Dividend portfolio – multi-period performance of portfolio selection based

solely on dividend yields

(DRAFT)

Bogusz Jelinski
September 29th, 2020

Abstract: What will happen to your investment if you ignore change of share prices while
calculating both returns and risk during stocks portfolio selection? Is it profitable in long term to
sell a share when its price has started to skyrocket? Backtesting of a decade-long investment with
periodic portfolio rebalancing shows a continuous advantage of such radical and non-intuitive
approach over the classic formula of return-on-investment. The idea to concentrate on dividend
yield in portfolio selection theory, contrarian to popular models, is at least thirty years old, here it is
tested and some practical aspects are analyzed. With the help of a computer simulation it is possible
to show the impact of both different frequencies of portfolio rebalance, which was huge, and
different risk levels. All simulations have been logged to files so that it is possible to trace changes
in portfolios including dividends paid and splits without running the simulator, which is published
on GitHub as open-source with an instruction manual and can easily be run by anyone.

Keywords: dividend contrarian dynamic portfolio selection computer simulation backtesting

Key Takeaways:
- historical prices of stocks, their past performance might not be the best hint on how to invest
- a naive method to pick out the best stocks for a portfolio (opposed to nonlinear programming) can
provide very good results
- the naive method gives better results in low risk strategies, risk-return tradeoff has not been
observed,
- there is no evidence that portfolio selection based on dividend yield performs worse in bearish or
bullish periods. Yields were quite consistent throughout decades.

