1

Toward Determining the Optimal

Investment Strategy for Retirement ∗
Javier Estrada
IESE Business School, Department of Finance, Av. Pearson 21, 08034 Barcelona, Spain
Tel: +34 93 253 4200, Fax: +34 93 253 4343, Email: jestrada@iese.edu

Mark Kritzman
Windham Capital Management, 800 Boylston Street, 30th Floor, Boston, MA 02199, USA
Tel: +1 617 419-3900, Fax: +1 617 236-5034, Email: mkritzman@windhamcapital.com

Abstract. Investors who are about to retire are first and foremost concerned with supporting their
spending needs throughout retirement. But they also derive satisfaction from growing their wealth
beyond what is needed to support consumption in order to leave a bequest to their heirs or chosen
charities. The predominant metric for evaluating retirement investment strategies is the failure rate.
However, it fails to distinguish between strategies that fail early in retirement from those that fail
near the end of retirement. Moreover, it fails to account for potential bequests. To overcome these
shortcomings we propose a new metric, the coverage ratio, which is more comprehensive and
informative than the failure rate. In addition, we propose a utility function to evaluate the coverage
ratio, which penalizes shortfalls more than it rewards surpluses. Finally, we use the framework we
propose to determine the optimal allocation to stocks and bonds using both historical and simulated
returns.
December, 2018

1. Introduction
The most important attribute of a retirement investment strategy is that it provides a
return that is sufficiently large and reliable to support spending throughout retirement. Beyond
this primary concern, investors also derive satisfaction from generating surplus wealth in order
to leave a bequest to their heirs or chosen charities.
A critical issue for retirees and advisors is to determine whether a retirement investment
strategy is better than another, and ultimately which is the best one. The predominant metric for
such evaluation is the failure rate, which measures how frequently a strategy failed to sustain a
withdrawal plan over all the (historical or simulated) retirement periods considered. This failure
rate has at least two critical shortcomings. First, it fails to distinguish failures that occur near the
beginning of retirement from those that occur near the end. Second, it fails to account for
surpluses that could be left as a bequest.

∗
We would like to thank Jack Rader for his comments. Fran García and Cel Kulasekaran provided valuable
research assistance. IESE’s Center for International Finance (CIF) kindly provided partial support for this
research. The views expressed below and any errors that may remain are entirely our own.

Electronic copy available at: https://ssrn.com/abstract=3303153

 2

To overcome these limitations of the failure rate, we propose a new metric for evaluating
retirement investment strategies called the coverage ratio, which measures the number of years
of withdrawals (during and after retirement) supported by a strategy, relative to the length of the
retirement period considered. In addition, we recognize that investors are more averse to failures
than they are attracted to surpluses. We, therefore, propose a particular utility function for
evaluating the coverage ratio which penalizes failures more than it rewards surpluses. Finally, we
apply our approach to evaluate static investment strategies based on both historical and
simulated returns.
For the sake of concreteness, we illustrate our approach by focusing on the retirement
period, and particularly on the choice among competing asset mixes. However, our approach is
more general and could be used to choose an optimal withdrawal rate during the retirement
period, or a saving rate or asset allocation during the accumulation period.
The remainder of the paper is organized as follows. In Section 2 we introduce the coverage
ratio, and we show how to evaluate it based on a kinked utility function, which we believe is a
plausible description of investor preferences. In Section 3 we apply our framework to 11 static
strategies based on historical returns across 21 countries and the world market. In Section 4 we
apply our framework using simulation to evaluate static strategies given different assumptions
for the length of the retirement period and the distribution of stock returns. We conclude in
section 5.

