Skip to main content


Log in

Bayesian estimation of the long-run trend of the US economy

  • Published:
Empirical Economics Aims and scope Submit manuscript


The main purpose of this paper is to scrutinize the long-run trend of the US real GDP during the post-war period. In the empirical analysis, we introduce multivariate unobserved components models that accommodate time-varying volatility bounded by an economically reasonable range. After accounting for the cointegration relationship among the real GDP, consumption, and investment, we find the following. (i) There was an abrupt and significant downturn in the long-run growth rate in 2007. (ii) The annualized long-run growth rate of the real GDP declined to approximately 1.69%, decreasing from a peak of nearly 3.21% during the 2000s. (iii) The bounded volatility assumption enables a trend-cycle decomposition of the US real GDP that matches the NBER’s recession dates.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Price excludes VAT (USA)
Tax calculation will be finalised during checkout.

Instant access to the full article PDF.

Fig. 1
Fig. 2
Fig. 3
Fig. 4
Fig. 5
Fig. 6
Fig. 7

Similar content being viewed by others


  1. Gerlach et al. (2000) and Huber et al. (2019) design time-varying parameters by using mixture innovations. The merit of their approaches is that both gradual and abrupt changes in the model parameters can be estimated within a unified model framework. For the estimation, Gerlach et al. (2000) propose to integrate all latent state variables, except for those associated with the mixture innovations, by employing a Kalman filter-based algorithm. The estimation algorithm described by Gerlach et al. (2000) is computationally demanding due to the integration step. Huber et al. (2019) developed a new model and a corresponding estimation algorithm to address the computational bottleneck. In future research, it could be worthwhile to model \(\mu _t\) with the mixture innovation models reported by Gerlach et al. (2000) and Huber et al. (2019).

  2. In addition to the aforementioned studies, Sinclair (2009), who studied how US output and the unemployment rate are related in the short run and long run, assume that individual cycle components follow AR(2) processes.

  3. The trend-cycle decomposition of the US real GDP is highly sensitive to the zero correlation assumption. See Morley et al. (2003) for empirical evidence.

  4. Notably, conventional Bayesian estimation methods applied to stochastic volatility models, such as the model developed by Kim et al. (1998), are not applicable to our nonlinear state-space model due to the inequality restriction on the volatility process in Eq. (9).

  5. This study uses quarterly data because the U.S. Bureau of Economic Analysis (, which announces the US national accounting data, does not provide monthly observations of the real GDP and investment.

  6. We verified that priors more diffuse than those used in this study do not change the main empirical results.

  7. The mean of the inverse Wishart distribution is \( \mathrm{diag}(0.7^2,0.7^2,0.7^2,3^2 ) \over 6 - 4 - 1 \).

  8. The expected duration of each state is computed by \(1 \over {1 - \pi }\), where \(\pi \) represents the corresponding transition probability. The standard deviation of the prior beta distribution is \(0.0139 = \sqrt{ (98 + 2) \over (98 + 2)^2 (98 + 2 +1) }\). Thus, at approximately 68% probability, the prior covers a transition probability of [0.966, 0.994] and an expected state duration of \([30 \ \text {quaters},166 \ \text {quaters}]\).

  9. The annualized standard deviation of the real consumption growth during the 1970s was 1.87%.

  10. The empirical result shown in Table 1 is robust to using different initial values.

  11. We evaluate \(g(Y_{\tau +h} \vert \theta _M^{(r)}, Y_{1:\tau },M)\) using the Kalman filter algorithm. In the computation, \(\Sigma _\tau ^{(r)}\) was also included in the conditional information set. For notation simplicity, we suppress the time-varying covariance matrix in the likelihood function.

  12. In conventional structural break models, one should decide in advance how many structural breaks exist before and after the prediction period. To overcome this drawback, Koop and Potter (2007) developed an advanced model that does not need to know the number of breaks prior to the estimation and prediction. Interested readers should refer to Koop and Potter (2007) and the references therein.

  13. The posterior means of the MUC-SB1 and MUC-SB2 models are similar to the estimates reported in other related studies, such as Gordon (2014) and Fernald et al. (2016), who predict long-run growth of approximately 1.5%. Antolin-Diaz et al. (2017) estimate the time-varying growth rate of the real GDP using a dynamic factor model with a large data set. These authors claim that the long-run growth rate is near 2%, which is similar to the result of the MUC-RW model.

  14. One main obstacle to obtaining reliable estimates of the structural breaks lies in the lack of information regarding the low-frequency trend component. Bai et al. (1998) theoretically demonstrate that the confidence interval of a structural break date does not decrease even asymptotically as the sample size increases; however, the confidence interval is inversely related to the dimension of time-series data that share a common break point.

  15. Smoothed estimates at time t can be significantly affected by events in subsequent time periods.

  16. Fernald et al. (2016) argues that low productivity growth is the main reason for the recent slow growth of the US economy and is expected to last over the long run because the high growth in education attainment that underlies high productivity is no longer achievable.

