smoothness forward rate smoothing. This graph shows the quarterly evolution of Japanese government bond zero coupon yields over time:
This graph shows the evolution of the first quarterly forward rate (the forward that applies from the 91^{st} day through the 182^{nd} day) over the same time period:
We use three statistical tests to determine whether or not the hypothesis of normality should be rejected at the 5% level for two sets of data:
- The absolute level of zero coupon bond yields over the 1974 to 2015 time period
- The quarterly changes in 160 quarterly forward rates making up the 40 year Japanese government bond yield curve.
The statistical tests we use include the Shapiro-Wilk test, theShapiro-Francia test, and the skew test, all of which are available in common statistical packages.
The chart above shows the p-values for these three statistical tests for the first twelve quarterly maturities. The null hypothesis of normality is rejected by all 3 tests for 90 of the 160 quarterly zero coupon yield maturities. For quarterly changes in forward rates, the null hypothesis of normality is rejected by all 3 tests for 104 of the 160 maturities.
Non-parametric test 3: As commonly implemented, one factor term structure models imply that all yields will either (a) rise, (b) fall, or (c) remain unchanged. In Chapter 3 of Advanced Financial Management (second edition, 2013), van Deventer, Imai and Mesler show that this implication of one factor term structure models is rarely true in the U.S. Treasury market. We perform the same test using 10,540 days of zero coupon bond yields for the Japanese government bond yield curve. We analyze the daily shifts in the 480 different monthly zero coupon bond yields on each day. The results are given here:
The results were rarely consistent with the implications of one factor term structure models. Yield curve shifts were all positive, all negative, or all zero 7.15%, 0.56%, and 0% of the time, a total of 7.71% of all business days. The predominant yield curve shift was a twist, with a mix of positive changes, negative changes, or zero changes. These figures are similar to those for the U.S. Treasury and Government of Canada yield curves. These twists, which happen 92.29% of the time in Japan, cannot be modeled at all with one factor term structure models.
Non-parametric test 4 : A closely related test is discussed in Chapter 3 of van Deventer, Imai and Mesler. One factor term structure models cannot create a yield curve that has multiple humps in it. One simply has to count the humps in the Japanese government bond yield curve to show that this is another serious problem with one factor term structure models:
The number of days with 0 or 1 humps (defined as the sum of local minima and maxima on that day’s yield curve) was 41.54% of the total observations in the data set. The remainder of the data set, 58.46% of the total, has yield curves with shapes that are inconsistent with a one factor term structure model.
Fitting a Multi-Factor Heath Jarrow and Morton Term Structure Model to Japanese Government Bond Yields
Given the poor performance of one factor models on the non-parametric tests above, it is no surprise that bank regulators are turning to multi-factor models around the world. The Federal Reserve’s Comprehensive Capital Analysis and Review stress testing regime has included 3 points on the U.S. Treasury yield curve since 2014. The Bank for International Settlements has required that at least six interest rate risk factors be used to model market risk since 2010. Adrian, Crump and Moench of the Federal Reserve Bank of New York use five factors in their U.S. Treasury term structure model.
We now fit a multi-factor Heath, Jarrow and Morton model to quarterly Japanese government bond zero coupon yield data from 1974 to June 30, 2015. The availability of Japanese government bond par coupon bond yields varied over time in accordance with the following summary:
The availability of data out to 40 years is fairly unique in government bond markets world-wide. Japanese government bill auction results are reported by the Japan Ministry of Finance beginning in April 1999. The procedures used are described in detail in these documents:
Jarrow, Robert A. and Donald R. van Deventer, “Parameter Estimation for
Heath, Jarrow and Morton Term Structure Models,” Technical Guide, Version 2.0, Kamakura Corporation, June 30, 2015.
Jarrow, Robert A. and Donald R. van Deventer, Appendix C, Version 2.0: “Japanese Government Bond Yields,” to “Parameter Estimation for Heath, Jarrow and Morton Term Structure Models,” Technical Guide, Kamakura Corporation, June 30, 2015.
Jarrow, Robert A. and Donald R. van Deventer, “Monte Carlo Simulation in a Multi-Factor Heath, Jarrow and Morton Term Structure Model,” Technical Guide, Version 4.0, Kamakura Corporation, June 16, 2015.
We followed these steps to estimate the parameters of the model:
- We extract the zero coupon yields and zero coupon bond prices for all quarterly maturities out to forty years for all daily observations. This is done using Kamakura Risk Manager, version 8.1, using the maximum smoothness forward rate approach.