2. The Coverage Ratio

Both academics and practitioners have invested considerable effort toward deriving
measures to help retirees and advisors determine the appropriate blend of risky and safe assets
for managing wealth throughout retirement. The most common measure is the failure rate, which
is the fraction of all the (historical or simulated) retirement periods considered in which an
investment strategy failed to support spending through the end of retirement. An important
shortcoming of the failure rate is that it assumes investors are indifferent between running out of
money early in retirement or near the end of retirement. In addition, it fails to account for
potential surpluses that could be applied toward a bequest.1

1 To overcome the first limitation, Estrada (2017) introduced the concept of shortfall years to complement
the failure rate; the latter measures how often a strategy failed, and the former measures by how much it
failed. To overcome the second limitation, and therefore to account for funds left as bequest, Estrada
(2018a and 2018b) introduced the risk-adjusted success ratio (RAS) and the downside risk-adjusted
success ratio (D-RAS), both of which are linked to the variable ‘years sustained’ (i.e., the average number
of years a strategy sustained withdrawals both when failing and when succeeding). The coverage ratio we
introduced in this article is a simpler and more intuitive version of the years sustained variable, and the
utility function we consider here accounts for risk as both RAS and D-RAS aim to do but further takes into
account an investor’s attitude toward risk

Electronic copy available at: https://ssrn.com/abstract=3303153

 3

To overcome both limitations, and ultimately to enable a more comprehensive evaluation

of retirement strategies, we suggest that investors and advisors focus on the metric we introduce
here, the coverage ratio, which aims to capture the number of years of withdrawals (during and
after retirement) supported by a strategy, relative to the length of the retirement period
considered. Importantly, our coverage ratio accounts for the years a strategy sustained
withdrawals both when failing (not supporting withdrawals through the entire retirement
period) and when succeeding (leaving a bequest).
Formally, let Yt be the number of years of inflation-adjusted withdrawals sustained by a
strategy, both during and after the retirement period, and L be the length of the retirement period
considered. Then we define the coverage ratio in retirement period t (Ct) as

Ct = Yt/L (1)

By definition, C<1 indicates that the strategy depleted the portfolio before the end of the
retirement period; C>1 indicates that the strategy sustained withdrawals through the entire
retirement period and left a bequest; and C=1 indicates that the strategy sustained withdrawals
exactly through the end of the retirement period and left no bequest.
To illustrate, consider a 30-year retirement period, a $1,000 retirement portfolio, annual
inflation-adjusted withdrawals of $40, and three strategies. The first strategy depletes a portfolio
in 24 years, the second does so in exactly 30 years, and the third sustains withdrawals for 30
years and leaves a bequest of $240 (which can support another six years of $40 withdrawals).
Then, Yt would be 24, 30, and 36, for the first, second, and third strategies, and Ct would
respectively be 0.8, 1.0, and 1.2.
However, the coverage ratio by itself is incomplete for evaluating the suitability of
alternative investment strategies, because investors are much more likely to be displeased with
outcomes that fall short of fully supporting retirement spending than pleased with outcomes that
generate surpluses. In order to account for this asymmetry in preferences, we evaluate a
strategy’s coverage ratio within a utility-based framework. We propose a kinked utility function
that assumes as the coverage ratio increases above 1, the investor derives increasing satisfaction
but at a decreasing rate, and as the coverage ratio falls below 1, the investor experiences a steep
and linear decline in utility. Formally, our utility function is given by

𝑈(𝐶) = for C ≥ 1

𝑈(𝐶) = − 𝜆(1 − 𝐶) for C < 1 (2)

Electronic copy available at: https://ssrn.com/abstract=3303153

 4

where U denotes utility; C denotes the coverage ratio; γ is the coefficient of risk aversion, which
determines the curvature of the slope when C>1; and λ is a linear penalty coefficient when C<1.
This utility function is depicted in Exhibit 1.

Exhibit 1: Kinked Utility Function of the Coverage Ratio

0.10

0.05
Utility

0.00

-0.05

-0.10
0.90 0.95 1.00 1.05 1.10
Coverage Ratio

This utility function has a very appealing property when we locate the kink at a coverage
ratio of 1. When a strategy exactly funds a retirement period (C=1), utility equals zero. When a
strategy leaves a bequest (C>1), utility is positive and increases as the coverage ratio increases
above 100% but at a diminishing rate. And when a strategy fails before the end of the retirement
period (C<1), utility declines steeply and linearly. If we were to set the kink below 1, the coverage
ratio would convey negative utility before the penalty function takes effect. We would thus be
adding an unintended penalty, and one that is difficult to interpret; it is therefore very clean to
locate the kink at 1.
While we are confident of the general form of our proposed utility function, there are no
universally appropriate values for the curvature of the utility function above the kink or the slope
of the penalty function below the kink. Higher values of γ, which defines the degree of curvature,
cause utility to increase more slowly as the coverage ratio increases; and higher values for λ
penalize spending shortfalls more severely than lower values.

