Optimal Portfolio Strategy to Control Maximum Drawdown

The Case of Risk-based Active Management with Dynamic Asset Allocation

Z. George Yang1 and Liang Zhong2

February 25th, 20123

1
Z. George Yang, Ph.D., is the Director of Research at Flexible Plan Investments, Ltd. in Bloomfield Hills,
Michigan, and the contacting author. E-mail: gyang@[Link].

2
Liang Zhong is a Ph.D. Candidate at Department of Economics, Lerner School of Business at University
of Delaware, Newark, Delaware. E-mail: lzhong@[Link].

3
This article only reflects the opinions of the authors’ own, not of the affiliated employer company or
educational institution. The authors are responsible for any errors.

Electronic copy available at: [Link]

 Optimal Portfolio Strategy to Control Maximum Drawdown

The Case of Risk-based Active Management with Dynamic Asset Allocation

Abstract

Draw-down losses from a previously reached maximum portfolio wealth level, is

an important risk measure for investment management. In this study, we present a

discrete trading strategy to directly control a portfolio’s maximum percentage of

drawdown within a target level while maximizing the portfolio’s long term growth rate.

Modifying the continuously rebalancing models proposed by Grossman and Zhou

[1] [2]
(1993) and Cvitanic and Karatzas (1995) , we define the loss control target as a

Rolling Economic Drawdown (REDD) with a constant look-back window progressing in

time. Additionally, investor’s risk aversion in the power law wealth utility function is

specified as complement of the maximum loss limit to construct a portfolio’s risk-return

efficient frontier.

We test the dynamic strategy using data of three broad asset class indexes: S&P

500 Total Return Index (SPTR), Barclays Capital 20+ Year US Treasury Bond Index

(TLT) and Dow-Jones UBS Commodity Total Return Index (DJUBS), with 3-month US

Treasury Bill as the risk-free asset. Over a test period of the past 20 years (1992-2011), a

simplified risk-based out-of-sample dynamic asset allocation portfolio is robust against

variations in capital market expectation inputs, and out-performs the in-the-sample

calibrated model portfolio and traditional asset allocation portfolios significantly.

Electronic copy available at: [Link]

 Optimal Portfolio Strategy to Control Maximum Drawdown

The Case of Risk-based Active Management with Dynamic Asset Allocation

Executive Summary

Draw-down, the percentage of loss from a previous peak wealth level, is a critical

risk measure for investment management. Originated from the high water-mark practice

of the hedge fund industry, the portfolio’s maximum draw-down metric has profound

implication for manager incentives, performance measurement, and the broad objective

of client retention for any investment management practice.

Without an explicit loss control mechanism, traditional passive asset allocation

portfolios experienced large drawdowns during the 2008-2009 financial crises. In this

study, we propose an active quantitative trading strategy to directly control portfolio’s

drawdown within a target level while maximizing long term portfolio wealth growth rate.

With 40-year S&P 500 Index data including the post-crisis 2009-2011 market

recovery, we can demonstrate that the classical mathematical framework of continuous

drawdown control by Grossman and Zhou (1993) did not achieve optimality under

discrete trading extension. This is due to the model’s implied one-step trend-following

nature, but sizing position with a long term anchored drawdown look-back.

We introduce a new concept of Rolling Economic Drawdown (REDD) to define a

maximum loss limit. Counting investor’s opportunity loss at risk free rate (e.g. the 3-

month T-Bill yield), REDD looks back the portfolio wealth history during a rolling time

window of a fixed length. By specifying investor’s risk aversion as the complement of

drawdown loss limit, we are able to derive explicit formula of optimal dynamic asset
allocation weights for one or multiple risky assets. They depend on capital market

expectations of long term expected Sharpe ratios, volatilities and correlation coefficients.

Further, an adaptive position scaling by ratio   REDD is the key to control portfolio
1  REDD

maximum drawdown below a target level  . This is our active management proposal of a

new Rolling Economic Drawdown-Controlled Optimal Portfolio Strategy (REDD-COPS).

REDD-COPS are calibrated with historical data of three broad asset class indexes:

S&P 500 Total Return Index (SPTR), Barclays Capital 20+ Year US Treasury Bond

Index (TLT) and Dow-Jones UBS Commodity Total Return Index (DJUBS) as risky assets

up to six decades but with focus on the past 20+ years. We use a rolling time window of

one year to define REDD, considering investor’s liquidity and behavioral constraints, and

an expected time scale of market decline cycle before recovery.

Systematic active management through REDD-COPS is applied progressively in

back-test: first only to position sizing with calibrated constant parameter inputs, secondly

as an out-of-sample risk-based dynamic asset allocation scheme. Compared to traditional

Stock/Bond asset allocation benchmarks, such as a 60/40 Balanced Portfolio, Minimum

Variance Portfolio (MVP) or a Leveraged Risk Parity Portfolio (RPP) over recent 20+

years, the calibrated REDD-COPS achieves dominance on risk-return efficient frontiers.

Adding commodities index (DJUBS) as a third asset in the active scheme further

improves diversification benefit. Out-performing all calibrated models, the out-of-sample

risk-based REDD-COPS achieved around 5% excess annualized return over traditional

risk-based asset allocation portfolios with similar maximum drawdowns. For a 20-year

test period (1992-2011), the following table lists the performance comparison.
 SPTR-TLT SPTR-TLT-DJUBS
Single Index Buy & Hold Fixed Allocation Portfolio Risk-based REDD-COPS
60% 15% 20% 25%
Balanced Levered REDD REDD REDD
SPTR TLT DJUBS (60-40) MVP RPP Target Target Target
Annualized Return 7.81% 8.93% 5.65% 8.79% 9.04% 11.29% 11.06% 13.71% 16.45%
Annualized Std Deviation 15.00% 11.37% 15.23% 9.46% 8.40% 13.35% 10.29% 13.98% 17.90%
Sharpe Ratio 0.301 0.495 0.154 0.580 0.683 0.598 0.754 0.745 0.734
Average REDD 7.21% 5.64% 8.01% 4.31% 3.68% 5.45% 3.69% 5.01% 6.41%
Max REDD 46.76% 21.51% 54.48% 26.68% 20.51% 26.04% 14.65% 19.68% 24.85%
Max Drawdown 50.95% 21.40% 54.26% 28.63% 19.09% 28.16% 13.98% 19.00% 24.10%
Maximum Exposure 100% 100% 100% 100% 100% 160% 292.3% 396.8% 507.9%
Average Exposure 100% 100% 100% 100% 100% 160% 143.3% 195.2% 250.3%
Ending Multiple 4.501 5.533 3.001 5.389 5.641 8.489 8.143 13.065 21.014

Unlike the common active management using technical or fundamental analysis,

active trading scheme in risk-based REDD-COPS makes simple theoretical assumptions

about capital markets: long term un-correlated asset classes exist; their returns

statistically follow random walk and they have long term stable market price of risk.

Despite actual market’s deviations from the simplified premises in the last 20 years

(1992-2011), the risk-based REDD-COPS demonstrate robustness of out-performance in

back test. It is an active asset allocation framework for practical portfolio management.

Besides the benefit of a clear mandate to construct suitable client portfolios, REDD-

COPS methodology can be applied to design invest-able securities, such as principal-

guaranteed investment product, target risk asset allocation ETF, and target-date mutual

fund with glide path, etc. With active mechanisms derived solely from risk statistics and

risk target, the current proposal of REDD-COPS demonstrates the prowess of risk based

investment management: driven by active risk control, rather than performance chasing,

an investment portfolio will out-perform consistently, for the long term.

Acronyms and Symbols

MPT Modern Portfolio Theory

MVO Mean Variance Optimization

MVP Minimum Variance Portfolio

RPP Risk Parity Portfolio

EM Economic Maximum (portfolio wealth in history)

REM Rolling Economic Max

EDD Economic Drawdown

REDD Rolling Economic Drawdown

DD-COPS Draw-down Controlled Optimal Portfolio Strategy

REDD-COPS Rolling Economic Draw-down Controlled Optimal Portfolio Strategy

H Rolling time window length in REDD definition

W Portfolio Wealth Level

 Complement of Risk Aversion in the power law investor utility function

  Target level of portfolio maximum draw-down, maximum loss limit

s, t,   Discrete time variables

R Long term annualized return

  Volatility (evaluated from annualized standard deviation)

 Correlation coefficient

rf Average risk free interest rate (e.g. 3-month T-Bill yield) for compounding

x Portfolio weight

 Market price of risk, defined through Sharpe ratio

h Look back time window for volatility calculation

1. Introduction

Drawdown is a critically important risk measure for investment management. The

common definition of drawdown of a portfolio is the percentage loss of current wealth
(Wt) from a prior all-time high (Mt): DDt  1  Wt M t . The portfolio’s downside risk of a
prolonged drawdown matters not only to the investors’ financial well-being, but also to
the investment manager’s business survival in an immediate term.