  17. Notably, \(m_t = \mu _t\) in MUC-RW and \(m_t = s_t\) in MUC-SB1 and MUC-SB2.

  18. \(g_{\theta }(Y_t \vert \beta _{t}^{(n)}, m_{t}^{(n)}, h_{t}^{(n)}) = 1\) under the proposed sampling scheme. No observations are available at time \(t=0\). Therefore, a user-specific importance distribution or an unconditional distribution can be used to generate all elements of \(\beta _0\) at \(t=0\).

  19. We confirm via simulations that the Bayesian method developed here performs well as a practical matter even under the degenerate transition and measurement equations of our UC models and provides reliable estimates of the break date and other latent variables.


  • Andrieu C, Doucet A, Holenstein R (2010) Particle Markov chain Monte Carlo methods. J R Stat Soc Ser B (Stat Methodol) 72:269–342

    Article  Google Scholar 

  • Antolin-Diaz J, Drechsel T, Petrella I (2017) Tracking the slowdown in long-run GDP growth. Rev Econ Stat 99:343–356

    Article  Google Scholar 

  • Bai J, Lumsdaine RL, Stock JH (1998) Testing for and dating common breaks in multivariate time series. Rev Econ Stud 65:395–432

    Article  Google Scholar 

  • Berger T, Everaert G, Vierke H (2016) Testing for time variation in an unobserved components model for the US economy. J Econ Dyn Control 69:179–208

    Article  Google Scholar 

  • Carter CK, Kohn R (1994) On Gibbs sampling for state space models. Biometrika 81:541–553

    Article  Google Scholar 

  • Chan JCC, Grant AL (2016) On the observed-data deviance information criterion for volatility modeling. J Financial Econometrics 14:772–802

    Article  Google Scholar 

  • Chan JCC, Koop G, Potter SM (2016) A bounded model of time variation in trend inflation, NAIRU and the Phillips curve. J Appl Econometrics 31:551–565

    Article  Google Scholar 

  • Chib S (1995) Marginal likelihood from the Gibbs output. J Am Stat Assoc 90:1313–1321

    Article  Google Scholar 

  • Chib S (1998) Estimation and comparison of multiple change-point models. J Econometrics 86:221–241

    Article  Google Scholar 

  • Del Negro M, Giannone D, Giannoni MP, Tambalotti A (2019) Global trends in interest rates. J Int Econ 118:248–262

    Article  Google Scholar 

  • Fernald J et al (2016) What is the new normal for US growth? FRBSF Econ Lett 30

  • Fernald JG, Stock JH, Hall RE, Watson MW (2017) The disappointing recovery of output after 2009. Brookings papers on economic activity

  • Frühwirth-Schnatter S (1994) Data augmentation and dynamic linear models. J Time Ser Anal 15:183–202

    Article  Google Scholar 

  • Gerlach R, Carter C, Kohn R (2000) Efficient Bayesian inference for dynamic mixture models. J Am Stat Assoc 95:819–828

    Article  Google Scholar 

  • Geweke J, Amisano G (2010) Comparing and evaluating Bayesian predictive distributions of asset returns. Int J Forecast 26:216–230

    Article  Google Scholar 

  • Gordon RJ (2014) The demise of US economic growth: restatement, rebuttal, and reflections. Technical report, National Bureau of Economic Research

  • Grant AL, Chan JCC (2017a) A Bayesian model comparison for trend-cycle decompositions of output. J Money Credit Bank 49:525–552

  • Grant AL, Chan JCC (2017b) Reconciling output gaps: unobserved components model and Hodrick–Prescott filter. J Econ Dyn Control 75:114–121

  • Hall RE (1978) Stochastic implications of the life cycle-permanent income hypothesis: theory and evidence. J Polit Econ 86:971–987

    Article  Google Scholar 

  • Holston K, Laubach T, Williams JC (2017) Measuring the natural rate of interest: international trends and determinants. J Int Econ 108:S59–S75

    Article  Google Scholar 

  • Huang Y-F, Luo S, Startz R (2016) Are recoveries all the same: GDP and TFP? Available at SSRN 2988263

  • Huber F, Kastner G, Feldkircher M (2019) Should I stay or should I go? A latent threshold approach to large-scale mixture innovation models. J Appl Econometrics 34:621–640

    Article  Google Scholar 

  • Kamber G, Morley J, Wong B (2018) Intuitive and reliable estimates of the output gap from a Beveridge–Nelson filter. Rev Econ Stat 100:550–566

    Article  Google Scholar 

  • Kim C-J (2008) Markov-switching and the Beveridge–Nelson decomposition: has US output persistence changed since 1984? J Econometrics 146:227–240

    Article  Google Scholar 

  • Kim J, Chon S (2018) Why are Bayesian trend-cycle decompositions of US real GDP so different? Empir Econ 1–16

  • Kim C-J, Kim J (2020) Trend-cycle decompositions of real GDP revisited: classical and Bayesian perspectives on an unsolved puzzle. Macroecon Dyn 1–25

  • Kim C-J, Nelson CR (1998) Business cycle turning points, a new coincident index, and tests of duration dependence based on a dynamic factor model with regime switching. Rev Econ Stat 80:188–201

    Article  Google Scholar 

  • Kim C-J, Nelson CR et al (1999) State-space models with regime switching: classical and Gibbs-sampling approaches with applications, vol 1. MIT Press Books, Cambridge