3. Historical Results
We first apply our approach using returns from the Dimson-Marsh-Staunton database,
described in detail in Dimson, Marsh, and Staunton (2002, 2016). It contains annual returns for
stocks and long-term government bonds over the 1900-2014 period for 21 countries and the
world market. Returns are adjusted by each country’s inflation rate, and they are expressed in
local currency (except for the world market, in dollars). All returns account for capital gains and
losses as well as cash flows such as dividends and coupons.

Electronic copy available at: https://ssrn.com/abstract=3303153

 5

The analysis is based on a $1,000 portfolio at the beginning of retirement, a 4% initial

withdrawal rate, annual inflation-adjusted withdrawals, and a 30-year retirement period. At the
beginning of each year the annual withdrawal is made, the portfolio is then rebalanced to the
target asset allocation for the year, and then it compounds at the observed return of stocks and
bonds for that year. This process is repeated at the beginning of each year during the 30-year
retirement period, at the end of which the portfolio has a terminal wealth or bequest that may be
positive or 0. The first 30-year retirement period considered is 1900-1929 and the last one is
1985-2014, for a total of 86 rolling (overlapping) periods.
The analysis considers 11 static (annually-rebalanced) stock-bond allocations ranging
from 100% (all stocks, no bonds) to 0% (no stocks, all bonds), with nine allocations (90%, 80%,
…, 20%, 10%) in between, all indicated by the proportion of stocks in the portfolio, with the
balance allocated to bonds.
For each of the 11 strategies and 22 markets we consider we proceed as follows. We first
calculate the coverage ratio for each of the 86 rolling 30-year retirement periods in our sample.
Then we calculate the utility an investor derives from each coverage ratio. Finally, we calculate
expected utility by averaging the utilities in the previous step across the 86 retirement periods.
This process yields the average utility an investor obtains from a given strategy in a given market.
We implement the process using the kinked utility function described earlier, assuming that the
risk aversion coefficient γ equals 0.9999, and the linear penalty coefficient λ equals 10.2 Exhibit 2
summarizes the results of our empirical analysis.

Exhibit 2: Optimal Stock Allocation Based on Historical Returns

This exhibit shows the allocation to stocks (S) selected by the utility function in (2) for γ=0.9999 and λ=10, with the
rest allocated to bonds. The 11 asset allocations considered range between 100 (all stocks) and 0 (no stocks), with nine
allocations (90, 80, …, 20, 10) in between. All strategies are evaluated over 86 rolling 30-year retirement periods
between 1900-1929 and 1985-2014; a starting capital of $1,000; a 4% initial withdrawal rate; subsequent annual
withdrawals adjusted by inflation; and annual rebalancing.
Country S (%) Country S (%)
Australia 100 Netherlands 90
Austria 80 New Zealand 100
Belgium 100 Norway 90
Canada 100 Portugal 60
Denmark 90 South Africa 100
Finland 100 Spain 70
France 100 Sweden 60
Germany 100 Switzerland 70
Ireland 100 UK 100
Italy 100 USA 100
Japan 90 World 100

2As γ approaches 1, utility equals the natural logarithm of the coverage ratio; hence, we are effectively
using a log-wealth utility function for the coverage ratio above 1.

Electronic copy available at: https://ssrn.com/abstract=3303153

 6

Exhibit 2 reveals that our approach results in the selection of relatively aggressive
strategies, with an average allocation of 91% to stocks and 9% to bonds. In over half of the
markets, including the U.S. and the world market, the strategy selected is the most aggressive of
those considered, namely, 100% stocks. It is important to note that the optimal allocations we
report for each country do not suggest what each country’s future optimal allocation will be,
unless its historical market conditions prevail into the future. We report results across many
countries in order to provide a broad perspective and to show how results differ based on
different market conditions. Moreover, we recognize that our results are based on overlapping
retirement periods and are, therefore, less reliable than they would be if we had data for many
independent retirement periods.
As we already mentioned, we are confident about the shape of the utility function we
propose, and we think our choice of parameters for the base case (γ=0.9999 and λ=10) is sensible.
That said, in Exhibit 3 we explore the sensitivity of the results in our base case to changes in the
value of these two parameters, just for the U.S. market. As the exhibit shows, and as expected, the
optimal asset allocation does become more conservative as either γ or λ increase. However, even
for high values of both coefficients, the optimal strategy never allocates less than 80% to stocks.