- We drop the daily observations that are not the last observation of the quarter, to avoid overlapping quarterly observations and the resulting autocorrelated errors that would stem from that.
- We calculate the continuously compounded changes in forward returns as described in the parameter technical guide.
- We drop the first observation after a change in the Japan Ministry of Finance data regime, because a change in data availability can impact the shape of the full yield curve.
- We then begin the process of creating the orthogonalized risk factors that drive interest rates. These factors are assumed to be uncorrelated independent random variables that have a normal distribution with mean zero and standard deviation of 1.
- In the estimation process, we added factors to the model as long as each new factor provided incremental explanatory power.
We use the resulting parameters and accuracy tests to address the hypothesis that a one factor model is “good enough” for modeling Japanese government bond yields.
Proof That One Factor Models Are Not Sufficient for Best Practice Risk Management
We now test the hypotheses about one factor term structure models.
Strong form of hypothesis: One factor term structure models are so accurate that there are no other variables than the first factor that have statistically significant explanatory power.
The following graph shows that a one factor term structure model omits a very large number of risk factors driving Japanese government bond yields:
Other than the first quarterly forward rate in the yield curve, there are as many as 25 explanatory variables that drive the 160 quarterly segments of the yield curve. The final Japanese government bond term structure model from Kamakura Risk Information services has 16 independent risk factors that drive yields, and these factors also appear in combination with rate level variables. A total of 35 related candidate explanatory variables were used in the estimation process.
Conclusion: The strong form of the hypothesis is overwhelmingly rejected by the data on Japanese government bond yields.
We now turn to the weaker hypothesis.
Weaker form of hypothesis: There are other factors beyond the first factor that are statistically significant, but their impact is very modest and the benefits of using more than one factor are very minor.
To address this hypothesis, we graph the adjusted r-squared of a “regime change” one factor model which combines normally distributed and rate dependent one factor models with a best practice model which includes all statistically significant factors. The results are shown here:
The adjusted r-squared for the best practice model is plotted in blue and is near 100% for nearly all 160 quarterly segments of the yield curve. The one factor model, by contrast, does a stunningly poor job of fitting quarterly movements in the quarterly forward rates. The adjusted r-squared is good, of course, for the first forward rate since the short rate is the standard risk factor in a one factor term structure model. Beyond the first quarter, however, explanatory power is extremely low. The adjusted r-squared of the one factor model never exceeds 45% after the first forward and is far below that level at most maturities. The changes in the shape of both r-squared lines are in part due to changes in the Japan Ministry of Finance data regime.
This result should not come as a surprise to a serious analyst, because it is very similar to the results of the best practice Heath, Jarrow and Morton term structure model for U.S. Treasuries and Government of Canada yields.
We can confirm the low explanatory power of a one factor model with a one line principal components analysis in a common statistical package. The “PCA” analysis is not constrained to choose the short rate as the explanatory variable in a one factor model. In fact, there are many other factors that would be stronger candidates for a single factor model. The results of the principal components analysis on quarterly movements in Japanese government bond yields are shown here:
The results show that at least 16 or 17 factors are needed to model the Japanese government bond yield curve with cumulative accuracy comparable to the confidence levels most large financial institutions would use for value at risk analysis. The first factor explains only 54% of quarterly forward rate movements. Readers should beware of the fact that principal components analysis uses only that part of the data set for which data exists at all maturities tested.
Conclusion: The weak form of the hypothesis is also overwhelmingly rejected by the data on Japanese government bond yields.
Can a One Factor Model be “Tweaked” with One or Two More Factors?
Many large financial institutions have been using one factor models for such a long time that hope springs eternal that they can be fixed with a small “tweak,” a second or third factor. In the next graph we show the adjusted r-squareds for 1, 2, 3, 6 and “all” factors in a model of the Japanese government bond yield curve.
The results show that even a six factor model leaves a big chunk of yield curve movements unexplained. We note again that some of the sharp movements in r-squareds are due to changes in Japan Ministry of Finance data regime. We speculate that yield curve smoothing techniques used by the Ministry of Finance may also contribute to some of the changes in r-squareds. In short, however, in the 21^{st} century, with modern big data technology, using all factors that matter, instead of just a few of them, is a simple step forward.