    Google Scholar 

  • Kim C-J, Piger J (2002) Common stochastic trends, common cycles, and asymmetry in economic fluctuations. J Monet Econ 49:1189–1211

    Article  Google Scholar 

  • Kim S, Shephard N, Chib S (1998) Stochastic volatility: likelihood inference and comparison with ARCH models. Rev Econ Stud 65:361–393

    Article  Google Scholar 

  • Koop G, Potter SM (2007) Estimation and forecasting in models with multiple breaks. Rev Econ Stud 74:763–789

    Article  Google Scholar 

  • Lindsten F, Jordan MI, Schön TB (2014) Particle Gibbs with ancestor sampling. J Mach Learn Res 15:2145–2184

    Google Scholar 

  • Lindsten F, Bunch P, Singh SS, Schön TB (2015) Particle ancestor sampling for near-degenerate or intractable state transition models. arXiv:1505.06356

  • Luo S, Startz R (2014) Is it one break or ongoing permanent shocks that explains US real GDP? J Monet Econ 66:155–163

    Article  Google Scholar 

  • McConnell MM, Perez-Quiros G (2000) Output fluctuations in the United States: what has changed since the early 1980’s? Am Econ Rev 90:1464–1476

    Article  Google Scholar 

  • Morley J, Piger J (2012) The asymmetric business cycle. Rev Econ Stat 94:208–221

    Article  Google Scholar 

  • Morley JC, Nelson CR, Zivot E (2003) Why are the Beveridge–Nelson and unobserved-components decompositions of GDP so different? Rev Econ Stat 85:235–243

    Article  Google Scholar 

  • Perron P (1989) The great crash, the oil price shock, and the unit root hypothesis. Econometrica J Econometric Soc 1361–1401

  • Perron P, Wada T (2009) Let’s take a break: trends and cycles in US real GDP. J Monet Econ 56:749–765

    Article  Google Scholar 

  • Sinclair TM (2009) The relationships between permanent and transitory movements in US output and the unemployment rate. J Money Credit Bank 41:529–542

    Article  Google Scholar 

  • Stock JH, Watson MW (1989) New indexes of coincident and leading economic indicators. NBER Macroecon Annu 4:351–394

    Article  Google Scholar 

  • Stock JH, Watson MW (2002) Has the business cycle changed and why? NBER Macroecon Annu 17:159–218

    Article  Google Scholar 

Download references

Author information

Authors and Affiliations


Corresponding author

Correspondence to Jaeho Kim.

Ethics declarations

Conflicts of interest

The authors declare that they have no conflicts of interest.

Ethical approval

This article does not contain any studies involving human participants or animals performed by any of the authors.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Appendix A: Bayesian estimation

Appendix A: Bayesian estimation

The proposed UC model is cast into the following state space representation to gain the joint smoothing distribution of the latent state variables:

$$\begin{aligned} Y_t= & {} D + H\beta _{t} \end{aligned}$$
$$\begin{aligned} \beta _t= & {} {\tilde{\mu }}_{t} + F\beta _{t-1} + G\epsilon _t, \end{aligned}$$

where \(Y_t = [ y_t \; c_t \; i_t]'\), \(D = [0 \; {\bar{c}} \; {\bar{i}}]'\), \(\beta _t = [x_t \; z_t \; z_{t-1} \; g_{c,t} \; g_{c,t-1} \; g_{i,t} \; g_{i,t-1}]'\), \({\tilde{\mu }}_{t} = [\mu _{t} \; 0 \; 0 \; 0 \; 0 \; 0 \; 0]'\), \(\epsilon _t = [u_{t} \; e_t \; \eta _{c,t}\; \eta _{i,t}]'\),

$$\begin{aligned} H = \begin{bmatrix} 1 &{} 1 &{} 0 &{} 0 &{} 0 &{} 0 &{} 0\\ \gamma _c &{} 0 &{} 0 &{} 1 &{} 0 &{} 0 &{}0 \\ \gamma _i &{} 0 &{} 0 &{} 0 &{} 0 &{} 1 &{}0 \end{bmatrix}, \ F = \begin{bmatrix} 1 &{} 0 &{} 0 &{} 0 &{} 0 &{} 0 &{} 0 \\ 0 &{} \phi _1 &{} \phi _2 &{} 0 &{} 0 &{} 0 &{} 0 \\ 0 &{} 1 &{} 0 &{} 0 &{} 0 &{} 0 &{} 0 \\ 0 &{} 0 &{} 0 &{} \gamma _{c,1} &{} \gamma _{c,2} &{} 0 &{} 0 \\ 0 &{} 0 &{} 0 &{} 1 &{} 0 &{} 0 &{} 0 \\ 0 &{} 0 &{} 0 &{} 0 &{} 0 &{} \gamma _{i,1} &{} \gamma _{i,2} \\ 0 &{} 0 &{} 0 &{} 0 &{} 0 &{} 1 &{} 0 \end{bmatrix}, G = \begin{bmatrix} 1 &{} 0 &{} 0 &{} 0\\ 0 &{} 1 &{} 0 &{} 0 \\ 0 &{} 0 &{} 0 &{} 0 \\ 0 &{} 0 &{} 1 &{} 0 \\ 0 &{} 0 &{} 0 &{} 0 \\ 0 &{} 0 &{} 0 &{} 1 \\ 0 &{} 0 &{} 0 &{} 0\end{bmatrix}. \end{aligned}$$