Collecting performance based fee, usually at 20% of annual profit, hedge fund
managers have been known for the high watermark practice. If the portfolio ends a year
that is lower in value than any of the previous year, the manager will not get any
performance based compensation. Also drawdown causes account terminations or
redemptions. Losing account assets undermines any investment business or even lead to
its demise. When it comes to tolerating large drawdowns for a separate managed account,
there is no or little regard to whether the strategy is valid in long run with high expected
risk adjusted return – it simply cannot survive to that point!

Maximum drawdown challenges a client’s financial and psychological tolerance.

According to Chekhlov et al (2005) [3], 50% drawdown is unlikely to be tolerated in any
average account, and an account may be closed if drawdown breaches 20% or has lasted
over two years. For passive index investors, financial markets have been tough:
maximum drawdown over 50% occurred for both the Dow Jones Industrial Average
(DJIA) and the S&P 500 Index during the recent 2008-2009 financial crisis. A 50%
drawdown, as pointed out by Zhou and Zhu (2009)[4], however, is 90% probable to
happen over a century even if the stock markets are simply modeled as random walk.

Despite being widely used, diversification through passive asset allocation was not
effective to avoid large drawdowns. During a market crisis, risky asset classes can exhibit
the “contagion” effect: highly correlated losses across the board lead to large drawdowns.
[5]
Markowitz (1952) modern portfolio theory (MPT) and Mean Variance Optimization
(MVO) methodology defined risk as return standard deviation, a path-independent
statistical attribute. Without an explicit mechanism to control maximum drawdown, it
was not uncommon for a traditional balanced (60% stock + 40% bond) portfolio suffering
maximum drawdown loss of 30% during the 2008-2009 financial crises.

[1]
Grossman and Zhou (1993) pioneered the mathematical frame-work of portfolio
optimization under the dynamic floor constraint to control maximum drawdown,
[6]
extending the constant floor portfolio optimization by Black and Perold (1987) – the
basis of the theory and practice of Constant Portion Portfolio Insurance (CPPI). In an
award-winning Ph.D. thesis work, Grossman and Zhou (1993) approached the problem
with Expected Utility Theory and defined portfolio optimality as maximizing long term
growth rate in power law wealth utility function ( U  W   ). The model assumed
continuously rebalancing between a risk free asset and single risky asset, which has
random walk return dynamics. Their drawdown calculation accounted the economic
decay of portfolio value at the risk free rate r f 4. An Economic Drawdown (EDD) was

defined by EDDt   1 Wt / EM t  , where an Economic Max (EM) since inception is

calculated as EM t   Max1  rf t  s Ws . The continuous Drawdown-Controlled Optimal

0 s t

Portfolio Strategy (DD-COPS) has the portfolio fraction allocated to single risky asset as:

    12     EDDt 
xt       (1)
 1       1  EDDt  

where  is the drawdown limit, 1    is the investor risk aversion, and   R  r f   is

the long term expected Sharpe ratio of the risky asset (R and  are its long term expected
return and volatility). The rest of the portfolio is allocated to the risk free asset. The risky

asset allocation has a leverage factor     2  , and is scaled by a timing ratio   EDDt  ,
1

 1    1  EDDt 

which adaptively controls drawdown. The case of   100% gives Merton’s (1971) [7]

We take r f as the average realized risk free rate over a period of length t  s  and write the compounding
4

factor as 1  rf  t s
in Economic Max definition. When the risk free rate is changing, we will use
t

 1  r  t  to compound, where r denotes the risk free interest rate during i’th discrete time interval t.
i s
i i
unconstrained optimal portfolio leverage x   1    2  where drift   R  rf  1 2   2 in a

continuous random walk model. Optimal leverage from Kelly’s criterion formulae
x    2 is a further special case that an investor has a myopic logarithmic utility
function with   0 . Cvitanic and Karatzas (1995) [2]
extended the continuous optimal
drawdown control strategy to a case of multiple risky assets. Under discrete time interval
rebalance and leverage-limited linear constraints, He (1999) [8] conducted market data test
and obtained piecewise parabola type risk-return efficient frontier.

However, Klass and Nowicki (2005) [9] countered that a discrete implementation can
result in the loss of portfolio optimality. The discrete trading of equation (1) uses updated
current portfolio value, but the allocation takes full effect with a finite t time delay. Loss
of portfolio optimality is due to the finite delay and an anchored long term drawdown
look-back. Market cycle from decline to recovery can be much longer than the discrete
rebalance frequency. Compared to a continuous rebalancing, discrete trading under-
weights risky assets during a market downward spiral. Due to long term memory effect of
the drawdown control, less exposure to risky asset can cascade into the rebound phase of
the market cycle, leading to lower long term accumulated returns. The recent 2007 to
2011 market cycle provides us the opportunity to test the loss of optimality problem.
With S&P 500 Index as the risky asset, Appendix 1 gives a data-based demonstration.

2. Analytical Formulation

Economic Drawdown (EDD) in Grossman and Zhou (1993)’s model, reflects an

idealistic mental accounting from sophisticated investors: how much better off if they
have exited the risky asset completely at a retrospective perfect time in history, when a
risk free rate compounded historical high was achieved. However, not all investors
invested or had memory since time zero – there are portfolio inception difference among
investors. There are also liquidity constraints: not all investors can exit at a perfect time.
Hedge fund’s initial 1-year lock-up and quarterly redemption window, mutual fund
minimum holding period or redemption penalty are examples of restrictions. Practically,
at time of a market cycle bottom, using a drawdown reference lower than Economic Max
(EM) can improve performance as a forward looking market timing mechanism.
 We propose an alternative to the anchored time window (since inception) for
drawdown calculation: a constant rolling time window. Define a Rolling Economic Max
(REM) at time t, looking back at portfolio wealth history for a rolling window of length H:

REM (t , H )  Max [(1  rf )t  s Ws ] (2)

t  H  s t

Rolling Economic Max (REM) with a reset feature can be alternatively defined as:

 
 Max Wt  s 1  r f

s , t   H
 0 s  H
REM t , H    where  t   max s REM s  Ws 
  ,
s t (3)


Max Wt ,Wt  H 1  r f

H
t   H

Equation (3) is the same as (2) to define REM for most of the time, but it compounds
portfolio value since H period ago at risk-free rate if it has not been renewed by the
portfolio value over time H. Equation (2) is used later unless pointed out otherwise.

From REM, the Rolling Economic Draw-down (REDD) is consequently defined as:

Wt
REDD(t , H )  1  (4)
REM (t , H )

Intuitively, a drawdown look-back period H somewhat shorter than or similar to the

market decline cycle is the key to achieve optimality. Substituting EDD with a lower
REDD in equation (1), we have higher risky asset allocation to improve portfolio return
during a market rebound phase. In the examples followed, we’ll use H  1 year throughout.

Our second change to the original continuous models is about investor risk profile
characterization in a dynamic asset allocation framework. Equation (1) has two risk
tolerance parameters: drawdown control target  and risk aversion complement  . Gross
and Zhou (1993) did not make a direct connection between them. A straightforward
conjecture is that a more risk-averse investor should have a lower drawdown loss
tolerance. We propose in this study simply    to use the dynamic floor percentage as
the risk aversion parameter in power utility function. In comparison, Kelly’s formulae
( x    2 ) implies 100% drawdown tolerance (   1 ) whereas assuming logarithmic
wealth utility (   0 ), the conservative limit of power law utility functions.
 We propose a Rolling Economic Drawdown – Controlled Optimal Portfolio
Strategy (REDD-COPS): a dynamic allocation policy between risky asset(s) and a risk
free asset, such that the long term wealth growth rate is maximized under a dynamic
constraint REDD   . The dynamic portfolio fraction allocated to single risky asset is:

     12     REDDt , H  
xt  Max0,  2  
  (5)
  1     1  REDDt , H   
The risk free asset (e.g. 3-month T-Bill) accounts for the rest of portfolio
allocation x f  1  xt . In discrete trading, equation (5) disallows shorting risky asset but

has the portfolio all in risk free asset for any out-of-control time when REDD   . Larger
expected Sharpe ratio , smaller volatility of the risky asset, or a higher drawdown
tolerance , leads to higher or leveraged level of investments in the risky asset. Maximum
weighting in single risky asset is    12    1   2  , realized when REDD = 0.