Exhibit 3: Sensitivity Analysis (U.S. Market)

This exhibit shows the allocation to stocks (in %) selected by the utility function in (2) for different values of the risk
aversion (γ) and penalty (λ) coefficients, with the rest allocated to bonds. The 11 asset allocations considered range
between 100 (all stocks) and 0 (no stocks), with nine allocations (90, 80, …, 20, 10) in between. All strategies are
evaluated over 86 rolling 30-year retirement periods between 1900-1929 and 1985-2014; a starting capital of $1,000;
a 4% initial withdrawal rate; subsequent annual withdrawals adjusted by inflation; and annual rebalancing.
λ γ (Degree of Curvature)
(Penalty) 0.5 0.9999 1.5 2.0 3.0 5.0
10 100 100 100 90 80 80
30 100 90 90 80 80 80
50 100 90 80 80 80 80
70 90 80 80 80 80 80
90 90 80 80 80 80 80

We suspect that our approach yields relatively aggressive strategies for two reasons.
First, the equity risk premium was relatively high in most markets in our sample; and second, we
use a 30-year retirement period, and for such long investment horizons stocks have far more
often than not outperformed bonds.3

4. Simulated Results
We now present results based on simulated returns, which have several advantages
relative to results that are based on actual historical returns. First, simulation allows us to apply

3 Across the 21 markets in our sample, the average arithmetic mean return, geometric mean return, and
standard deviation were 7.3%, 4.6%, and 23.9%.

Electronic copy available at: https://ssrn.com/abstract=3303153

 7

our approach using different assumptions for the distribution of stock returns. Moreover,
simulated retirement periods are independent of each other unlike historical retirement periods
which overlap; hence, the simulated results are more statistically reliable. Finally, simulation
allows us to eliminate any biases that may have arisen from the particular sequence of returns
that occurred historically.
The sequence of returns could have a drastic impact on the progression of wealth
throughout retirement. In the early years of retirement, there is greater wealth because less of it
has been spent. Thus, relatively high (low) returns in the early (later) years will favor (hinder)
the compounding of capital. Whereas historical returns allow only for the return sequences that
actually occurred, simulation captures a much broader range of potential return sequences.
We employ Monte Carlo simulation to generate 25,000 annual real returns for stocks and
bonds drawing from two uncorrelated normal distributions, one with a particular mean and
standard deviation to characterize stocks and another with different mean and standard
deviation to characterize bonds. We always assume that the distribution of bond returns has a
mean (real) return of 2% and a standard deviation of 3%. However, we allow the mean (real)
return and standard deviation for stock returns to vary.
For each set of the 25,000 simulated returns we compute for each 30-year retirement
period the coverage ratio for stock allocations ranging between 0% and 100%, in 10%
increments, with the balance allocated to bonds. As in the historical analysis, we assume a $1,000
retirement portfolio, an initial withdrawal rate of 4%, and subsequent annual withdrawals
adjusted by inflation.
In our base case we compute the utility of each strategy’s coverage ratio for each of the
25,000 trials assuming, as in the historical analysis, γ=0.9999 (essentially log-wealth utility for
coverage ratios higher than 1) and λ=10 (penalizing coverage ratios lower than 1 more than we
reward coverage ratios higher than 1). We store these utilities for each of the 25,000 trials and
compute the average utility for each of the mixes of stocks and bonds. We then identify the asset
allocation with the highest average utility. Finally, we repeat this exercise for stock and bond
return distributions that have different combinations of means and standard deviations for
stocks. Exhibit 4 presents these results.
Exhibit 4 reveals that the optimal allocation to stocks is significantly lower than the stock
allocations suggested by historical returns when we assume lower mean returns and standard
deviation for stocks than those that prevailed historically. These simulated results may provide
better guidance for investors who believe that the historical record represents a relatively
favorable pass through history that is unlikely to recur.