A similar plot of the root mean squared errors for 1 factor, 2 factor, 3 factor, 6 factor, and all factor term structure models shows the danger of half steps in improving interest rate risk technology:
We close with this plot of which maturities on the yield curve are statistically significant in predicting forward rate movements at each of the 160 quarterly segments on the 40 year Japanese government bond yield curve. Statistical significance is represented by a dot at the combination of yield curve risk factor (by maturity, on the vertical axis) and quarterly forward rate number. The lack of a dot means that risk factor maturity is not statistically significant. An orange dot represents interest rate volatility that is constant or “affine.” A green dot represents interest rate volatility that is proportional to the level of interest rates. A blue dot represents interest rate volatility that is linear, combining both constant and proportional impacts on interest rate volatility.
At shorter term forward rates, the linear specification for interest rate volatility is the dominant specification. As maturities lengthen to maturities where low rate experience is very limited, the measured interest rate volatility is orange, or constant. We caution readers, especially those in high interest rate environments, that the constant volatility result is very likely to be rejected as experience with low rates becomes more common.
Note also that a “regime change” one factor model includes only those statistically significant variables on the bottom row of the chart. The explanatory power of such a model is very low because the variables on all of the other rows have been omitted.
Moving Forward with Modern Interest Rate Risk Technology
Kamakura Corporation facilitates client progress in interest rate modeling in multiple ways via Kamakura Risk Information Services’ Macro Factor Sensitivity Products:
Research subscriptions to Heath, Jarrow and Morton term structure modeling
This is a good first step for regulatory agencies and financial institutions building their familiarity with modern interest rate risk technology. The subscription includes the Technical Guides describing the parameter estimation process, the underlying raw data, and the parameters themselves, updated annually. Models are available for all major government yield curves.
Production subscription to Heath, Jarrow and Morton term structure modeling
The production subscription includes formatting of parameters for use in Kamakura Risk Manager’s newest versions and immediate release of HJM parameter estimates as soon as the quality control process at Kamakura Corporation is completed.
Production subscription to HJM yield scenarios
Kamakura Risk Information Services also generates the scenarios in-house and provides the scenarios in standard Kamakura Risk Manager format for all major government yield curves at customized frequencies (daily, weekly, monthly, quarterly) for individual clients. Transfer of data is by file transfer protocol technology.
Heath, Jarrow and Morton Training
Kamakura Corporation, led by Managing Director Robert A. Jarrow (Cornell University) provides training in modern Heath, Jarrow and Morton interest rate risk technology for both clients and potential clients. Professor Jarrow usually participates by video link in these training sessions.
For inquiries about these and other products, please contact your Kamakura representative or e-mail Kamakura at info@kamakuraco.com.
Further Reading for the Technically Inclined Reader
References for random interest rate modeling are given here:
Heath, David, Robert A. Jarrow and Andrew Morton, "Bond Pricing and the Term Structure of Interest Rates: A Discrete Time Approach," Journal of Financial and Quantitative Analysis, 1990, pp. 419-440.
Heath, David, Robert A. Jarrow and Andrew Morton, "Contingent Claims Valuation with a Random Evolution of Interest Rates," The Review of Futures Markets, 9 (1), 1990, pp.54 -76.
Heath, David, Robert A. Jarrow and Andrew Morton, ”Bond Pricing and the Term Structure of Interest Rates: A New Methodology for Contingent Claim Valuation,” Econometrica, 60(1), 1992, pp. 77-105.
Heath, David, Robert A. Jarrow and Andrew Morton, "Easier Done than Said", RISK Magazine, October, 1992.
References for non-parametric methods of model testing are given here:
Bharath, Sreedhar and Tyler Shumway, "Forecasting Default with the Merton Distance to Default Model," Review of Financial Studies, May 2008, pp. 1339-1369.
Jarrow, Robert, Donald R. van Deventer and Xiaoming. Wang, “A Robust Test of Merton’s Structural Model for Credit Risk,” Journal of Risk, Fall 2003, pp. 39-58.
References for modeling traded securities (like bank stocks) in a random interest rate framework are given here:
Amin, Kaushik and Robert A. Jarrow, "Pricing American Options on Risky Assets in a Stochastic Interest Rate Economy," Mathematical Finance, October 1992, pp. 217-237.
Jarrow, Robert A. “Amin and Jarrow with Defaults,” Kamakura Corporation and Cornell University Working Paper, March 18, 2013.
The behavior of credit spreads when interest rates vary is discussed in these papers:
Campbell, John Y. & Glen B. Taksler, "Equity Volatility and Corporate Bond Yields," Journal of Finance, vol. 58(6), December 2003, pages 2321-2350.
Elton, Edwin J., Martin J. Gruber, Deepak Agrawal, and Christopher Mann, “Explaining the Rate Spread on Corporate Bonds,” Journal of Finance, February 2001, pp. 247-277.