discussion of the dynamics of expected returns: evidence ... · ⋆ this paper introduces a new...
TRANSCRIPT
![Page 1: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/1.jpg)
Discussion ofThe Dynamics of Expected Returns: Evidence from
Multi-Scale Time Series Modeling
Anne Opschoor
Vrije Universiteit Amsterdam
9th ECB Workshop on Forecasting TechniquesFrankfurt, June 3-4, 2016
Anne Opschoor (VU) Frankfurt, June 2016 1 / 7
![Page 2: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/2.jpg)
Summary
⋆ This paper introduces a new framework that combines scale-specificinformation from multiple predictors for extracting and modeling thedynamics of latent short-term expected returns. The framework isbased on multi-scale model of Ferreira et al.(2006).
Anne Opschoor (VU) Frankfurt, June 2016 2 / 7
![Page 3: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/3.jpg)
Summary
⋆ This paper introduces a new framework that combines scale-specificinformation from multiple predictors for extracting and modeling thedynamics of latent short-term expected returns. The framework isbased on multi-scale model of Ferreira et al.(2006).
⋆ More specifically, the model combines information from the logdividend-price ratio and the consumption-wealth ratio at differenthorizons.
Anne Opschoor (VU) Frankfurt, June 2016 2 / 7
![Page 4: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/4.jpg)
Summary
⋆ This paper introduces a new framework that combines scale-specificinformation from multiple predictors for extracting and modeling thedynamics of latent short-term expected returns. The framework isbased on multi-scale model of Ferreira et al.(2006).
⋆ More specifically, the model combines information from the logdividend-price ratio and the consumption-wealth ratio at differenthorizons.
⋆ Using this framework results in long-lasting effects on expectedreturns. The degree of persistence (ACF) and model-implied forecastsdiffer form standard ARMA models.
Anne Opschoor (VU) Frankfurt, June 2016 2 / 7
![Page 5: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/5.jpg)
Summary
⋆ This paper introduces a new framework that combines scale-specificinformation from multiple predictors for extracting and modeling thedynamics of latent short-term expected returns. The framework isbased on multi-scale model of Ferreira et al.(2006).
⋆ More specifically, the model combines information from the logdividend-price ratio and the consumption-wealth ratio at differenthorizons.
⋆ Using this framework results in long-lasting effects on expectedreturns. The degree of persistence (ACF) and model-implied forecastsdiffer form standard ARMA models.
⋆ Bayesian estimation is done by a MCMC algorithm. A simulationstudy shows that the method works pretty well.
Anne Opschoor (VU) Frankfurt, June 2016 2 / 7
![Page 6: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/6.jpg)
Summary
⋆ Nice idea to combine different scales into one model
Anne Opschoor (VU) Frankfurt, June 2016 3 / 7
![Page 7: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/7.jpg)
Summary
⋆ Nice idea to combine different scales into one model
⋆ Empirical exercise shows that the new model outperforms otherbenchmarks with increasing horizon (both point and density forecasts)
Anne Opschoor (VU) Frankfurt, June 2016 3 / 7
![Page 8: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/8.jpg)
Summary
⋆ Nice idea to combine different scales into one model
⋆ Empirical exercise shows that the new model outperforms otherbenchmarks with increasing horizon (both point and density forecasts)
⋆ Also the optimal allocation in a portfolio-decision problem changes ifinformation from both scales is taken into account.
Anne Opschoor (VU) Frankfurt, June 2016 3 / 7
![Page 9: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/9.jpg)
Comments: The new framework
⋆ The link function is the same as in Ferreira et al.(2006). That is, z isthe average of non-overlapping groups of m consecutive x values.Interpretation? Expected return is the average of log-dividend ratiovalues?
Anne Opschoor (VU) Frankfurt, June 2016 4 / 7
![Page 10: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/10.jpg)
Comments: The new framework
⋆ The link function is the same as in Ferreira et al.(2006). That is, z isthe average of non-overlapping groups of m consecutive x values.Interpretation? Expected return is the average of log-dividend ratiovalues?
⋆ Normality is assumed for both processes x and z . Good for derivingconditional distributions. What is the impact of this assumption?Moreover, how can we justify this?
Anne Opschoor (VU) Frankfurt, June 2016 4 / 7
![Page 11: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/11.jpg)
Comments: The new framework
⋆ The link function is the same as in Ferreira et al.(2006). That is, z isthe average of non-overlapping groups of m consecutive x values.Interpretation? Expected return is the average of log-dividend ratiovalues?
⋆ Normality is assumed for both processes x and z . Good for derivingconditional distributions. What is the impact of this assumption?Moreover, how can we justify this?
⋆ Both processes have also constant variances σ2x
and σ2z. Is this
realistic?
Anne Opschoor (VU) Frankfurt, June 2016 4 / 7
![Page 12: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/12.jpg)
Comments: Persistence, φx and ARMA models
⋆ The paper shows interesting figures to highlight persistence of thenew framework for φx = φz = 0.9 (Figure 2). The empirical exerciseimplies that in particular the setting of σ2
x/σ2
z= 3 is interesting.
Moreover φ̂x = 0.67.
Anne Opschoor (VU) Frankfurt, June 2016 5 / 7
![Page 13: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/13.jpg)
Comments: Persistence, φx and ARMA models
⋆ The paper shows interesting figures to highlight persistence of thenew framework for φx = φz = 0.9 (Figure 2). The empirical exerciseimplies that in particular the setting of σ2
x/σ2
z= 3 is interesting.
Moreover φ̂x = 0.67.
⋆ What is the effect on the autocorrelations in this particular settingcompared to the AR(1)?? How harmful is this ? Could this bequantified?