Electronic copy available at: https://ssrn.com/abstract=3303153

 8

Exhibit 4:
Optimal Stock Allocation Based on Simulated Returns for 30-Year Retirement Periods
This exhibit shows the allocation to stocks (in %) selected by the utility function in (2) for γ=0.9999 and λ=10, with the
rest allocated to bonds, given different values for the mean and standard deviation of stock returns. The 11 asset
allocations considered range between 100 (all stocks) and 0 (no stocks), with nine allocations (90, 80, …, 20, 10) in
between. All strategies are evaluated over 86 rolling 30-year retirement periods between 1900-1929 and 1985-2014;
a starting capital of $1,000; a 4% initial withdrawal rate; subsequent annual withdrawals adjusted by inflation; and
annual rebalancing. The mean and standard deviation for bond returns are assumed to be 2% and 3%.
Stocks Standard Deviation of Stocks
Mean 15% 20% 25% 30%
3% 50 30 20 10
4% 100 60 40 30
5% 100 80 60 40
6% 100 100 70 50

We next explore whether the aggressive stock allocations suggested by historical returns
may result, in part, from our choice of a 30-year retirement period. We recognize that many
investors may be well into retirement and hence face a shorter horizon than 30 years. We
therefore produce results for horizons of 20, 10, and five years. We follow the same steps in our
simulation as we described earlier with one important exception. The initial wealth for each of
these shorter horizons is determined by the strategy returns and withdrawals that led up to that
point in time assuming the strategy was originally initiated at the inception of a 30-year horizon.
More precisely, when we simulate each of the 25,000 paths for a particular stock-bond
allocation for a remaining 20-year horizon, the wealth at the beginning of each path is determined
by the performance of the stock-bond allocation, net of the annual withdrawals, during the prior
10 years. Obviously, the level of wealth 10 years into an individual’s retirement period will be
specific to the returns and spending that actually occurred. Our results are based on 25,000 paths;
hence they represent what one might expect on average.
Exhibit 5, together with Exhibit 4, reveals that the optimal allocation to stocks differs in
accordance with conventional wisdom. If we hold constant standard deviation and horizon, the
optimal stock allocation rises with the expected return of stocks. If we hold constant the expected
return of stocks and horizon, the optimal stock allocation falls as the standard deviation of stock
returns increases. And if we hold constant both the expected return and standard deviation, the
optimal allocation to stocks falls as the retirement period becomes shorter. In some cases optimal
allocations do not change from one cell to another; this is because we only allow the asset
allocation to change in 10% increments. If we were to allow for more granular changes in the
stock-bond mix, we would see differences across all combinations of expected return, standard
deviation, and retirement period.

Electronic copy available at: https://ssrn.com/abstract=3303153

 9

Exhibit 5:
Optimal Stock Allocation Based on Simulated Returns for Different Retirement Periods
This exhibit shows the allocation to stocks (in %), for different lengths of the retirement period, selected by the utility
function in (2) for γ=0.9999 and λ=10, with the rest allocated to bonds, given different values for the mean and standard
deviation of stock returns. The 11 asset allocations considered range between 100 (all stocks) and 0 (no stocks), with
nine allocations (90, 80, …, 20, 10) in between. All strategies are evaluated over 86 rolling 30-year retirement periods
between 1900-1929 and 1985-2014; a starting capital of $1,000; a 4% initial withdrawal rate; subsequent annual
withdrawals adjusted by inflation; and annual rebalancing. The mean and standard deviation for bond returns are
assumed to be 2% and 3%.
20-Year Horizon (Years 11-30)
Stocks Standard Deviation of Stocks
Mean 15% 20% 25% 30%
3% 40 20 10 10
4% 70 40 30 20
5% 100 60 40 30
6% 100 70 50 40
10-Year Horizon (Years 21-30)
Stocks Standard Deviation of Stocks
Mean 15% 20% 25% 30%
3% 20 10 10 0
4% 30 20 10 10
5% 40 20 20 10
6% 60 30 20 10
5-Year Horizon (Years 26-30)
Stocks Standard Deviation of Stocks
Mean 15% 20% 25% 30%
3% 10 10 0 0
4% 10 10 10 0
5% 20 10 10 10
6% 40 20 10 10