Introduction
Harry Markowitz’ work (1959) contributed vastly to the development of financial economics and
established underpinnings for Modern Portfolio Theory – a framework for the construction of
investment portfolios. There were many followers, new models and theories, both normative ones
describing how investors should behave and positive ones like CAPM, which proposes hypothesizes
how markets actually behave. MPT1 includes numerous implicit and explicit assumptions about
markets and investors, and criticisms emerged soon after the theory got its audience. For example
asset returns have shown to be far from the normal distribution, with fat tails and high kurtosis.
Even Markowitz pointed out limitations of his own work by warning us that when past
performances of securities are used as inputs, the outputs of the analysis are portfolios which
performed particularly well in the past.
One of criticisms refers to risk-return tradeoff. Haugen and Heins (1975) found that risk
might not generate a special reward. Murphy (1977) cited several studies, which concluded that
there might not be any stable relationship between return and risk. Fama and French stated after
having examined 9500 stocks across three decades that risk measured by beta was not able to
predict performance. It has to be noted that despite scientific evidence, much wider than mentioned
1 This acronym will be used later on to refer to the initial Markowitz’s model. Modern Portfolio Theory contains a
much wider spectrum of models and ideas.
1
above, one can still find books on investment which state without deeper reflection or any
reservation that «there is a risk-return trade-off» (like Bradford, 2009, p. 208). Fama and French
also found that shares with lower P/E ratios and price to book ratios provided the highest returns,
stocks are more bound to these measurements than to beta. Similar observations were made among
others by Basu (1977), Rosenberg (1985), Bhandari (1988), but you could find such claims already
in Graham and Dodd (1934). The alleged risk-return tradeoff can however be observed in some
results of simulations presented below.
A fundamental and still debated question is how risk ought to be measured. The original
variance approach began to be troublesome as markets were becoming more volatile. It has forced
scientists and financial institutions to invent other risk measures, like for example JP Morgan’s
Value-at-Risk or copula proposed by Sklar (1973). Although variance will be used here, techniques
used for simulations allow for further research on the impact of different risk definitions on
portfolio performance.
While estimating risks and returns using the historical performance of stocks it has to be
decided which time frame should be used to get the sample. It was seen as MPT’s drawback by
Fabozzi (2002) and the results of simulations presented here show a vast impact of this decision.
Estimation error attracted attention of many researchers, a good overview was given by DeMiguel
(2009). DeMiguel’s work is worth mentioning not only because it is similar to what is proposed in
this article – the 1/N, naive strategy is used as a benchmark for other approaches. It shows that the
estimation window needed for the sample-based mean-variance strategy and its extensions to
outperform the 1/N benchmark is around 3000 months for a portfolio with 25 assets and about
6000 months for a portfolio with 50 assets. Estimation errors were claimed to be the cause of poor
performance of sample-based portfolios. The results shown below contradict such claims in long
term – MPT portfolios performed much better than S&P index. Another contradiction found is that
there have been claims that MPT portfolios tend to have large quantities of individual assets,
Affleck-Graves (1996) found that weights are limited to about 40 percent. In results presented
below many risky portfolios produced by model solver did contain just one asset during
backtesting.
Concentration on dividends as the main source of wealth and on other «value» strategies
based on different intrinsic properties was one of the mainstream ideas at least since Fama and
French published their studies on the power of dividends yield to explain stock returns. Their
findings suggested that it increases with return horizon. Two decades later Bekaert (2006) wrote that
the ability of the dividend yield to predict excess returns is best visible at short horizons. Another
contribution was provided by Lakonishok (1994) who postulated that over longer horizons, value
strategies have outperformed glamour2 strategies quite consistently. He also considered possible
explanations for this extraordinary performance – value strategies were contrarian3 to naive
strategies based on past earnings, investors get overly excited by past and are irrespectful to mean
reversion (Kahneman, 1982, p. 417). Lakonishok’s evidence does not support the hypothesis that
value strategies are fundamentally riskier. Let us mention a few more researchers interested in value
portfolios. A study done by Keppler (1991), which examined the link between price-to-cash flow
and investment returns found that the most profitable strategy was investment in the lowest price-to-
cash flow quartile. Arnott (2003) examined different components of equity returns in 200 years. His
conclusion was that dividends were responsible for most of returns «dwarfing» rise of prices.
Interestingly some studies independently found that the ninth decile outperforms the highest
yielding one.
2 fast growing
3 Because this term is used not only for portfolios controlled by dividend yields «dividend portfolio» will be used
instead
2
 MPT is a single-period, static theory on the choice of portfolio, although Markowitz also
tried to extend his optimization criterion with the expected utility of wealth after many reinvestment
periods and used SIMSCRIPT for simulations (Rubinstein, 2002). This approach of a repeated game
between the nature and the investor, sometimes called «dynamic hedging» or backtesting, is used
during simulations here4.

Two models will be compared – the classic MPT and a model based solely on dividend yields,
ignoring changes in asset valuations and even ignoring optimization – weights will be split evenly.
Portfolios will be continuously adjusted (rebalanced) by buying and selling shares in response to
variations of prices and dividends. The terminal wealth will be compared.

To sum up what needs to be addressed:

- does ignoring changes in stock valuations bring extra benefit over MPT?
- is this benefit consistent trough-out decades?
- does taking risk reward an investor, is there any risk-profit tradeoff?
- what is the impact of different sampling methods, different period lengths?
- what is the structure of MPT portfolios, does a weight exceed 50 percent and when?
- what values of strategy attributes (preferred risk, rebalance frequency) give best results?
- does the number of assets in dividend portfolio make a difference?

Classic mean-variance model

In order to help recognize the model used for simulations in the source code and help understand
the flow of the simulation let us sum up some widely known formulas for sample-based
Markowitz’s model. In order to invest in stocks one would choose a period of time in recent history,
divide this period into smaller parts and calculate returns on investment of a chosen set of stocks:

ROI = F i n al P r i c e− I ni t i al P r i c e+ D i v i d e nd s P a y o ut s−T r a n s a c t i o nC o s t s
I ni t i al P r i c e

