Smoothing and de-smoothing of returns

Prof. Jayanth R. Varma

Desmoothing when “smoothing parameter” α is known
Smoothing model
Rt = smoothed return at time t
Rt∗ = “true” return
Rt = (1 − α)Rt∗ + αRt−1
Var(R) = (1 − α)2 Var(R∗ ) + α2 Var(R) =

Desmoothing
Rt − αRt−1 = (1 − α)Rt∗ or
Rt − αRt−1
.
Rt∗ =
1−α
(1 − α)2
Var(R∗ )
1 − α2

(1 − α)2
0.04
1
=
= = 0.1111. The true variance is greater than the observed variance by a
2
1−α
0.36
9
factor of 9 while the true standard deviation will be greater by a factor of 3
If α = 0.8,

How to estimate the smoothing parameter α?
Assume market efficiency
In an efficient market (random walk hypothesis), returns should have no serial correlation. In this case,
we should desmooth the observed returns with α equal to the autocorrelation ρ of the observed returns.
This is because if the true return has zero autocorrelation and it is smoothed with smoothing
parameter α, then the autocorrelation ρ of the smoothed return Rt is α.
Cov(Rt , Rt−1 ) = Cov((1 − α)Rt∗ + αRt−1 , Rt−1 ) = α Var(R)
where we use the fact that Rt∗ is uncorrelated with Rs∗ for s < t and is therefore also uncorrelated with
Rs for s < t.
Rt − ρRt−1
We shall let R† denote the series obtained by desmoothing with α = ρ. Rt† =
or
1−ρ
Rt = (1 − ρ)Rt† + ρRt−1 where ρ is the auto correlation of the observed return.
Even if the true return is itself an autoregressive process (the random walk hypothesis is violated), the
series R† is useful because it has zero autocorrelation as seen below.

)
Cov(Rt† , Rt−1

=
=
=

Cov(Rt − ρRt−1 , Rt−1 − ρRt−2 )
(1 − ρ)2
1
[Cov(Rt , Rt−1 ) − ρ Cov(Rt , Rt−2 ) − ρ Cov(Rt−1 , Rt−1 ) + ρ2 Cov(Rt−1 , Rt−2 )]
(1 − ρ)2
Var(R)
[ρ − ρ3 − ρ + ρ3 ] = 0
(1 − ρ)2

where we have used the fact that Cov(Rt , Rt−2 ) = ρ2 (exponential decline of the autocorrelations of
higher order for any autoregressive process).

Replicate the estimated volatility of the asset
It is often desired to choose an α to match some estimate of the volatility of the “true” returns. For
example, in case of real estate, the estimated volatility can be half of equities or average of stocks and
bonds or survey expectations of risk relative to equities/bonds. The α required to replicate a desired
volatility can be obtained by trial and error or by solving a non linear equation as shown below.
Rt − αRt−1
Suppose the true return is given by Rt∗ =
for α 6= ρ.
1−α
Rt − αRt−1 = (1 − ρ)Rt† + ρRt−1 − αRt−1 = (1 − ρ)Rt† + (ρ − α)Rt−1 .
1

2. and desmoothing these returns with α = 0. 1 + α2 − 2αρ ∗ Cov(Rt∗ .411 σ† 26.000 σ∗ 12. σ †2 be the variance of the artificial uncorrelated (random walk) series (Rt† ) and σ ∗2 be the variance of the true returns (Rt∗ ). We have get σ ∗2 = + (ρ − α)2 or 2 = = 2 2 2 (1 − α) (1 − ρ) σ (1 − α) 1 + α2 − 2α obtained a non linear equation for α in terms of the known or assumed variances σ 2 and σ ∗2 and the known autocorrelation ρ. Rt−1 − αRt−2 ) (1 − α)2 (ρ − α)σ 2 − (ρ − α)ασ 2 ρ (ρ − α)(1 − αρ) 2 = σ (1 − α)2 (1 − α)2 (ρ − α)(1 − αρ) (1 − α)2 (ρ − α)(1 − αρ) ∗2 σ ∗2 = σ (1 − α)2 1 + α2 − 2αρ 1 + α2 − 2αρ Other Methods of estimating α 1. We t t−1 t (1 − [α)2 1 − ρ2 ] 2 2 2 ∗2 2 2 σ (1 − ρ) (1 − ρ ) σ 1 + α − 2αρ 1 + α − 2αρ .458 ρ 0.750 ρ∗ 0. it is seen that σ = σ .4): σ 9.000 Note: ˆ σ is the standard deviation and ρ is the autocorrelation of the observed return series without any desmoothing ˆ σ ∗ is the standard deviation and ρ∗ is the autocorrelation of the “true” return series obtained by desmoothing with α = 0.472 σ† 26. This equation can be solved numerically. we may solve for α that maximizes the correlation with listed property (REIT) which is assumed to be free of smoothing biases. 2 .445 ρ† 0. Replicate or maximise correlations to other assets: In case of real estate for example.50 ˆ σ † is the standard deviation and ρ† is the autocorrelation of the random walk series obtained by desmoothing with α = ρ.995 σ∗ 12. Numerical Simulation Given below are the results of simulating 5. Since R = (1 − ρ)R + ρR . Rt−1 ) = = = 1 Cov((1 − ρ)Rt† + (ρ − α)Rt−1 .000 returns from a normal distribution with σ = 10 and ρ = 0. the theoretical values using the formulas derived earlier are: σ 10.Var(Rt∗ ) = (1 − ρ)2 Var(Rt† ) + (ρ − α)2 Var(Rt−1 ) (1 − α)2 where we use the fact that Rt† is uncorrelated with Rs† for s < t and is therefore also uncorrelated with Rs for s < t. (ρ − α)(1 − αρ) The autocorrelation of R∗ is as computed below.000 For comparison.756 ρ∗ 0.437 ρ† 0. Replicate professional asset portfolios which are assumed to be efficient: This is a reverse optimization problem to find the α which produce variance and covariance estimates that generate optimal portfolios that are closest to observed portfolios.75.168 ρ 0. We then have (1 − ρ)2 σ †2 + (ρ − α)2 σ 2 (1 − ρ)2 †2 † 2 σ ∗2 = . Let σ 2 denote the variance of the observed smoothed returns (Rt ).