For two risky assets in REDD-COPS, dynamic asset allocation weights are:

 x1  1  1  12  1    2  12  2   1   1    REDD 

x   2   1       1    
 Max 0, 
2   (6)
 2 1   2 2 2 1 2 1 2  1    1  REDD 

The portion of risk free asset is x f  1  x1  x2 . It is obvious that return correlation efficient

between two risky assets impacts position size and leverage in the dynamic asset
allocation process. When risky assets have positive Sharpe ratios (no shorting in
any risky asset needs   0 or   1  12 1  2  12  2   1  (   0 ). In Appendix 2, we

derive equation (6), and the explicit asset allocation formula for three risky assets case.

3. Historical Calibration

Traditional asset allocation process needs to estimate asset class returns, standard
deviations and correlations as input parameters to the portfolio optimization process. One
can calculate these fixed input parameters from historical data sample themselves to get
in-the-sample optimized performance – the best possible fixed asset allocation outcome
during the historical sampling period.
 For discrete REDD-COPS, current and historical portfolio values are used in the
forward asset allocation weight calculation, even though the asset class parameters of ,
 and can still be estimated as constants from in-the-sample process. We call this
historical data Calibrated REDD-COPS model. The purpose of model calibration is to
demonstrate the value added by the active forward drawdown control mechanism,
beyond what traditional asset allocation can do theoretically from in-the-sample
parameter specification.

3.1 Single Risky Asset – May not be enough

For a period of 60.5 years ending 2011/6/30, we use S&P 500 Total Return Index
(SPTR) as the single risky asset and 3-month US T-bill as the risk free asset to test a
Calibrated REDD-COPS. At market close of the last trading day every month, the
portfolio is rebalanced between SPTR and T-bill according to equation (5). The
allocations are held through the end of next month then repeat the process. The  and
parameters in equation (5) are calculated from SPTR Index’s monthly return streams
for the same 60.5-year period (see Table 1 for values of  and in the column for SPTR).

Table 1: Calibrated REDD-COPS of Single Risky Asset (SPTR): 1/1951-6/2011*

SPTR 1/3 REDD-COPS 25% REDD-COPS 20% REDD-COPS 15% REDD-COPS
Annualized Return 10.81% 11.50% 9.79% 8.80% 7.82%
Annualized Std Deviation 14.61% 14.16% 10.03% 7.80% 5.73%
Max REDD 46.76% 29.43% 21.59% 17.10% 12.72%
Average REDD 6.19% 6.42% 4.58% 3.58% 2.63%
Max Drawdown 50.95% 37.67% 26.53% 19.89% 14.29%
Sharpe Ratio 0.403 0.465 0.486 0.497 0.506
Average Total Exposure 100% 103.74% 73.56% 57.32% 42.08%
Max Total Exposure 100% 122.30% 86.97% 67.94% 50.04%
Ending Multiple (60.5 years) 498.0 726.6 284.6 164.2 95.3
Skew -0.418 -0.511 -0.504 -0.452 -0.444
Excessive Kurtosis 1.769 2.348 2.354 2.364 2.300

* In the first year 1951, REDD is evaluated by anchoring look-back to the end of 1950. The fixed
rolling one-year look-back starts at the beginning of 1952. “Ending Multiple” refers to the
portfolio dollar value on 2011/6/30 if it started as one dollar on 1950/12/31 without any deposit or
withdrawal / tax during the 60.5 years. “Total Exposure” refers the portfolio portion x invested in
risky asset SPTR. The “Average Total Exposure” and “Average REDD” reported are the simple
arithmetic averages of the 726 monthly end values. Risk free rate for Sharpe ratio calculation is
4.92%, the average 3-month T-bill yield over the 60.5 years.
 From Table 1, all “Max REDD” have been controlled within the target levels
(and tested, despite that the SPTR index deviates from the normal
return distribution with negative skew and excess kurtosis. The ex post Maximum
Drawdown stays within target  for lower levels but is exceeded when  > 1/4. Compared
to SPTR, Calibrated REDD-COPS with single risky asset (1/3 in Table 1) reduces
Max REDD and Max Drawdown substantially while adding returns.

Figure 1: Growth of Calibrated REDD-COPS (Single Risky Asset SPTR 1951-2011/6)

1000
Rolling Economic Max (33.3% REDD-COPS)

Portfolio Wealth (Target 33.3% REDD-COPS)

SPTR

Rolling Economic Max (20% REDD-COPS)

Portfolio Wealth (Target 20% REDD-COPS)

100
3-month T-Bill

10

But the level of REDD tolerance required (1/3) is still high. Both the average
and maximum exposure for the 1/3 case exceed 100%, indicating a leveraged (up to
22.3%) portfolio that must borrow in 3-month T-bill, or long S&P 500 Index futures, or
use a leveraged index ETF. From the curve in Figure 1 where Rolling Economic Max is
different from portfolio wealth level, we observe that the longest non-zero REDD lasted
3.5 years until June 2003, and maximum REDD of 29.4% occurred in Oct. 2008. Figure 2
shows the volatile cycle of risky asset allocation, and its high correlation (coefficient of
0.72) to the 12-month trailing returns. 13 of the 16 months with annual trailing losses
more than 20% would have occurred during the 2002-2003 and 2008-2009 bear markets.
We will mostly focus on the recent 20.5-year period (1990/12/31-2011/6/30) hereafter.
 Figure 2: 12-month Rolling Returns and Monthly Allocation in SPTR
For Calibrated 1/3 REDD-COPS (12/31/1950 – 6/30/2011)
80% 200%
1-Year Rolling Return Portfolio % in SPTR
180%

60%
160%

140%
40%
1-Year Rolling Return

SPTR Position size

120%

20% 100%

80%

0%
60%

40%
-20%

20%

-40% 0%

3.2 Effects of REDD-COPS Trading Frequency – Monthly rebalance sufficient

More frequent portfolio rebalancing for REDD-COPS increases transaction costs
and may not improve performance, especially when the stock market index shows less
one-step trending due to noises at a shorter time scale. We compare frequencies of
monthly, weekly and daily rebalancing between SPTR Index and 3-month T-bill.

Table 2: Rebalance Frequency Effects*:  = 30% Calibrated REDD-COPS (SPTR/T-Bill)

Annualized Annualized Sharpe Maximum Maximum Average Average Maximum
Return Std Deviation Ratio Drawdown REDD REDD Exposure Exposure
Monthly 9.61% 11.71% 0.521 32.11% 25.60% 4.90% 76.72% 101.53%
Weekly 9.02% 11.30% 0.488 31.01% 24.72% 4.86% 76.34% 101.53%
Daily 8.65% 11.09% 0.465 31.13% 25.41% 5.01% 75.16% 101.53%
SPTR 9.22% 14.89% 0.384 50.95% 46.76% 6.18% 100% 100%

* 20.5 year test period (from 1/2/1991 to 6/30/2011) is used. All portfolio risk metrics: standard
deviation, simple drawdowns and REDD, are measured in discrete time intervals with monthly
returns and month-end portfolio wealth levels, even when the rebalance trading is done daily or
weekly. Regardless of rebalance frequencies, the input parameters, and , in equation (5) for
SPTR are evaluated the same in the sample period from its monthly metrics listed in Table 3. One
may choose to use the same data frequency of rebalance to calculate input and  parameters.
However daily or weekly measured volatility  is more than that from monthly measurement,
which leads to less leverage according to equation (5), but worse REDD-COPS performance.
 The results are shown in Table 2 and Figure 3 for a 20.5 year (1991/01/2-
2011/06/30) test period. Monthly rebalance outcome is shown to out-perform daily or
weekly trading scheme in both returns and risk-adjusted returns. Active allocation of
equation (5) has a 1-step momentum trading nature. Rising SPTR Index reduces REDD
and raises SPTR allocation over next period, which will further gain if the uptrend
continue. On the other hand, decline in SPTR leads to higher REDD, lower portfolio
exposure or less losses should the downtrend continue. For the period 1/1991-6/2011, the
monthly return time series of SPTR has a one-step-lag auto-correlation coefficient of
0.118, compared to weekly series’ -0.081 and daily series’ -0.062. Thus the out-
performance of monthly rebalance can be explained by a stronger tendency of gain or
loss to continue for the next period, compared to a weekly or daily rebalanced scheme.