Then an optimal portfolio would be found, which would require making some assumptions – is it
the expected return that should be maximized with an assumed, maximum level of accepted risk or
the risk that should be minimized with an assumed, minimum level of return? Having estimated the
mean values and the covariance matrix we can construct the classic model proposed by Markowitz:
N
E=∑ x i μ^i
i=1
N N
V =∑ ∑ ^
σ i j xi x j
i=1 j=1
N

∑ xi=1
i=1
x i≥0 f o r i=1 . . N

Depending on investors preferences (emin or vmax) you can have two variants of the model to solve:

V→M I N
E≥em i n
4 The simulator with a short manual is available on GitHub: https://github.com/boguszjelinski/stocks
3
or
E→ M A X
V ≤v m a x

The latter will be used in simulations. No short sale, dividends will be reinvested. As transaction
costs are negligible in a large scale, they will not be calculated.

Dividend portfolio
The dividend portfolio will be constructed differently. The ROI will ignore the change of valuation:

ROI = D i v i d e n d s P a y ou t s−T r an s ac t i o nC o s t s
I ni t i al P ri c e

The variance will be estimated for each stock, stocks exceeding vmax will be ignored, the rest will be
sorted by mean dividend yield. The best scoring ’n’ stocks will be picked up and evenly split in the
portfolio.

Simulation results
Let us assume we have 100000 USD, investment horizon is 16 years starting Jan 1st 2000. This
spans over two periods of both decline and growth, and gives a good insight in performance in
different market conditions – see Figure 2. We would like to invest in S&P 100 stocks. We can
rebalance portfolio every month or once a year, or somewhere in between. It means we have 12*16
rebalances when one month has been chosen, or 16 rebalances with period length of one year. We
can have different risk preferences. While gathering the sample (for assessing risk and return) let us
look into twelve preceding periods of the same length as the chosen rebalance period – a rolling
window sample. It sounds quite natural – someone interested in long-term investments would not
perceive a short-term fluctuations as valuable information. Sampling approaches are not in scope of
this work although impact of different sampling periods have been analyzed – see tables 2 and 3.
We ignore the possibility of estimation error.
With MPT we maximize expected return, we use a SCS solver to find the optimal portfolio5. Figure
1 shows the achieved terminal wealth depending on how often portfolios are rebuilt (in months) and
how high the acceptable risk is.

Figure 1. Terminal wealth of 42 strategies achieved with MPT and dividend portfolio.

5 https://github.com/cvxgrp/scs/
4
Dividend portfolios contained four best scoring assets. Simulations with two and ten assets showed
that four is a better choice, but differences were small.
It is clearly visible that these two portfolios react differently to risk. In MPT one could see the risk-
return tradeoff while in “dividend portfolio” the lowest risk prevails. The average wealth of 42
strategies is significantly higher in “dividend portfolio” (372884 vs 455621). Both MPT and
dividend portfolios outperformed S&P100 in 16 year period – see Figure 2. A more detailed
analysis of investment «trajectories» shows that the dividend portfolios coped with the 2000-2002
decline, but were not resistant to that in 2008 and 2009. Similar conclusion could be drawn for the
1990-2000 decade (not shown in the chart).

Figure. 2 MPT and dividend portfolios vs S&P100 – terminal wealth in thousands USD
(from Jan 1st 2000 till Jan 1st 2016, rebalance after 3 months)

Table 1 shows how the maximum risk in the MPT model affects portfolio structure – it has not been
observed that portfolios contain large quantities of individual assets, portfolios with just one asset
were often proposed by the solver for more risky portfolios (vide Affleck-Graves).
Table 2 and Table 3 show how differences in sampling (different lengths of period's
window) affects one chosen strategy (0.005/12) in MPT and dividend portfolios respectively. As
you can see the impact on the terminal wealth may be vast, but table is not meant to provide any
hint for investors due to probably significant estimation errors.

