80 CONFIDENCE-WEIGHTED MEAN REVERSION
where φ
= 2φ for CWMR-Var and φ
=
φ
√
U
t
for CWMR-Stdev. Since both λ
t+1
and φ
are positive, poor-performing assets (with lower values of x
t,i
) have higher
variance terms than good-performing ones (with higher x
t,i
). Note that denotes the
covariance matrix of brather than x.Thus,ahigher value means that the corresponding
mean is more volatile than others. Since we move the weights from good-performing
assets to poor-performing ones, the latter would change more than the former, that
is, the latter has higher volatility. In the next update of μ, assets with high volatility
would actively magnify the movement magnitude.
To better illustrate the updates, we give running updates based on a classic exam-
ple (Cover and Gluss 1986). Let a market consist of cash and one volatility asset, and
the sequence of x is
1,
1
2
,
1, 2
,
1,
1
2
,.... Obviously, market strategy can gain
nothing since no asset grows in the long run. The best CRP strategy, with b =
1
2
,
1
2
,
grows to
9
8
n
2
at the end of n periods. However, starting with μ
0
=
1
2
,
1
2
, the CWMR
strategy can grow to
3
4
×2
n−1
2
after n-th periods. Table 10.2 shows the running details
for the initial five periods, and further details can be derived. On each period t +1,
the mean moves toward the last mean and also moves far away by the excess return
vector (x
t
−¯x
t
1), and its magnitude is determined by both λ
t
and
t
. Note that, in
this example, μ before projection is out of the simplex domain and is forced sparse
via normalization, which is not an usual case in real tests. In summary, both the first-
and second-order information contribute to CWMR’s success.
Then, let us compare deterministic CWMR with the stochastic version (Line 4
in Algorithm 10.2), which draws a portfolio based on both the mean and covariance
matrix. Interestingly, negatively affects CWMR’s performance in several aspects.
Firstly, a stochastic b drawn from the distribution is always different from the optimal
mean μ, which obviously causes performance divergences. Given that converges
to the zero matrix (see the recursive updates in the two propositions), the distribution
of b conditioning on the data converges to the point mass at the mean parameter
value, μ = lim
t
μ
t
. Thus, drawing weights b from the distribution (the stochastic
version) is suboptimal, since we already have an estimate of μ. It is better to choose b
as either the mode or mean (incidentally, the same for the Gaussian case), which is
actually deterministic CWMR. Another effect caused by the stochastic behavior is
Table 10.2 A running example of CWMR-Stdev on the Cover’s game
t x
t
b
t
b
t
x
t
λ
t
x
t
−¯x
t
1 diag(
t
) μ
t
0 (0.25, 0.25) (0.5, 0.5)
1 (1.0, 0.5) (0.5, 0.5) 0.75 40.78 (0.25, −0.25) (0.10, 0.40) (0.0, 1.0)
2 (1.0, 2.0) (0.0, 1.0) 2.00 61.61 (−0.80, 0.20) (0.40, 0.10) (1.0, 0.0)
3 (1.0, 0.5) (1.0, 0.0) 1.00 75.56 (0.10, −0.40) (0.10, 0.40) (0.0, 1.0)
4 (1.0, 2.0) (0.0, 1.0) 2.00 61.61 (−0.80, 0.20) (0.40, 0.10) (1.0, 0.0)
5 (1.0, 0.5) (1.0, 0.0) 1.00 75.56 (0.10, −0.40) (0.10, 0.40) (0.0, 1.0)
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
T&F Cat #K23731 — K23731_C010 — page 80 — 9/28/2015 — 21:24