Figure 3: Dollar Growth of  = 30% Calibrated REDD-COPS (SPTR/T-bill)

7.0
Monthly Rebalance

6.0 Weekly Rebalance

Daily Rebalance
5.0

SPTR
Portfolio Wealth ($)

4.0

3.0

2.0

1.0

0.0

3.3 Two Risky Assets Case – Dominating Efficient Frontier

Over a 20.5-year test period, we use monthly data of S&P 500 Total Return Index
(SPTR) and Barclays Capital 20+ Year US Treasuries Index (TLT) to calibrate REDD-
COPS of two risky assets. The 3-month T-Bill is still used as the risk free asset.
 Table 3: MPT Parameters for SPTR and TLT (Monthly data from 1/1991-6/2011)
Return Correlation Annualized risk free rate
SPTR TLT (monthly) (average 3-month T-bills)
Annualized Return 9.22% 8.15%  -0.069 r f = 3.5%
Annualized Std Deviation 14.89% 10.73%

Figure 4: Return vs. Risk: Fixed Allocation Portfolios of SPTR and TLT (1/1991-6/2011)

10.0%
Efficient Frontier

9.8% MVP
MRP
9.6% RPP
Portfolios under MVP
9.4%
Portfolios below MRP
Annualized Return (%)

9.2%
60-40 (SPTR-TLT) Balanced Portfolio
SPTR Index
9.0%

8.8%

8.6%

8.4%

8.2%
Barclay 20+ Year US Treasuries Index (TLT)
8.0%
6% 7% 8% 9% 10% 11% 12% 13% 14% 15% 16%
Risk (Annualized Standard Deviation %)

Efficient frontier of SPTR and TLT from traditional fixed asset allocations is
plotted in Figure 4. With multi-period rebalance at end of every month, there is a
Maximum Return Portfolio (MRP) of SPTR (77.3%) and TLT (22.7%) blend that
achieved the highest annualized return among all fixed allocation portfolios over the 20.5
years. The portfolios between Minimum Variance Portfolio (MVP)5 and MRP form the
efficient frontier (solid blue curve in Figure 4).

5  12, 2  1 2
Two-asset fixed allocation MVP has portfolio weighting w1, 2  .When correlation
 12   22  2 1 2
coefficient < 0, w1 > 0 and w2 > 0 (no shorting) is guaranteed. For 1 >  > 0,  < leads to no
shortingWith SPTR and TLT index data over the recent 20.5 years (Table 3), MVP is a blend about 35.2%
in SPTR and 64.8% in TLT that has a minimal level of 8.41% in annualized standard deviation (Table 4).
 The two-asset Risk Parity Portfolio (RPP)6 on the efficient frontier turns out to be
very close to the tangential portfolio for the example period. Thus, a blend of 3-month T-
bill and RPP (green line in Figure 4) has better Sharpe ratio than any efficient frontier
portfolio. The blend can be a Leveraged Risk Parity Portfolio (LRPP) [10] that borrows 3-
month T-bill for over 100% exposure in RPP.

Figure 5: Return-Risk Profile of Calibrated REDD-COPS (1/1991-6/2011)

14%
REDD-COPS

Leveraged RPP

13% Leveraged MVP 25% REDD-COPS

Efficient Frontier

MVP
12%
Annualized Return (%)

MRP

RPP
20% REDD-COPS
11% Portfolios under MVP

Portfolios below MRP

10%

15% REDD-COPS
9%

8%
6% 7% 8% 9% 10% 11% 12% 13% 14% 15%

Risk (Annualized Standard Deviation %)

Calibrated REDD-COPS are tested for wide range of drawdown target  for the
20.5 years ending 6/30/2011. Figure 5 shows risk-return frontier of REDD-COPS (black
line) with equation (3) used to define REM in REDD7. Leverage can also be applied for

6  2,1
The two-asset RPP has portfolio weighting as w1, 2  . With parameters in Table 3 for SPTR and
1   2
TLT indexes for the recent 20.5 years, RPP is about 41.9% SPTR and 58.1% TLT blend, delivering 9.07%
in annualized return and 8.51% of annualized standard deviation – a Sharpe ratio of 0.655. That is a better
risk-adjusted return than the 0.598 Sharpe ratio of the 60-40 Balanced Portfolio (60% SPTR and 40% TLT
blend - rebalanced monthly) for the same period. Table 4 lists performance statistics of RPP.
7
If the simpler equation (2) is applied to define REM value in equation (4), the REDD-COPS frontier with
SPTR/TLT in the example period happens to be very close to match the LRPP frontier (Fig. 5’s green line).
MVP and RPP by borrowing at risk free rate. Practically, leverage can be implemented
through using T-bills as margin collateral and investing in index futures contracts of risky
assets for more than 100% the nominal portfolio value. Leveraged Risk Parity Portfolios
(LRPP – green line in Figure 5) and Leveraged Minimum Variance Portfolios (LMVP –
brown line in Figure 5) are extensions from their baseline fixed allocation portfolios.
Calibrated REDD-COPS frontier (black line in Figure 5) dominates LRPP, LMVP and
Efficient Frontier from fixed allocation (blue line) in the risk-return space over 20.5 years.
We summarize performance of the calibrated portfolios in Table 4 along with those for
the SPTR and TLT indexes.

Table 4: Performance Statistics of Two Risky Assets (SPTR and TLT) based
Calibrated REDD-COPS and Fixed Allocation Portfolios (1/1991-6/2011)
Single Index SPTR-TLT SPTR-TLT
Buy & Hold Fixed Allocation Portfolio REDD-COPS
60% 15% 20% 25%
60/40 Portfolio 60/40 Portfolio Levered REDD REDD REDD
SPTR TLT (no-rebalance) (monthly-rebalance) MVP RPP RPP Target Target Target
Annualized Return 9.22% 8.15% 8.82% 9.26% 8.97% 9.07% 12.16% 9.03% 10.91% 12.85%
Annualized Std Deviation 14.89% 10.73% 10.47% 9.64% 8.41% 8.51% 13.58% 8.32% 11.29% 14.46%
Max REDD 46.76% 21.51% 27.14% 26.68% 15.67% 16.25% 26.04% 13.06% 17.65% 22.42%
Average REDD 6.18% 4.79% 4.11% 3.52% 2.92% 2.87% 4.63% 3.38% 4.59% 5.88%
Max Drawdown 50.95% 21.40% 30.30% 28.63% 15.65% 17.22% 28.16% 14.89% 20.45% 26.07%
Sharpe Ratio 0.384 0.433 0.508 0.598 0.650 0.655 0.638 0.665 0.656 0.646
Average Total Exposure 100% 100% 100% 100% 100% 100% 160% 93.99% 127.84% 163.80%
Max Total Exposure 100% 100% 100% 100% 100% 100% 160% 125.47% 170.35% 218.04%
Ending Multiple (20.5 years) 6.098 4.985 5.653 6.149 5.820 5.934 10.515 5.886 8.362 11.915

From equation (6), active trading for Calibrated REDD-COPS is driven by portfolio
  REDD
scaling factor of , superimposed to allocations in risky asset class indexes SPTR
1  REDD
and TLT. Relative asset allocations in SPTR and TLT in Calibrated REDD-COPS are
decided by input parametersand , the same as MVP and LRPP portfolios. The ratio

between allocations to SPTR and TLT is

1  12  1  2  12  2  1 
40.61%
from
2  12  2  1  12  1   2 59.39%

equation (6), very close to the asset allocation ratios between SPTR and TLT for MVP and
RPP portfolios (see Footnote 5 and 6). Thus, the performance difference between
Calibrated REDD-COPS and MVP or LRPP is mostly due to market timing of drawdown
control to allocate between risky assets and risk free asset.
 With similar portfolio’s volatilities (annualized standard deviations), =15% REDD-
COPS returned more than MVP (comparing green columns in Table 4). =25% REDD-
COPS out-performed 60% leveraged RPP with less Maximum Drawdown (comparing
yellow columns in Table 4). Active trading through drawdown control mechanism in
Calibrated REDD-COPS adds value in the example 20.5-year back test period.

3.4 Three Risky Assets Case – Additional Diversification Benefit

In addition to SPTR and TLT Indexes, we introduce into Calibrated REDD-COPS a
third risky asset class index: Dow-Jones UBS-Commodity Total Return Index (DJUBS).
Despite the index’s lower Sharpe ratio (DJUBS = 0.157) over the 20.5 year period (since
inception in 1/1991) than that of SPTR or TLT, it can bring diversification benefit and
raise risk adjusted return when blending with a SPTR-TLT fixed allocation portfolio
under Mean-Variance Optimization (MVO)8.