Anne Opschoor (VU) Frankfurt, June 2016 5 / 7
![Page 14: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/14.jpg)
Comments: Persistence, φx and ARMA models
⋆ The paper shows interesting figures to highlight persistence of thenew framework for φx = φz = 0.9 (Figure 2). The empirical exerciseimplies that in particular the setting of σ2
x/σ2
z= 3 is interesting.
Moreover φ̂x = 0.67.
⋆ What is the effect on the autocorrelations in this particular settingcompared to the AR(1)?? How harmful is this ? Could this bequantified?
⋆ In addition to this, all Figures are all based on basic AR(1) dynamics.How do things change if we come up with for example an ARMA(2,1)model?
Anne Opschoor (VU) Frankfurt, June 2016 5 / 7
![Page 15: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/15.jpg)
Comments: Persistence, φx and ARMA models
⋆ The paper shows interesting figures to highlight persistence of thenew framework for φx = φz = 0.9 (Figure 2). The empirical exerciseimplies that in particular the setting of σ2
x/σ2
z= 3 is interesting.
Moreover φ̂x = 0.67.
⋆ What is the effect on the autocorrelations in this particular settingcompared to the AR(1)?? How harmful is this ? Could this bequantified?
⋆ In addition to this, all Figures are all based on basic AR(1) dynamics.How do things change if we come up with for example an ARMA(2,1)model?
⋆ Moreover: In the simulation setting, Figure 5 (panel B) shows resultson estimating back the ACF of the extracted expected returns. Theposterior mean is on average 0.10 lower than the theoretical ACF forthe multi-scale process for almost all lags.All in all, how sensitive is this??
Anne Opschoor (VU) Frankfurt, June 2016 5 / 7
![Page 16: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/16.jpg)
Comments: Empirical exercise
⋆ What is the impact of the coarsening window m on the main results??In particular the window for the log dividend-ratio. Difficult to seewhich m is optimal.
Anne Opschoor (VU) Frankfurt, June 2016 6 / 7
![Page 17: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/17.jpg)
Comments: Empirical exercise
⋆ What is the impact of the coarsening window m on the main results??In particular the window for the log dividend-ratio. Difficult to seewhich m is optimal.
⋆ What about possible structural breaks in both series through time?(since the sample runs from 1952 - 2013).
Anne Opschoor (VU) Frankfurt, June 2016 6 / 7
![Page 18: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/18.jpg)
Comments: Empirical exercise
⋆ What is the impact of the coarsening window m on the main results??In particular the window for the log dividend-ratio. Difficult to seewhich m is optimal.
⋆ What about possible structural breaks in both series through time?(since the sample runs from 1952 - 2013).
⋆ There is a huge literature on combining forecasts for real returns, dueto unstable relationships. Does it play a role here?
Anne Opschoor (VU) Frankfurt, June 2016 6 / 7
![Page 19: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/19.jpg)
Comments: Empirical exercise (II)
⋆ Results show better MSE and predictive log-scores, in particular forlarge horizons. What about the statistical significance of thesedifferences, especially when h = 12 and h = 24?
Anne Opschoor (VU) Frankfurt, June 2016 7 / 7
![Page 20: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/20.jpg)
Comments: Empirical exercise (II)
⋆ Results show better MSE and predictive log-scores, in particular forlarge horizons. What about the statistical significance of thesedifferences, especially when h = 12 and h = 24?
⋆ One of the benchmarks is the AR(1) model estimated on extracted
returns. Obviously the model is beaten (although not for h = 12!!)Implication?
Anne Opschoor (VU) Frankfurt, June 2016 7 / 7
![Page 21: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/21.jpg)
Comments: Empirical exercise (II)
⋆ Results show better MSE and predictive log-scores, in particular forlarge horizons. What about the statistical significance of thesedifferences, especially when h = 12 and h = 24?
⋆ One of the benchmarks is the AR(1) model estimated on extracted
returns. Obviously the model is beaten (although not for h = 12!!)Implication?
⋆ Why not compare the results with the model of Pastor andStambaugh (2009), to investigate the added value of ‘scale-specific’predictors?
Anne Opschoor (VU) Frankfurt, June 2016 7 / 7
![Page 22: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/22.jpg)
Comments: Empirical exercise (II)
⋆ Results show better MSE and predictive log-scores, in particular forlarge horizons. What about the statistical significance of thesedifferences, especially when h = 12 and h = 24?
⋆ One of the benchmarks is the AR(1) model estimated on extracted
returns. Obviously the model is beaten (although not for h = 12!!)Implication?
⋆ Why not compare the results with the model of Pastor andStambaugh (2009), to investigate the added value of ‘scale-specific’predictors?
⋆ Optimal allocation differs: what about the ex-post utility?? Do wegain by incorporating scale-specific information?
Anne Opschoor (VU) Frankfurt, June 2016 7 / 7
![Page 23: Discussion of The Dynamics of Expected Returns: Evidence ... · ⋆ This paper introduces a new framework that combines scale-specific information from multiple predictors for extracting](https://reader035.vdocuments.us/reader035/viewer/2022070717/5edc1304ad6a402d66669640/html5/thumbnails/23.jpg)
Comments: Empirical exercise (II)
⋆ Results show better MSE and predictive log-scores, in particular forlarge horizons. What about the statistical significance of thesedifferences, especially when h = 12 and h = 24?
⋆ One of the benchmarks is the AR(1) model estimated on extracted
returns. Obviously the model is beaten (although not for h = 12!!)Implication?
⋆ Why not compare the results with the model of Pastor andStambaugh (2009), to investigate the added value of ‘scale-specific’predictors?
⋆ Optimal allocation differs: what about the ex-post utility?? Do wegain by incorporating scale-specific information?
⋆ Does the allocation differs over time? (crisis periods etc.)
Anne Opschoor (VU) Frankfurt, June 2016 7 / 7