Throughout this appendix, we adhere to the following notations of the latent variables: \(\beta _{\kappa :\tau } = \{\beta _\kappa ,\beta _{\kappa +1},\ldots ,\beta _\tau \}\), and \(m_{\kappa :\tau } = \{m_\kappa ,m_{\kappa +1},\ldots ,m_\tau \}\) for \(\tau \ge \kappa \). Regarding the stochastic volatility processes of the shocks in \(\epsilon _t\), we use \(h_{\kappa :\tau } = \{h_\kappa ,h_{\kappa +1},\ldots ,h_\tau \}\), where \(h_t = \{h_{u,t},h_{e,t},h_{\eta _c,t},h_{\eta _i,t}\}\). For MUC-RW, \(m_t = \mu _t\) is assumed to follow the random walk process in Eq. (5). In MUC-SB1 and MUC-SB2, \(m_t = s_t\) is assumed to follow the structural break process in Eq. (6).

1.1 Appendix A.1: posterior simulation of the model parameters

Kim and Chon (2018) show that in univariate UC models, the variance in the permanent shock could be severely biased if the autoregressive and long-run mean growth rate parameters are sampled separately without considering the correlation between the permanent and transitory shocks. Therefore, we sample the model parameters of our multivariate UC models simultaneously using their joint posterior distribution as described in Kim and Chon (2018). In this appendix, we focus on the estimation of MUC-SB1 and MUC-SB2. We can easily estimate the model parameters of MUC-RW by slightly modifying the algorithm explained in this appendix.

Consider the following seemingly unrelated regression (SUR) model derived from the transition equations of the proposed UC model conditional on \(\{\beta _{0:T},s_{0:T},h_{0:T} \}\):

$$\begin{aligned} \begin{bmatrix} \Delta x_t^* \\ z_t^* \\ c_t^* \\ i_t^*\end{bmatrix}&= \begin{bmatrix} I_{s_t =1}^* &{}I_{s_t =2}^* &{} \cdots &{} I_{s_t=K}^*&{} 0 &{} 0 &{} 0 &{} 0 &{} 0 &{} 0 \\ 0 &{} 0 &{} \cdots &{} 0 &{} z_{t-1}^* &{} z_{t-2}^* &{} 0 &{} 0 &{} 0 &{} 0 \\ 0 &{} 0 &{} \cdots &{} 0 &{} 0 &{} 0 &{} g_{c,t-1}^* &{} g_{c,t-2}^* &{} 0 &{} 0 \\ 0 &{} 0 &{} \cdots &{} 0 &{} 0 &{} 0 &{} 0 &{} 0 &{} g_{i,t-1}^* &{} g_{i,t-2}^* \end{bmatrix} \begin{bmatrix} \mu _1 \\ \mu _2 \\ \vdots \\ \mu _K \\ \phi _1 \\ \phi _2 \\ \gamma _{c,1} \\ \gamma _{c,2} \\ \gamma _{i,1} \\ \gamma _{i,2} \end{bmatrix} \nonumber \\&\qquad + \begin{bmatrix} u_{t}^* \\ e_{t}^* \\ \eta _{c,t}^* \\ \eta _{i,t}^* \end{bmatrix}, \end{aligned}$$
$$\begin{aligned}&({\tilde{Y}}_t = {\tilde{X}}_t \Gamma + {\tilde{\epsilon }}_t), \nonumber \\&\quad \begin{bmatrix} u_{t}^* \\ e_{t}^* \\ \eta _{c,t}^* \\ \eta _{i,t}^* \end{bmatrix} \sim N\left( \begin{bmatrix} 0 \\ 0 \\ 0 \\ 0 \end{bmatrix}, \begin{bmatrix} \sigma ^2_{u,1} &{} \rho _{u,e} \sigma _{u,1}\sigma _{e,1} &{} \rho _{u,\eta _c} \sigma _{u,1}\sigma _{\eta _{c},1} &{} \rho _{u,\eta _i} \sigma _{u,1}\sigma _{\eta _i,1} \\ \rho _{u,e} \sigma _{e,1}\sigma _{u,1} &{} \sigma ^2_{e,1} &{} \rho _{e,\eta _c} \sigma _{e,1}\sigma _{\eta _c,1} &{} \rho _{e,\eta _i} \sigma _{e,1}\sigma _{\eta _i,1} \\ \rho _{u,\eta _c} \sigma _{\eta _c,1}\sigma _{u,1} &{} \rho _{e,\eta _c} \sigma _{\eta _c,1}\sigma _{e,1} &{} \sigma ^2_{\eta _c,1} &{} \rho _{\eta _c,\eta _i} \sigma _{\eta _c,1}\sigma _{\eta _i,1} \\ \rho _{u,\eta _i} \sigma _{\eta _i,1}\sigma _{u,1} &{} \rho _{e,\eta _i} \sigma _{\eta _i,1}\sigma _{e,1} &{} \rho _{\eta _c,\eta _i} \sigma _{\eta _c,1}\sigma _{\eta _i,1} &{} \sigma ^2_{\eta _i,1} \end{bmatrix}\right) , \nonumber \\&\quad ({\tilde{\epsilon }}_t \sim N(0, \Sigma _1)), \end{aligned}$$