Figure 6: Return/Risk of Calibrated REDD-COPS: More Risky Assets (1/1991-6/2011)

18%
3 Risky Assets REDD-COPS: SPTR, TLT & DJUBS
16%
2 Risky Assets REDD-COPS: SPTR & TLT 

14%
Single Risky Asset REDD-COPS: SPTR

12% 
Annulized Return


10%


8%  

6%


4%

2%

0%
0% 2% 4% 6% 8% 10% 12% 14% 16% 18% 20%

Annualized Standard Deviation

8
We examine that a relation among Sharpe ratios and correlation coefficient of the DJUBS Index and the
Tangential Portfolio (essentially Risk Parity Portfolio or RPP point on the SPTR-TLT efficient frontier in
Figure 4), DJUBS   DJUBS  RPP  RPP . This indicates that DJUBS is value adding under traditional MVO
framework over the 20.5 year test period.
 Over 20.5years, monthly return correlation coefficients among DJUBS, SPTR and
TLT are low:  DJUBS SPTR  0.275 ,  DJUBS TLT  0.085 and  SPTRTLT  0.069 . Applying these
correlation coefficients and the and parameters of SPTR and TLT (Table 4) and
DJUBS (Table 5) in equation A.3 (Appendix 2), we can rebalance the REDD-COPS
monthly among three risky asset class indexes SPTR-TLT-DJUBS and the risk free asset
3-month T-bill. Same as for the two risky assets case, Calibrated REDD-COPS with three
risky-assets only applies active trading to scale overall position size in risky assets by a
  REDD
factor . The relative weightings in three risky asset indexes, however, are fixed.
1  REDD
They have the ratios of: SPTR /TLT /DJUBS = C1/C2/C3 = 33.78%: 54.86%: 11.36%,
where C1, C2 and C3 are calculated according to equation A.3 (Appendix 2).

Including DJUBS in Calibrated REDD-COPS portfolio improves portfolio

efficiency to a better return-risk trade-off. As shown in Figure 6, under the same REDD
control limit, adding TLT to the SPTR based Calibrated REDD-COPS, achieves
significantly higher returns but also somewhat higher standard deviations. Adding
DJUBS to the SPTR-TLT based Calibrated REDD-COPS, results in higher annualized
returns and less portfolio return standard deviations when comparing Table 5 and Table 4.
Results of REDD-COPS in Table 5 use equation (3) for REM calculation. Realized
Maximum REDD breaks the drawdown control target slightly during 2009 in the 25%
and 30% example cases.

Table 5: Performance Statistics of Calibrated REDD-COPS of Three Risky Assets and

Buy-and-Hold SPTR, TLT or DJUBS Index (01/1991-06/2011)
Single Index SPTR-TLT-DJUBS
Buy & Hold REDD-COPS
10% 15% 20% 25% 30%
REDD REDD REDD REDD REDD
SPTR TLT DJUBS Target Target Target Target Target
Annualized Return 9.22% 8.15% 5.80% 7.19% 9.05% 10.95% 12.92% 14.96%
Annualized Std Deviation 14.89% 10.73% 14.70% 5.34% 8.10% 10.99% 14.08% 17.40%
Max REDD 46.76% 21.51% 54.48% 9.92% 14.93% 19.98% 25.09% 30.24%
Average REDD 6.18% 4.79% 7.34% 2.04% 3.09% 4.20% 5.37% 6.62%
Max Drawdown 50.95% 21.40% 54.26% 9.36% 14.46% 19.60% 24.77% 30.04%
Sharpe Ratio 0.384 0.433 0.157 0.690 0.686 0.678 0.669 0.659
Average Total Exposure 100% 100% 100% 71.07% 108.41% 147.67% 189.50% 234.69%
Max Total Exposure 100% 100% 100% 87.89% 133.53% 181.28% 232.04% 286.86%
Ending Multiple (20.5 years) 6.098 4.985 3.179 4.150 5.908 8.422 12.065 17.438
4. Risk-based Dynamic Asset Allocation to Control Drawdown

The Calibrated REDD-COPS using historical data has assumed three types of input
parameters of the risky assets as constants: Sharpe ratios (), correlation coefficients ()
and return standard deviations (). We now revisit the input parameters issue to design
REDD-COPS for practical implementation. Notice when correlation coefficients ij  0 ,

equations (6), A.1 and A.3 for multiple un-correlated risky assets are simplified. The
allocation weights all reduce to the form of equation (5) for each risky asset. We can
rewrite equation (5) for each un-correlated risky asset as:

 i 1  1   REDD t , H  
xi      Max  0,   (7)
  i t , h   1   1  REDD t , H  
2
2 

n
The allocation in risk free asset is still x f  1   xi . Equation (7) is an easily
i 1

implementable one-step forward trading rule: rebalance according to the portfolio value
and REDD at time t, and hold for the time period t , t  t  before the next rebalance trade
at time t  t  . The assumptions about capital markets are:

1. We can find n investable asset class indexes that are un-correlated over long term;

2. Long term expected Sharpe ratio of each risky asset class index is stable and can be
estimated as a constant in the dynamic asset allocation process of equation (7);

3. Each risky asset return follows random walk, but the return volatility i is not
constant and can be forecasted from returns during a preceding time window h.

In Appendix 3, these simplifications are examined with monthly market data in the
last 36 years (ending 12/30/2011) on three asset class indexes: SPTR, TLT and Goldman
Sachs Commodities Index (GSCI Total Return Index)9. Equation (7) is essentially a risk-
based dynamic asset allocation scheme, as  i t , h is the volatility risk measure updated at

9
DJUBS index has only 21 years of historical data (1991-2011). It has high correlation at about 0.898 to
the GSCI which has monthly data available since 1971.
time t from the most recent return series looking back time window h. Still, the dynamic
  1   REDD t , H 
asset allocation of  i   is further scaled by factor , the result of
  i t ,h  2  1  REDD t , H 

active portfolio position sizing to control drawdown. Additionally, investor drawdown

1
risk tolerance also weighs in as a leverage factor of . Similar to choosing look-back
1  2

period H for REDD, we will use h = 1 year, i.e. 12-month trailing standard deviation to
define the varying volatility risk in the examples.

We use SPTR, TLT and DJUBS indexes to back-test for the recent 20 years (1/1992-
12/2011). Based on historical statistics in Appendix 3, we use constant expected Sharpe
ratios  SPTR  0.4 ,  TLT  0.45 and  DJUBS  0.15 , and   20% as a baseline case. Close to
actual Sharpe ratios in Table 5, relative high expected capital market growth is assumed.
We call it Risk-based High 20% REDD-COPS.

Figure 7: Risk based High 20% REDD-COPS (  SPTR  0.4 ,  TLT  0.45 ,  GSCI  0.15 )
Monthly Rolling1-year Return & REDD (top) and Total Risk Exposure (bottom)
100%
1 Year Rolling Return
Rolling 1-year Return

80%

60% - REDD

40%

20%

0%

-20%

400%

350%
Total Exposure to Risky Assets

300%

250%

200%

150%

100%

50%

0%
 Figure 7 shows one year rolling return, REDD, and the portfolio total risk exposure in
tandem. One-step momentum trading is evident: increase (decrease) in total exposure to
risky assets follows the 1-year gain (loss) by one month. During 7 months (10/2008-
4/2009) of the last bear market, REDD-COPS would have less than 5% total risk
exposure. Overall, in 210 out of 240 months, the total portfolio exposure exceeded 100%.
Maximum leverage was close to 300% in 07/1993. Monthly average leverage is about
100% over last 20 years.

Figure 8: Risk Based High 20% REDD-COPS (1992-2011): Distribution of Relative

Portfolio Weights in Risky Assets (normalized to total 100%)

SPTR TLT DJUBS

100%

90%

80%

70%

60%

50%

40%

30%

20%

10%

0%

Within total risky asset exposure, distributions among the three indexes are shown in
Figure 8 under a normalized sum to 100%. Over 20 years, temporal variations for relative
allocations in each index are wide: SPTR (19.6% ~ 52.8%), TLT (29.2% ~ 66.5%) and
DJUBS (9.5% ~ 29.7%). REDD drawdowns were successfully controlled within 20%
over 20 years. REDD drawdown over 5% lasted 15 months in 7/2008 – 09/2009.