It is worth noting that although the sensitivity of the optimal allocation to stocks to the
length of the retirement period is consistent with conventional wisdom, it appears to run counter
to Samuelson’s famous contradiction to the notion of time diversification (Samuelson, 1963). In
this classic paper Samuelson argued that the optimal mix of risky and safe assets should be
invariant to the length of the investment horizon. He showed that under certain conditions
(essentially that investors have a power utility function and that returns are serially
independent), the likelihood of loss and the potential magnitude of loss exactly offset each other
as the investment horizon increases.
In our analysis we assume that returns are serially independent, but we do not assume
power utility across the entire range of returns. We assume that investors have power utility for
returns that generate a coverage greater than 1, but that they have linear utility for returns that
produce coverage ratios below 1. Stated more prosaically, we assume that investors face a critical
threshold (a coverage ratio equal to 1) at which their utility changes abruptly. Thus our
framework is not at all inconsistent with Samuelson’s observation about the connection between
time and risk.

Electronic copy available at: https://ssrn.com/abstract=3303153

 10

5. Conclusion
We propose a new framework for evaluating retirement investment strategies. We begin
with the assumption that investors are primarily concerned with generating returns that are
sufficiently large and stable to sustain spending throughout retirement, and that they derive
additional satisfaction from leaving a bequest.
We then introduce a new metric, the coverage ratio, which measures the number of years
of withdrawals (during and after retirement) supported by a strategy, relative to the length of the
retirement period considered. Importantly, this metric accounts for everything the failure rate
accounts for. In addition, it accounts for when a strategy fails if it does, and by how much it
succeeds when it does, neither of which is taken into account by the failure rate.
Moreover, we propose a particular kinked utility function for evaluating the coverage
ratio. We assume that utility rises at a diminishing rate as the coverage ratio rises above 1; that
is, with increases in the potential bequest, and that it falls sharply and linearly as the coverage
ratio falls below 1, which corresponds to increasing failure to support spending. This particular
utility function is consistent with our initial assumption that investors are primarily concerned
with supporting spending and secondarily concerned with leaving a bequest.
We then applied our framework to determine the optimal allocation to stocks and bonds
based on historical returns across 21 countries as well as the world market over a 110-year
period. Our analysis revealed that in most countries investors should have implemented very
aggressive strategies, with an average allocation to stocks of 91%. We argued that this strong
preference for stocks was caused by the favorable equity risk premium that prevailed across
these markets as well as by the (30-year) length of the retirement period we considered.
We next applied our framework using simulated returns. Simulation allowed us to
experiment with different assumptions for the expected return and risk of stocks and with
different retirement horizons. Our simulations revealed that the optimal allocation to stocks
varied in accordance with conventional wisdom. All else equal, the optimal allocation to stocks
was higher for higher expected returns for stocks, lower volatility, and longer retirement periods.
We believe that our framework for determining the optimal retirement investment
strategy is conceptually appealing and empirically reasonable. We also believe that it is superior
to a simplistic framework based solely on the limited information provided by the failure rate.
We thus encourage investors and particularly advisors to take it into account when making
financial decisions for retirement.

Electronic copy available at: https://ssrn.com/abstract=3303153

 11

References

Dimson, Elroy, Paul Marsh, and Mike Staunton (2002). Triumph of the Optimists – 101 Years of
Investment Returns. Princeton University Press.

Dimson, Elroy, Paul Marsh, and Mike Staunton (2016). Credit Suisse Global Investment Returns
Sourcebook 2016. Credit Suisse Research Institute, Zurich.

Estrada, Javier (2017). “Refining the Failure Rate.” Journal of Retirement, 4, 3, 63-76.

Estrada, Javier (2018a). “From Failure to Success: Replacing the Failure Rate.” Journal of Wealth
Management, Journal of Wealth Management, 20, 4, 9-21.

Estrada, Javier (2018b). “Replacing the Failure Rate: A Downside Risk Perspective.” Journal of
Retirement, 5, 4, 46-56.

Samuelson, Paul (1963). “Risk and Uncertainty: A Fallacy of Large Numbers.” Scientia, 98, 4, 108-
113.

Electronic copy available at: https://ssrn.com/abstract=3303153