where \(\Gamma = [\mu _1 \; \mu _2 \; \ldots \; \mu _K \; \phi _1 \; \phi _2 \; \gamma _{c,1} \; \gamma _{c,2} \; \gamma _{i,1} \; \gamma _{i,2}]'\); \(\Delta x_t^* = {(x_t - x_{t-1})\over \exp (h_{u,t}/2)}\); \(I_{s_t=k}^* = {I(s_t = k)\over \exp (h_{u,t}/2)}\) for \(k=1,2,\ldots ,K\); \(z_{t-j}^* = {z_{t-j}\over \exp (h_{e,t}/2)}, \ g_{c,t-j}^* = {g_{c,t-j}\over \exp (h_{\eta _c,t}/2)}, \ g_{i,t-j}^* = {g_{i,t-j}\over \exp (h_{\eta _i,t}/2)} \) for \(j=0,1,2\). With a normal prior distribution of \(\Gamma \) and an inverse Wishart prior of \(\Sigma _1\), we can obtain the conditional posterior distribution of \(\Gamma \) as follows:

$$\begin{aligned} \Gamma \vert Y_{1:T},\beta _{1:T},s_{1:T}, \theta _{-\Gamma } \sim N({{\bar{\Gamma }}}, {{\bar{\Omega }}_{\Gamma }})I_{\Gamma }, \end{aligned}$$

where \({\bar{\Gamma }} = {\bar{\Omega }}_\Gamma ({\underline{\Omega }}_\Gamma ^{-1}{\underline{\Gamma }} + {\tilde{X}}'(\Sigma _0^{-1} \otimes I){\tilde{X}} {\hat{\Gamma }})\), \({\bar{\Omega }}_\Gamma = ({\underline{\Omega }}_\Gamma ^{-1} + {\tilde{X}}'(\Sigma _0^{-1} \otimes I){\tilde{X}})^{-1}\), and \({\hat{\Gamma }} = ({\tilde{X}}'(\Sigma _0^{-1} \otimes I){\tilde{X}})^{-1} {\tilde{X}}'(\Sigma _0^{-1} \otimes I){\tilde{Y}}\); \(\theta _{-\Gamma }\) is the set of all model parameters, except for \(\Gamma \). \(\{ {\underline{\Gamma }}, {\underline{\Omega }}_{\Gamma }\}\) are the prior mean and covariance matrices of \(\Gamma \). The data matrices \(\{{\tilde{X}},{\tilde{Y}}\}\) are obtained by vertically stacking \(\{{\tilde{X}}_t,{\tilde{Y}}_t\}\). The indication function \(I_{\Gamma }\) assumes a value of 1 if both the regime identification and stationarity conditions of \(\{\mu ,\phi ,\gamma _c,\gamma _i\}\) hold and 0 otherwise. We carry out the posterior simulation od \(\Gamma \) using the multivariate normal distribution given in Eq. (A.5) with rejection sampling.

Conditional on \(\{\beta _{0:T},s_{0:T},h_{1:T} \}\) and \(\Gamma \), it is relatively simple to draw other model parameters. Conditional on \(\Gamma \), the covariance matrix \(\Sigma _1\) at \(t=1\) is simulated from a posterior inverse Wishart distribution with a conjugate prior \(IW(\Lambda _0,\nu _0)\), where \(\Lambda _0\) and \(\nu _0\) are the prior scale and degree of freedom parameters. We simulate the posterior samples of the transition probabilities of \(s_t\) from posterior beta distributions. The samples of \(\{\sigma _{h,u}^2\), \(\sigma _{h,e}^2\), and \(\sigma _{h,\eta _c}^2, \sigma _{h,\eta _i}^2\}\) are drawn from posterior inverse gamma distributions using conventionally used conjugate priors. All posterior sampling steps are standard. Therefore, we do not describe the derivation of the posterior distributions here.

The sampling scheme of the cointegration parameters \(\{ {\bar{c}}, \lambda _c,{\bar{i}}, \lambda _i\}\) is based on the following Cholesky decomposition:

$$\begin{aligned} \begin{aligned} \begin{bmatrix} \eta _{c,t} \\ \eta _{i,t} \end{bmatrix}&= P A_t A_t^{-1} \begin{bmatrix} u_t \\ e_t\\ \eta _{c,t} \\ \eta _{i,t} \end{bmatrix} = P A_t \begin{bmatrix} \nu _{1,t} \\ \nu _{2,t} \\ \nu _{3,t} \\ \nu _{4,t} \end{bmatrix} \\&= \begin{bmatrix} A_{t,3,1} \ \nu _{1,t} + A_{t,3,2} \ \nu _{2,t} \\ A_{t,4,1} \ \nu _{1,t} + A_{t,4,2} \ \nu _{2,t} \end{bmatrix} + \begin{bmatrix} A_{t,3,3} \ \nu _{3,t} \\ A_{t,4,3} \ \nu _{3,t} + A_{t,4,4} \ \nu _{4,t} \end{bmatrix}, \ \ \begin{bmatrix} \nu _{3,t} \\ \nu _{4,t} \end{bmatrix} \ \sim N(0,I_2) \end{aligned} \end{aligned}$$