To address fluctuation of Sharpe ratios (see Appendix 3), we perturb each Sharpe
ratio input in equation (6) by up to +0.15 from the baseline case, to examine REDD-
COPS performance. For example, for capital market expectations of lower growth
referencing the 20-year Sharpe ratios of 1992–2011 (see Appendix 3), we use  SPTR  0.3 ,

 TLT  0.35 and  DJUBS  0.05 . At  = 20%, we call it a Low 20% REDD-COPS. The only
other parameter of equation (6) is the REDD target . We vary  = 15%, 20% and 25%.

Figure 9 shows the risk-based REDD-COPS has the best growth in the last two
decades while annual returns year by year are listed in Table 6. The Calibrated REDD-
COPS uses asset allocation equation A.3 with constant parameters of expected index
Sharpe ratios and volatilities (listed in Table 5), and correlation coefficients of Section
3.4. During the back-test of 20 years (1992-2011), the risk-based High 20% REDD-
COPS out-performed the Low 20% REDD-COPS and the Calibrated REDD-COPS.

Figure 9: Dollar Growth of Risk Based 20% REDD-COPS and Benchmarks (1992-2011)
14
SPTR

12 TLT

DJUBS

10
60/40 SPTR/TLT

Calibrated 20% REDD-COPS

8
High 20% REDD-COPS

6 Low 20% REDD-COPS

From return risk trade-off, risk-based REDD-COPS dominate Calibrated REDD-

COPS as shown in Figure 10. It also demonstrates the robustness of risk-based REDD-
COPS: higher capital market expectation makes little difference to the portfolio
efficiency between return and standard deviation or monthly average REDD, but does
lead to higher return for the same level of ex post Maximum REDD or Drawdown.
 Table 6: Annual Returns: Risk-based 20% REDD-COPS and Benchmarks (1992-2011)

1992 1993 1994 1995 1996 1997 1998 1999 2000 2001
SPTR 7.62% 10.08% 1.32% 37.58% 22.96% 33.36% 28.58% 21.04% -9.10% -11.89%
TLT 7.93% 18.00% -8.22% 32.59% -1.46% 16.18% 14.22% -10.06% 21.50% 3.63%
DJUBS 3.70% -1.07% 16.61% 15.21% 23.17% -3.39% -27.03% 24.35% 31.84% -19.51%
60/40 SPTR/TLT 7.81% 13.26% -2.53% 35.68% 12.73% 26.45% 23.63% 7.76% 2.45% -5.23%
Calibrated REDD-COPS 8.27% 21.32% -5.69% 50.66% 9.41% 25.86% 18.23% -0.06% 12.73% -5.70%
High REDD-COPS 5.85% 33.11% -6.23% 86.45% 21.80% 24.42% 8.37% -4.06% 22.62% -7.20%
Low REDD-COPS 6.59% 27.82% -6.09% 62.10% 16.19% 23.66% 14.55% -2.91% 18.35% -4.58%

2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
SPTR -22.10% 28.68% 10.88% 4.91% 15.79% 5.49% -37.00% 26.46% 15.06% 2.11%
TLT 17.00% 1.80% 8.99% 8.57% 0.93% 10.15% 33.72% -21.40% 9.38% 33.84%
DJUBS 25.91% 23.93% 9.15% 21.36% 2.07% 16.23% -35.65% 18.91% 16.83% -13.32%
60/40 SPTR/TLT -7.26% 17.70% 10.28% 6.60% 9.74% 7.63% -13.59% 5.34% 14.01% 15.06%
Calibrated REDD-COPS 4.56% 17.80% 14.25% 12.62% 6.06% 12.31% -14.49% -3.47% 23.45% 32.20%
High REDD-COPS 5.94% 20.27% 20.65% 14.46% 11.87% 14.64% -17.81% 3.53% 26.20% 22.07%
Low REDD-COPS 3.57% 15.61% 16.42% 11.26% 10.89% 12.36% -12.20% 0.46% 19.58% 19.33%

Figure 10: Risk-Return of Risk-based REDD-COPS (1992-2011) = 2% to 30%

20% 20%

Hi REDD-COPS Hi REDD-COPS
18% 18%

16% Low REDD-COPS 16% Low REDD-COPS

14% 14%
Calibrated REDD-COPS Calibrated REDD-COPS
Annualized Return
Annualized Return

12% 12%

10% 10%

8% 8%

6% 6%

4% 4%

2% 2%

0% 0%
0% 5% 10% 15% 20% 25% 0% 1% 2% 3% 4% 5% 6% 7% 8%

Annualized Std Deviation Average REDD

20% 20%

Hi REDD-COPS Hi REDD-COPS
18% 18%

16% Low REDD-COPS 16% Low REDD-COPS

14% 14%
Calibrated REDD-COPS Calibrated REDD-COPS
Annualized Return
Annualized Return

12% 12%

10% 10%

8% 8%

6% 6%

4% 4%

2% 2%

0% 0%
0% 5% 10% 15% 20% 25% 30% 35% 0% 5% 10% 15% 20% 25% 30%

Maximum REDD Maximum Drawdown

Table 7: Risk-based 20% REDD-COPS: Robustness to Market Expectations (1992-2011)
Sum of 3 Expected  's 1 0.7 1.15 1.15 1.15 1.05 1.05 1 1 1 1 0.9 0.85 0.85
 Expected  SPTR 0.4 0.30 0.4 0.4 0.55 0.4 0.55 0.25 0.25 0.4 0.55 0.4 0.25 0.4
 Expected  TLT 0.45 0.35 0.6 0.45 0.45 0.6 0.45 0.45 0.6 0.3 0.3 0.45 0.45 0.3
 Expected  DJUBS 0.15 0.05 0.15 0.3 0.15 0.05 0.05 0.3 0.15 0.3 0.15 0.05 0.15 0.15
Annualized Return 13.71% 11.89% 14.70% 14.67% 14.60% 14.28% 14.07% 13.71% 13.65% 13.31% 13.19% 13.25% 12.65% 12.44%
Annualized Std Dev 13.98% 11.42% 15.63% 14.97% 15.54% 15.01% 15.03% 13.50% 14.42% 13.68% 14.18% 13.30% 12.49% 12.43%
Sharpe Ratio 0.745 0.752 0.730 0.759 0.727 0.731 0.717 0.771 0.718 0.732 0.698 0.748 0.749 0.735
Average REDD 5.01% 4.06% 5.82% 5.34% 5.59% 5.81% 5.54% 4.99% 5.49% 4.86% 5.24% 4.95% 4.51% 4.44%
Max REDD 19.68% 18.10% 20.05% 20.64% 20.76% 19.32% 20.32% 19.45% 18.62% 20.31% 20.43% 19.01% 18.13% 19.41%
Max Drawdown 19.00% 17.44% 19.29% 20.05% 20.52% 19.88% 21.25% 18.63% 17.89% 19.72% 21.50% 19.46% 17.40% 19.03%
Maximum Exposure 396.8% 305.6% 450.3% 464.2% 453.6% 425.5% 419.2% 402.6% 396.0% 422.7% 412.2% 367.0% 342.4% 350.6%
Average Exposure 195.2% 161.2% 211.8% 215.8% 210.7% 196.1% 195.1% 196.7% 193.7% 194.3% 188.6% 178.9% 176.6% 173.1%
Ending Multiple 13.065 9.458 15.537 15.444 15.275 14.430 13.918 13.055 12.927 12.183 11.919 12.034 10.831 10.437

The example Risk-based REDD-COPS appears robust against variations of capital

market expectations, i.e., the input parameters of expected asset class Sharpe ratios. Table
7 indicates stable ex post risk-adjusted returns from risk-based REDD-COPS: Sharpe
ratios ranges only from 0.698 to 0.759 over the 20-year period, despite that sum of three
input index Sharpe ratios varies widely from 0.7 to 1.15. Larger total expected Sharpe
ratios input does result in higher returns and higher “Average REDD” due to larger
“Average Exposure” in risky assets. Some cases of the Risk based 20% REDD-COPS in
Table 7 lose control of Maximum REDD within the 20% target despite the extent of
“violation” is less than 1%. This is due to higher expected Sharpe ratios in riskier assets
SPTR and DJUBS, or larger sum of three Sharpe ratios input. The Low 20% REDD-
COPS (sum of input Sharpe ratios at 0.7) still has about 2% margin of safety to reach the
Max REDD or Max Drawdown limit of 20% during the 2008-2009 bear market.

