It is well known that the slope of the term structure of interest rates contains information for forecasting the likelihood of a recession in the US.1 Indeed, as shown by Figure 1, whenever three-month interest rates exceeded ten-year yields in the last 55 years of data (and with one exception), the US economy subsequently slipped into recession.
Figure 1 Recessions and the term spread, 1962-2018
It is for this reason that many market commentators are worried that the gradual flattening of the US term structure in recent months is indicative of an increased risk of a recession.2 This flattening, as on many previous occasions, reflects rising short-term rates as the Federal Reserve is in the process of ‘normalising’ monetary policy. Indeed, the information content of the slope of the term structure, or term spread, for future economic activity is typically seen as resulting from monetary policy.
For that very reason, it is interesting to explore whether the slope of the term structure of interest contains information about future economic growth before the founding of the Fed in 1913. If the information content arises because of monetary policy, then one would expect that the term structure of interest rates only contains information about future economic conditions after purposeful monetary policy was introduced.
In a series of interesting papers, Bordo and Haubrich (2004, 2008a, 2008b) study the information for future real GNP growth contained in the term structure, using quarterly US data starting in 1875. Since there are no data on treasury securities before the 1920s, they use yields on private sector securities which presumably contain credit risk premia. However, they present evidence suggesting that, at least in recent data, the slope of the term structure computed using private sector yields is very strongly correlated with the slope of the term structure of risk-free treasury securities and is only weakly correlated with the risk spreads of corporate and treasury bond yields.
From our perspective, their most interesting finding is that the slope of the term structure contained considerable information for future real GNP growth also before the establishment of the Federal Reserve. This suggests that activist monetary policy is not the only factor giving rise to the predictive content.
In a recent paper (Gerlach and Stuart 2018), we provide additional evidence on the information content of the term structure for future economic activity. Our work differs from that of Bordo and Haubrich in three important ways.
First, we ask whether the term structure forecasts NBER recessions rather than real economic growth. National accounting was only invented after WWI and attempts to construct historical time series for earlier periods have yielded varying results.3 Estimates of growth computed from national accounts statistics from the 19th and the early 20th centuries should therefore be seen as debatable estimates rather than as hard data. Recessions, which are defined by movements in a cross-section of time series correlated with economic activity, may well be subject to smaller measurement errors and easier to spot ex post in the data. It is therefore interesting to explore whether and, if so, how well the slope of the term structure predicts them.
Second, the NBER data start in 1857, whereas the estimates of real GNP start in 1875, enabling us to expand the pre-Fed period by 18 years. Looking at a longer time period is of independent interest, and lengthening the sample increases the precision of the statistical estimates.
Third, the NBER recession dummies are monthly in contrast to the real GNP data which are quarterly. Whether monthly data are more informative than quarterly data is an empirical question best settled by studying both data types.
Bordo and Haubrich study quarterly real GNP data from 1875 onwards. In this paper we study the monthly NBER recession dummies from 1857 onward. Figure 2 shows real GNP growth, using the Balke and Gordon (1986) data that Bordo and Haubrich use, and the NBER recession dates. The relationship between growth and recessions seems weaker before the 1890s, perhaps because of measurement errors in real GNP growth.
Figure 2 NBER recession dummy and real GNP growth, 1857-1913
Bordo and Haubrich note that t-bills were only authorised by Congress in 1929 and the yield on longer-term government debt, because of irregular issuance and its use as backing for bank notes, was not representative of market yields. Indeed, yields on short term US government securities are only available on the St Louis Fed’s FRED database since 1920, and long-term US bonds since 1919.
Bordo and Haubrich therefore use a commercial paper rate to capture the short-term interest rate and a corporate bond yield to capture long interest rates from Balke and Gordon (1986). We use a call money rate and the yield on high quality railroad bonds that are available on the FRED database from 1857 onwards. As Figure 3 shows, they are much lower than the yields used by Bordo and Haubrich but strongly correlated: the correlation between the short yields is 0.51 and the correlation between the long yields is above 0.99.
Figure 3 Interest rates, 1857-1913
We estimate probit models using data for the period January 1857 to November 1913 when the Federal Reserve was founded. The dependent variable is the NBER recession dummy between one and 18 months ahead.4 The results can be summarised as follows:
- A flattening of the term spread increases the risk of recession. This effect is statistically significant up to a forecast horizon of 13 months.
- Whether the economy is currently in a recession or not is useful for predicting the state of the business cycle in the future. For forecast horizons of less than 13 months, being in recession now raises the likelihood of being in recession in the future. For forecast horizons of 13 months or more, being in recession now reduces the likelihood that the economy will in recession in the future.
- Including the short interest rate as a regressor in addition to the term spread does not improve forecasting performance.
- The forecasting performance is highest at a forecast horizon of one month. It declines gradually to almost zero at a horizon of 13 months and then rises modestly.
- The slope of the term structure does not embody all available information about the likelihood of a recession, as evidenced by the fact that the change over 12 months of an index of American railroad stock prices is significant if included in the analysis.
Overall, the analysis suggests two conclusions.
First, the term structure contained information about the likelihood of a future recession even before the establishment of the Federal Reserve. That suggests that the information content does not arise solely as a consequence of countercyclical monetary policy.
Second, the results fully support the findings in Bordo and Haubrich (2004, 2008a and 2008b) for the pre-Fed period. This is noteworthy since their sample period starts later, they use quarterly rather than monthly data and different time series on interest rates, forecast future real GNP growth rather than the likelihood of a recession, and use OLS regressions rather than probit models.
Balke, N and R J Gordon, (1986), “Appendix B: historical data”, in R J Gordon (ed.), The American Business Cycle: Continuity and Change, NBER Studies in Business Cycles, vol. 25, University of Chicago Press.
Bordo, M D and J G Haubrich (2004), “The yield curve, recessions and the credibility of the monetary regime: Long-run evidence 1875-1997”, NBER Working Paper No. 10431.
Bordo, M D and J G Haubrich (2008a), “Forecasting with the yield curve: Level, slope and output 1875-1997”, Economics Letters 99: 48-50.
Bordo, M D and J G Haubrich (2008b), “The yield curve as a predictor of growth: Long-run evidence 1875-1997”, Review of Economics and Statistics 90: 182-185.
Ellison, Mand A Tischbirek(2018), “Beauty contests and the term structure”, VoxEU.org, 10 May.
Faria, Gand FVerona, (2018), “The yield curve and the stock market: Mind the long run”, VoxEU.org, 9 May.
Gerlach, S and R Stuart, (2018), “The Slope of the Term Structure and Recessions: The Pre-Fed Evidence, 1857-1913”, CEPR Discussion Paper 13013.
Goodhart, C (2007), “Does a downward-sloping yield curve predict a recession?”, VoxEU.org, 24 September.
Romer, C D (1988), “World War I and the Postwar Depression: A Reinterpretation based on alternative estimates of GNP”, Journal of Monetary Economics 22: 91-115.
Romer, C D (1989), “The Prewar Business Cycle Reconsidered: New Estimates of Gross National Product, 1869-1908”, Journal of Political Economy 97(1): 1-37.
 The term structure is a recurring theme in Vox pieces. See, for instance, Ellison and Tischbirek (2018), Faria and Verona (2018)and Goodhart (2007).
 See, for instance, reporting by Bloomberg and Reuters.
 Romer (1988, 1989) argues that measurement error is important for US national income estimates in the pre- and interwar periods.
 Since the forecast errors are overlapping, we compute standard errors that are robust to serial correlation.