where \(P = \begin{bmatrix} 0 &{} 0 &{} 1 &{} 0 \\ 0 &{} 0 &{} 0 &{} 1\end{bmatrix}\) is a selection matrix, \(A_t\) is a (\(4 \times 4\)) lower triangular matrix such that \(A_tA_t' = \Sigma _t\), \(A_{t,i,j}\) is the (ij) element of \(A_{t}\), and \(\Sigma _t\) is the time-varying variance–covariance matrix of \([u_t \; e_t \; \eta _{c,t} \; \eta _{i,t}]'\). Given all other parameters, \(u_t\) (\(=y_t - x_t\)), and \(e_t\) (\(=z_t - \phi _1z_{t-1} - \phi _2z_{t-2})\), we can obtain \(A_t\), and the normalized errors \(\{ \nu _{1,t},\nu _{2,t}\}\). We can derive the posterior mean and covariance matrix of \(\{ {\bar{c}}, \lambda _c,{\bar{i}}, \lambda _i\}\) with a conjugate prior based on the following multiple regression representation:

$$\begin{aligned} \begin{aligned} \begin{bmatrix} c_t^* \\ i_t^* \end{bmatrix} - \begin{bmatrix} A_{t,3,1} \ \nu _{1,t} + A_{t,3,2} \ \nu _{2,t} \\ A_{t,4,1} \ \nu _{1,t} + A_{t,4,2} \ \nu _{2,t} \end{bmatrix}&= \begin{bmatrix} x_{c,0}^* &{} x_{c,1}^* &{} 0 &{} 0 \\ 0 &{} 0 &{} x_{i,0}^* &{} x_{i,1}^* \end{bmatrix} \ \begin{bmatrix} {\bar{c}} \\ \lambda _c \\ {\bar{i}} \\ \lambda _i \end{bmatrix} \\&\quad + \begin{bmatrix} A_{t,3,3} \ \nu _{3,t} \\ A_{t,4,3} \ \nu _{3,t} + A_{t,4,4} \ \nu _{4,t} \end{bmatrix}, \end{aligned} \end{aligned}$$

where \(c_t^* = (c_t -\gamma _{c,1} c_{t-1}-\gamma _{c,2} c_{t-2})\), \(i_t^* = (i_t -\gamma _{i,1} i_{t-1}-\gamma _{i,2} i_{t-2})\), \(x_{c,0}^* = (1-\gamma _{c,1} - \gamma _{c,2})\), \(x_{i,0}^* = (1-\gamma _{i,1} - \gamma _{i,2})\), \(x_{c,1}^* = (x_t -\gamma _{c,1} x_{t-1}-\gamma _{c,2} x_{t-2})\), \(x_{i,1}^* = (x_t -\gamma _{i,1} x_{t-1}-\gamma _{i,2} x_{t-2})\).

1.2 Appendix A.2: posterior simulation of the latent state variables

To draw the trend and cycle components of the UC model, we apply the Kalman smoothing algorithm described by Carter and Kohn (1994) conditional on \(m_{0:T}\)Footnote 17 and \(h_{1:T}\). Notably, the smoothing algorithm developed by Frühwirth-Schnatter (1994) can also be applied in the posterior sampling step. Because of the degenerate measurement and transition equations, we draw only the trend component \(x_t\) from the smoothing algorithm and recover the cycle components using Eqs. (1), (2), and (3). We do not explain the Gibbs sampling step in detail because it is a standard method.

This appendix derives a particle Gibbs (PG) sampling algorithm to draw \(m_{0:T}\) and \(h_{1:T}\) from their posterior joint distribution conditional on \(\beta _{0:T}\). Our estimation method is based on a study by Lindsten et al. (2015).

1.2.1 A.2.1. Sequential Monte Carlo method

Suppose that we have the following approximate filtering density for the latent variables:

$$\begin{aligned} p_{\theta }(m_{t-1},h_{t-1} \vert Y_{1:t-1},\beta _{1:t-1}) \approx \sum _{n = 1}^{N} {\hat{\omega }}_{t-1}^{(n)} \delta _{\{m_{t-1}^{(n)},h_{t-1}^{(n)}\}}(m_{t-1},h_{t-1}) \end{aligned}$$

where \(\{ m_{t-1}^{(n)},h_{t-1}^{(n)} \}_{n =1}^{N}\) and \(\{{\hat{\omega }}_{t-1}^{(n)} \}_{n=1}^N\) denote N particles of the latent variables and their corresponding normalized importance weights at time \(t-1\), and \(\delta _{\{m_{t-1}^{(n)},h_{t-1}^{(n)}\}}(m_{t-1},h_{t-1} )\) is a Dirac measure of each particle in \(\{ m_{t-1}^{(n)},h_{t-1}^{(n)}\}_{n=1}^N\). For forward filtering, an ancestor index denoted by \(a_t^{(n)}\) is drawn using the following importance weights prior to sampling a new particle \(\{m_t^{(n)},h_{t}^{(n)}\}\):

$$\begin{aligned} Pr(a_t^{(n)} = r) = {\hat{\omega }}_{t-1}^{(r)}, \end{aligned}$$

where \(r = 1,2,\ldots ,N\). Then, an importance distribution is used to generate the new particle \(\{ m_t^{(n)},h_t^{(n)}\}\) conditional on \(\{m_{t-1}^{(a_t^{(n)})},h_{t-1}^{(a_t^{(n)})}\}\), which is the \(a_t^{(n)}\)-th particle in the particle swarm at time \(t-1\).