5. Discussion

The drastic market cycle of sharp decline during the 2008-2009 financial crisis and
subsequent strong recovery in 2009-2011, provides us the opportunity to demonstrate that
the continuous drawdown controlled optimal portfolio theory of Grossman and Zhou
(1993) loses optimality under a straightforward discretized rebalancing extension. Our
proposal of Risk-based REDD-COPS depends on a shorter portfolio history by looking
back only a constant rolling time window. In what is essentially one-step momentum
trading at discrete time intervals, REDD-COPS avoids a cascading performance drag due
to the long memory of anchored drawdown, especially over a sharp V-shape market cycle.
 Intuitively, the REDD-COPS can time the market well when the market drawdown
cycle (from peak to bottom) is somewhat longer than (or at least matches) the rolling time
window defining REDD. Right before a market rally from low, a lower Economic Max in
the shorter time window results in larger allocation in risky assets to boost subsequent
returns. In empirical tests with market index data, we have used one year as the rolling
drawdown time window. During the 2008-2009 market crash, the drawdown cycle lasted
17 months for S&P 500 Index (SPTR) after its Oct. 2007 high, and 8 months for the Dow
Jones UBS Commodities Index (DJUBS) after its June 2008 high, until the March 2009
low. One major factor preventing further decline was the coordinated market intervention
from the world’s central banks. Subsequent liquidity injections from the Federal Reserve
also lifted US markets through mid-2011. Looking forward, a “Fed put” or strong
external support after a major crash, can avoid an immediate “second leg down”, and
more probably lead to a strong rebound. This mitigates the risk of shorter memory in our
opportunistic REDD control mechanism. Although the test examples use constant look-
back window to define REDD, a variable time window can be optimal due to macro-
economic cycle, monetary policy and speed of market impact from excess liquidities, etc.

Due to stronger serial correlation, monthly trading of SPTR in REDD-COPS has

better performance than weekly or daily rebalancing. Monthly rebalance makes
transaction cost less a concern10 and we use it extensively in REDD-COPS examples.

The three asset class indexes, SPTR, TLT and DJUBS, used in our back-test examples
have nearly zero correlations to each other overall in the past 20+ years. In practice, one
can perform Principal Component Analysis (PCA) over a pool of asset class indexes’
historical return series to find leading principal factors that are uncorrelated. Risk-based
active trading scheme of equation (7) can be applied more rigorously to the investable
principal factors, each of which is a combination of underlying asset class indexes.

10
Since 2010, some North American brokerages (e.g. TD Ameritrade), have allowed monthly trading of
broad index ETF’s (like IVV, TLT, DBC and BIL) free of transaction fee for retail accounts. This makes the
REDD-COPS with lower drawdown limit (<15%) practically possible for smaller account sizes.
Table 8: Performance of Risk-based High REDD-COPS versus Benchmarks (1992-2011)

SPTR-TLT SPTR-TLT-DJUBS
Single Index Buy & Hold Fixed Allocation Portfolio Risk-based REDD-COPS
60% 15% 20% 25%
Balanced Levered REDD REDD REDD
SPTR TLT DJUBS (60-40) MVP RPP Target Target Target
Annualized Return 7.81% 8.93% 5.65% 8.79% 9.04% 11.29% 11.06% 13.71% 16.45%
Annualized Std Deviation 15.00% 11.37% 15.23% 9.46% 8.40% 13.35% 10.29% 13.98% 17.90%
Sharpe Ratio 0.301 0.495 0.154 0.580 0.683 0.598 0.754 0.745 0.734
Average REDD 7.21% 5.64% 8.01% 4.31% 3.68% 5.45% 3.69% 5.01% 6.41%
Max REDD 46.76% 21.51% 54.48% 26.68% 20.51% 26.04% 14.65% 19.68% 24.85%
Max Drawdown 50.95% 21.40% 54.26% 28.63% 19.09% 28.16% 13.98% 19.00% 24.10%
Maximum Exposure 100% 100% 100% 100% 100% 160% 292.3% 396.8% 507.9%
Average Exposure 100% 100% 100% 100% 100% 160% 143.3% 195.2% 250.3%
Ending Multiple 4.501 5.533 3.001 5.389 5.641 8.489 8.143 13.065 21.014

Theoretically, the optimal relative asset allocation weights in REDD-COPS

formula (  1      1 portion in equation A.1) is the same as the unconstrained investor

T

wealth utility optimization extension from Modern Portfolio Theory (MPT). We

introduce active management into the relative asset allocation weights by allowing the
inverse of each asset class volatility to change at portfolio trading frequency looking back
a rolling period (e.g. h = 12 months). The rest of the relative asset allocation variables
among uncorrelated asset classes are simply their expected Sharpe ratios. We choose to
specify them as input parameters using long term historical estimations.

Traditional asset allocation process based on MPT struggled with sensitivities to

parameter inputs. For a wide range variation of constant Sharpe ratio parameter inputs,
performance of the Risk based REDD-COPS examples (with SPTR-TLT-DJUBS indexes)
is robust over 20 years (1992-2011). With the same REDD drawdown control mechanism,
efficient frontier of the Risk-based REDD-COPS even dominates that of the Calibrated
REDD-COPS which has constant volatility and correlation inputs from in-the-sample
estimations. This indicates that the dynamic asset allocation using active volatility
estimates adds value in the back-test period. Still, majority of the market timing gains can
be attributed to the active position sizing factor   REDD for REDD drawdown control.
1  REDD

Table 8 summarizes the performance comparison of Risk-based REDD-COPS (  SPTR  0.4 ,

 TLT  0.45 and  DJUBS  0.15 ) against traditional asset allocation portfolios and indexes. In
general, maximum drawdown or REDD are successfully controlled within target limit.
 We find relative high leveraged level of exposure in risky asset class indexes are
needed for Risk-based REDD-COPS implementation. Cost of borrowing for leveraged
investment in risky asset class indexes is assumed at risk free rate. In reality, leverage can
reduce the portfolio returns as funding cost can be higher than 3-month T-Bill rate. Time
decay in derivative prices can degrade performance even if the leverage is realized
through index futures contract or long dated index options. Still the Risk-based REDD-
COPS’ margin of out-performance is substantially high. For example, about 5% in
annualized excess return of Risk-based REDD-COPS over a passive asset allocation
benchmark portfolio with similar maximum drawdown loss, is achieved for the test
period 1992-2011(see Table 8), much more than the potential cost of market friction from
extra leverage transactions. In practice, a retail investor can implement the example risk-
based REDD-COPS of lower drawdown target ( < 15%) with leveraged index ETF and
mutual funds. For institutional investors or managers of global macro or managed futures
strategies, REDD-COPS with higher drawdown limit ( > 20%) can be considered given
the access to higher leverage through derivative markets.

6. Conclusion

At the highest level of investment management, asset allocation is responsible for

most of the portfolio performance [11]. As an active asset allocation framework, our Risk-
based REDD-COPS proposed three new methodologies for practical portfolio
management: 1. Rolling Economic Drawdown (REDD) for portfolio loss control; 2. Risk
aversion defined as the dynamic floor of maximum loss tolerance; 3. Dynamic asset
allocation implemented through the inverse of asset class volatility estimation.

Besides the benefit of a clear mandate to construct suitable client portfolios, the
methodologies of REDD-COPS can be applied to design invest-able securities, such as a
principal-guaranteed investment product, target risk ETF, and target-date asset allocation
mutual fund with a glide path, etc. With active trading mechanisms derived solely from
risk statistics and risk target, the current proposal of REDD-COPS demonstrates the
prowess of risk based investment management: driven by active risk control, rather than
performance chasing, the portfolio will out-perform, consistently, for the long term.
References:

[1] Grossman, Sanford and Zhou, Zhongquan (1993), “Optimal Investment Strategies for
Controlling Draw-downs”, Mathematical Finance Vol. 3 (3), pp. 241-276.

[2] Cvitanic, Jasksa and Karatzas, Ioannis (1995), “On Portfolio Optimization under
Drawdown Constraints”, IMA Lecture Notes in Mathematics & Applications 65, pp. 77-
88.

[3] Chekhlov, A., Uryasev S., and M. Zabarankin (2005), “Drawdown Measure in
Portfolio Optimization”, International Journal of Theoretical and Applied Finance, Vol.
8, No. 1, pp. 13–58

[4] Zhou, G. F and Zhu, Y. Z., (2010) “Why is the Recent Financial Crisis a ''Once-In-A-
Century' Event?” Financial Analysts Journal, Vol. 66, No. 1, 2010

[5] Markowitz, H. M. (1952), “Portfolio Selection”, Journal of Finance, Vol. 7 (1952) 1,

77-91.