To construct an importance distribution, we employ the transition densities of \(h_t\) and \(m_t\) as follows:

$$\begin{aligned} q(m_t ,h_t\vert m_{t-1},h_{t-1}) = f_{\theta }(m_t \vert m_{t-1})f_{\theta }(h_t \vert h_{t-1}). \end{aligned}$$

The corresponding unnormalized importance weight is given byFootnote 18 as follows:

$$\begin{aligned} {\bar{\omega }}_t^{(n)} = f_{\theta }(\beta _t \vert \beta _{t-1}, m_{t}^{(n)}, h_{t}^{(n)})I(h_{u,t}), \end{aligned}$$

where \(I(h_{u,t})=1\) if the stochastic volatility of the permanent shock is within the prespecified bounds and \(I(h_{u,t})=0\) otherwise. After sampling the ancestor index and the new particle and evaluating the unnormalized importance weight of all N particles, we obtain \(\{a_{t}^{(n)}\}_{n=1}^{N}\), \(\{s_t^{(n)},h_t^{(n)} \}_{n=1}^N\), and \(\{{\bar{\omega }}_t^{(n)} \}_{n=1}^N\). The importance weight of \(\{m_t^{(n)},h_t^{(n)}\}\) is computed through self-normalization as follows:

$$\begin{aligned} {\hat{\omega }}_t^{(n)} = {{\bar{\omega }}_t^{(n)} \over \sum _{i=1}^N {\bar{\omega }}_t^{(i)}} \end{aligned}$$

for \(n=1,2,\ldots ,N\). Notably, \(Y_t\) is irrelevant for the importance weight given by \(\beta _t\) and \(\beta _{t-1}\) because there is no measurement error in Eq. (A.1). By continuously performing the particle sampling steps up to time T, we can collect \(M_{0:T} = \{m_{0}^{(n)},m_{1}^{(n)},\ldots ,m_{T}^{(n)}\}_{n=1}^N\); \(H_{1:T} = \{h_{1}^{(n)},\ldots ,h_{T}^{(n)}\}_{n=1}^N\); \(A_{1:T} = \{a_{1}^{(n)},a_{2}^{(n)},\ldots ,a_{T}^{(n)}\}_{n=1}^N\). This algorithm is referred to as the sequential Monte Carlo (SMC) method. Recall that \(h_{1}^{(n)}=1\).

1.2.2 A.2.2. Particle Gibbs sampler with ancestor sampling

Among \(\{M_{0:T},H_{1:T},A_{1:T}\}\) generated by the SMC method, we sample one particular particle path known as a reference particle trajectory. The reference particle trajectory is sampled by first drawing the index \(b_T \in \{1,2,\ldots ,N \}\) and then sequentially recovering \(b_{t-1} = a_t^{(b_t)}\) for \(t=T,\ldots ,1\). Here, the index \(b_t\) indicates the locations of the individual particles in the chosen reference particle trajectory as follows:

$$\begin{aligned} \{m_0^{(b_0)},m_1^{(b_1)},\ldots ,m_T^{(b_T)},h_1^{(b_1)},\ldots ,h_T^{(b_T)} \}. \end{aligned}$$

For notational simplicity, we label the reference particle trajectory \(\{m_{0:T}^{(b_{0:T})},h_{1:T}^{(b_{1:T})}\}\). The probability of \(b_T = r\) for \(r = 1,2,\ldots ,N\) is determined by the importance weight at time T as follows:

$$\begin{aligned} Pr(b_T = r \vert \theta , \beta _{0:T}, M_{0:T},A_{1:T}) = {\hat{w}}_T^{(r)}. \end{aligned}$$

Then, we update the remaining particles and the ancestor indices denoted by \(\{ {M_{0:T}^{(-b_{0:T})}},\) \({H_{1:T}^{(-b_{1:T})}}, A_{1:T}^{(-b_{0:T})}\}\) conditional on \(\{m_{0:T}^{(b_{0:T})},h_{1:T}^{(b_{1:T})}, b_{0:T}\}\) that we obtain in the previous step. The update is simple. \(N-1\) particles and the corresponding ancestor indices are sampled by treating the accepted \(\{m_{t}^{(b_{t})},h_{t}^{(b_{t})}, b_{t-1}\}\) as the N-th particle and the N-th ancestor index for \(t=0,1,\ldots ,T\). This step is called the conditional sequential Monte Carlo (CSMC) method. A detailed exposition of its implementation is provided in the outline of the proposed PG algorithm.