[6] F. Black and A.F. Perold (1992), "Theory of constant proportion portfolio insurance",
Journal of Economic Dynamics and Control, 1992, Vol. 16, issue 3-4, pp. 403-426

[7] Merton, R. C., “Optimal Consumption and Portfolio Rules in a Continuous Time
Model”, Journal of Economic Theory, Vol. 3 (1971), pp. 373-413.

[8] He, Guang-Liang (1999), “Drawdown Controlled Optimal Portfolio Selection with
Linear Constraints on Portfolio Weights”, working paper abstract on SSRN
([Link] and private communication (2011).

[9] Klass, Michael J. and Krzysztof Nowicki (2005), “The Grossman and Zhou
investment strategy is not always optimal”, Statistics and Probability Letters (74), pp.
245-252.

[10] Qian, Edward (2006), “On the Financial Interpretation of Risk: Risk Budgets do Add
up”, Journal of Investment Management, Vol. 4, No. 4, Quarter 4, 2006.

[11] Brinson, P., L. Randolph Hood and Gilbert L. Beebower (1986), “Determinants of
Portfolio Performance”, Financial Analysts Journal (July/August 1986). Vol. 42, No. 4,
pp. 39-44
 Appendix 1

Historical Data Demonstration: Discrete DD-COPS loses Optimality

Under the same constraint as discrete DD-COPS that Economic Drawdown (EDD)
stays below a target level another tradable allocation scheme is needed to show better
long term accumulated return. Here is the set up:

 Use S&P 500 Total Return Index (SPTR) as the single risky asset, and 3-month
US T-bill as the risk free bond. Take  = 20% as the target of EDD constraint.
 The reference alternative portfolio assumes a fixed allocation, 30% in SPTR and
70% in 3-month T-bill. Call it the 30-70 (Stock/T-Bill) Portfolio.
 Both DD-COPS and 30-70 (Stock/T-Bill) Portfolio uses monthly rebalance, trades
at market close value of SPTR Index on last trading day of each month.
 The time span is from 1/2/1971 to 6/30/2011. In equation (1), the long term return
R, standard deviation and risk free rate r f , are constant parameters evaluated

over the period (see Table A1).

Figure A.1: Growth of 20% DD-COPS vs. 30/70 Stock/T-Bill Portfolio for 40.5 Years
20
3-month T-Bill

18 SPTR

Economic Max (20% DD-COPS)

16
Portfolio Wealth (20% DD-COPS)
14
Economic Max (30-70 Portfolio)

12 30-70 (Stock/T-Bill) Portfolio Wealth

10

0
 From the above Figure, DD-COPS did not make any new Economic Max after
2000. Even when DD-COPS value reached a new high in late 2007, it is still below the T-
bill yield compounded high watermark from year 2000. Although the finite EDD
extended into the 2008-2009 market decline helped to limit the Max Drawdown to just
12%, it further reduced stock index exposure during 2009-2011 when S&P 500 Index
rose sharply. As a result of the cascading effect, the DD-COPS never made a new high in
2011 while the 30/70 Stock/T-bill portfolio did.

Table A1: 20% DD-COPS vs. the 30-70 Stock/T-Bill Portfolio and SPTR Index
Performance Statistics for 40.5 Years (1971-2011/6/30)
30-70 20% Target
SPTR Stock/T-Bill DD-COPS
Annualized Return 10.18% 7.27% 7.10%
Annualized Std Deviation 15.51% 4.69% 4.89%
Max EDD 55.65% 19.77% 18.31%
Average EDD 15.59% 4.39% 8.41%
Max Drawdown 50.95% 17.35% 12.01%
Sharpe Ratio 0.291 0.343 0.294
Average Total Exposure 100% 30% 30.61%
Max Total Exposure 100% 30% 49.50%
Ending Multiple (40.5 years) 50.69 17.18 16.10
Skew -0.448 -0.454 -1.037
Excessive Kurtosis 1.996 2.135 7.620

As shown in Table A1, the 40-year annualized return and “ending multiples” from
the 20% target DD-COPS turns out to be less than that from the referenced fixed
allocation scheme of the 30-70 Stock/T-Bill Portfolio. This example thus demonstrates a
situation that DD-COPS loses optimality in the long term portfolio growth rate.
 Appendix 2

Derivation of Multiple Risky Assets REDD-COPS Allocation Weights

Taking    in Cvitanic and Karatzas (1994) [2], the column vector [x] of risky
asset allocation weights of an REDD-COPS with multiple risky assets, is given by:

xT 

T



   1     1   max 0,
1

  REDD 
 1   2 1  REDD 
 (A.1)

where the matric is the lower triangular Cholesky decomposition of the return
 
covariance matrix   ij i j    T . The  vector is the return drift of risky assets

that each component i  Ri  r f  12  i2 where Ri is i’th risky asset’s long term return.

 1 0   1 1 0 
We have   
  2
 ,  1  
1    2 
2

   1 1  
2
 1  2 1   
2

for case of two risky

x  R r
assets. Substitute them into equation A.1 for  1  . With 1,2  1,2 f , it yields eqn. (6):
 x2   1,2

 x1  1  1  12  1    2  12     2   1   1    REDD 
x   2   1       1       
 Max 0,  
 2 1   2  1  1  REDD 
2
2 2 1 2 1 2

For three-dimensional case (three risky assets), we have:

 1 0 0 
 
   2 12  2 1  122 0 
(A.2)
 
 3 13  3  23  1213  1  12
2
 3 1  132   23  1213 2 1  122  


We can derive  and substitute into equation A.1, after some algebra, we have:
 x1   C1 
x   1     REDD 
 C2   Max 0,
1
 2 1 2 
2  
 x3  C3   1    1  REDD 
(A.3)
 2  122  23
2
 31
2
 2122313

   
 C1   1  12  1  1   23  2  12  2  2313  12   3  12  3  2312  13   1 
2

 
13 
C      1   1   2    1          1        
 2  2 2 2 1 2 1 13 23 12 3 2 3 13 12 23  2

   
C3   3  12  3  1  122  1  12  1  2312  13   2  12  2 1312   23   3 
 Appendix 3
Historical Correlation and Sharpe Ratios on Three Asset Class Indexes
We take a rolling window of look-back 20 years and compute monthly return
correlation coefficients at the end of each month. There are three pairs among three asset
class indexes: SPTR, TLT and GSCI – all three using the total return index series. Since
the TLT index monthly return data started the latest of the three in Jan 1976, we only have
this 20-year rolling correlation for the past 16 years ending Dec. 2011. See Figure A.2
below.

Figure A.2: History of the Rolling 20-year Return Correlation Coefficients

0.5
SPTR-GSCI

0.4
TLT-GSCI

0.3
SPTR-TLT

0.2

0.1

-0.1

-0.2

-0.3

-0.4

-0.5
Jan-91 Jan-96 Jan-01 Jan-06 Jan-11

During 16 years, the arithmetic averages of the 20-year trailing correlation

coefficients are  SPTRGSCI  0.019 ,  SPTRTLT  0.181 and  TLT GSCI  0.051 . The correlation
coefficients are also bounded between -0.2 and 0.4 at all sampled time. So for a 20-year
window, the un-correlated or low correlation assumption about SPTR, TLT and GSCI
indexes is held reasonably well in historical observations.
 We take the same approach of 20-year look-back to compute Sharpe ratios of the
three indexes at the end of each month. The risk free rate from 3-month T-bill is also
computed from the compounded return during the 20-year window. Risky asset class
index’s return (R) and standard deviation (), and risk free rate (rf) are all annualized to
compute Sharpe ratio (). See Figure A.3 below for the 21-year history of the Sharpe
ratios of SPTR and GSCI indexes and 16- year history of TLT index Sharpe ratio. The
averages are  SPTR  0.446 (on 252 monthly samples)  TLT  0.432 (on 192 monthly samples)
and  GSCI  0.166 (on 252 monthly samples). Notice the fluctuations and range for Sharpe
ratios. Thus we also compute the standard deviations of the Sharpe ratios from these
samples.   SPTR  0.146 ,   TLT  0.131 ,   GSCI  0.127 .

Figure A.2 Histories of 20-year Rolling Asset Class Index Sharpe Ratios

0.8
SPTR

0.7
TLT

0.6 GSCI

0.5

0.4

0.3

0.2

0.1

-0.1
Jan-91 Jan-96 Jan-01 Jan-06 Jan-11