The ancestor sampling method updates the ancestor index \(b_{t-1} \in \{1,2,\ldots ,N \}\) conditional on \(\{ m_{t}^{(b_{t})},h_{t}^{(b_t)}, b_t \}\). The sampling procedure is carried out for the entire period. To construct a valid MCMC step, consider the following target density for the ancestor sampling, which is the backward kernel at time t:

$$\begin{aligned} \Psi (b_{t-1}) \propto f_{\theta }(m_{t}^{(b_{t})} \vert m_{t-1}^{(b_{t-1})}) f_{\theta }(h_{t}^{(b_{t})} \vert h_{t-1}^{(b_{t-1})}) {\hat{\omega }}_{t-1}^{(b_{t-1})}. \end{aligned}$$

Because the normalization constant of the backward kernel is unknown, we employ the importance sampling method again to draw \(b_{t-1}\). The normalized importance weights of \(l = 1,2,\ldots ,N\) are given by

$$\begin{aligned} {\hat{\tau }}_t^{(l)} = {{\bar{\tau }}^{(l)}_{t} \over \sum _{j=1}^{N} {\bar{\tau }}^{(j)}_{t} }, \end{aligned}$$


$$\begin{aligned} {\bar{\tau }}^{(l)}_{t} = f_{\theta }(m_{t}^{(b_{t})} \vert m_{t-1}^{(l)}) f_{\theta }(h_{t}^{(b_{t})} \vert h_{t-1}^{(l)}) {\hat{\omega }}_{t-1}^{(l)}. \end{aligned}$$

Importance sampling is performed conditional on the previously accepted reference particle trajectory. This step refers to the conditional importance sampling (CIS) algorithm.

The PG algorithm derived in this appendix is an extension of the PG algorithm developed by Lindsten et al. (2014) to estimate a nonlinear switching state space model. The uniform ergodicity of the suggested algorithm is demonstrated in Lindsten et al. (2014). The following PG algorithm provides new posterior samples of \(m_{0:T}\) and \(h_{1:T}\) conditional on previously accepted \(\{m_{0:T}^{(b_{0:T})},h_{1:T}^{(b_{1:T})},b_{0:T}\}\) and \(\beta _{0:T}\):Footnote 19

\({\underline{\mathbf{Conditional ~SMC ~with ~Ancestor ~Sampling}}}\)

  • For \(t=0\),

    1. (i-1)

      Draw \(\{m_0^{(n)}\}_{n=1}^{N-1}\) from the \(f_\theta (m_0)\) and set \(h_0 = 1\).

    2. (i-2)

      Update \(\{m_0^{(b_0)},h_0^{(b_0)}\}\) using CIS algorithm.

      Set \(\{m_0^{(N)},h_0^{(N)}\} = \{m_0^{(b_0)},h_0^{(b_0)}\}\).

    3. (i-3)

      Set \(\{{\hat{\omega }}_0^{(n)} = {1\over N}\}_{n=1}^{N}\).

  • Iterate steps (ii-1), (ii-2), (ii-3), (ii-4), and (ii-5) for \(t=1,2,\ldots ,T\).

    1. (ii-1)

      Draw ancestor indices, \(\{ a_t^{(n)}\}_{n=1}^{N_p-1}\) with the normalized probabilities, \(\{ {\hat{\omega }}_{t-1}^{(n)} \}_{n=1}^{N_p}\).

    2. (ii-2)

      Draw \(\{m_t^{(n)},h_t^{(n)}\}_{n=1}^{N-1}, \) from \(f_{\theta }(m_t \vert m_{t-1}^{(a_t^{(n)})})\) and \(f_{\theta }(h_t \vert h_{t-1}^{(a_t^{(n)})})\).

    3. (ii-3)

      Run CIS algorithm to update \(b_{t-1}\).

      Set \(a_t^{(N)} = b_{t-1}\) and \(\{m_t^{(N)},h_t^{(N)}\} = \{m_t^{(b_t)},h_t^{(b_t)}\}\).

    4. (ii-5)

      Calculate the unnormalized weights \({\bar{\omega }}^{(n)}_{t}\) and obtain the normalized weights \({\hat{\omega }}_t^{(n)} = {{\bar{\omega }}^{(n)}_{t} \over \sum _{j=1}^N {\bar{\omega }}^{(j)}_{t} }\) of \(n = 1,2,\ldots ,N\).

  • Perform steps (iii-1), and (iii-2) for \(t=T\).

    1. (iii-1)

      Draw \(b_T^*\) with \(\{ {\hat{\omega }}_{T}^{(n)} \}_{n=1}^N\) and construct \(b_{0:T-1}^*\) using \(b_{t-1}^* = a_t^{(b_t^*)}\) for \(t = T,T-1,\ldots ,1\).

    2. (iii-2)

      Construct a new reference particle trajectory \(\{h_{1,1:T}^{(b_{1:T}^*)}\}\).

      Set \(\{h_{i,1:T}^{(b_{1:T})},b_{0:T}\} = \{h_{1,1:T}^{(b_{1:T}^*)},b_{0:T}^*\}\).

Rights and permissions

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Kim, J., Chon, S. Bayesian estimation of the long-run trend of the US economy. Empir Econ 62, 461–485 (2022).

Download citation

  • Received:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI:


JEL classification