5
 [%] Number of different tickers in portfolios
V≤ 1 2 3 4 5 6 7 8 9
0.005 3.3 10.0 22.5 35.0 18.3 5.8 1.7 2.5 0.8
0.015 17.5 30.8 30.8 12.5 5.0 0.0 2.5 0.8 0.0
0.075 75.8 14.2 6.7 2.5 0.8 0.0 0.0 0.0 0.0
0.375 92.5 2.5 2.5 1.7 0.8 0.0 0.0 0.0 0.0

Table 1. Structure of MPT portfolios depending on maximum risk – frequencies of particular sizes
in 120 rebalances, as a percentage (start on Jan 1st 2000, rebalance after 1 month)

Number of periods in sample

6 12 18 24
3 358 330 364 463
Length of
period in
months

6 453 207 258 315

9 267 241 429 324
12 340 270 303 286

Table 2. Impact of different sampling methods on terminal wealth in MPT

(in thousands; start on Jan 1st 2000, V≤0.005, rebalance after 12 months, 16 rebalances)

Number of periods in sample

6 12 18 24
3 258 468 428 243
Length of
period in
months

6 420 251 317 657

9 339 459 639 725
12 358 649 766 438

Table 3. Impact of different sampling methods on terminal wealth, «dividend portfolio»

(in thousands; start on Jan 1st 2000, V≤0.005, rebalance after 12 months, 16 rebalances)

Figure 3 is an example of documentation generated by the simulator in order to examine its flow
and correctness of historical data6. It shows a dividend portfolio purchased Jul 1st 2010 and its value
three months later. A 1000000 / 937989 split was considered on VZ (Jul 2nd 2010).

6 Data downloaded from Yahoo had some missing or wrong data in dividends and splits. Output generated by Python
simulator differs.
6
 Figure 3. Example of the simulator log

An obvious question would be if the dividend advantage is consistent throughout decades.

Figures 4 and 5 show the realized yearly dividends from the «dividend portfolio» between Jan 1st
1980 and the end of 2015, for V≤0.005 and V≤0.375 scenarios respectively (arithmetic mean of
four assets). Values tend to fluctuate around 5%, with some extraordinary positive deviations
(mean=0.05, variance=0.0002 and mean=0.06 and variance=0.001 respectively). It looks like a
trustworthy foundation for the dividend portfolio.

Figure 4. Dividend yields for V≤0.005 realized in 36 years - «dividend portfolio» strategy
(yearly values, start on Jan 1st 1980, rebalance after 3 months)

7
 Figure 5. Dividend yields for V≤0.375 realized in 36 years - «dividend portfolio» strategy
(yearly values, start on Jan 1st 1980, rebalance after 3 months)

Computational considerations and reproducibility

There are essential differences in datasets distributed freely in the Internet and these differences
have had vital impact on terminal wealth. As an example – while comparing Yahoo and Tiingo
datasets one third of paid dividends are different by more than 10 cents, 15% of splits are different
by more than 0.1 (10%), there are differences in stock prices too. Differences are smaller when
comparing Tiingo to Quandl, but still not negligible. The same effect was observed while running
simulations in different languages (Python vs Java) – small differences in implementations have
affected the terminal wealth quite significantly. All simulations are available in GitHub and these
differences can be reviewed.

Conclusions
The dividend portfolio even in a primitive form (no solver) outperforms the classic portfolio theory
quite consistently in low risk scenarios and it is generally a viable alternative to MPT. There is no
evidence that the dividend portfolio performs worse in bearish or bullish periods. The contrary –
achieved dividend yields were steady throughout decades and taking less risk did not punish the
investor. The risk-return tradeoff is visible in results for MPT but the structure of risky MPT
portfolios tends to contain few assets making such strategies unacceptable to most investors.
There are several topics that might need further research or improvements – reducing
estimation errors, other risk measures and techniques like dividend arbitrage. Does semi-variance
give better results? Does the ninth decile perform better than the last one? Why was there a
difference in performance in two decline periods which were under scrutiny? Maybe the optimal
parameters of strategies differ in bearish and bullish periods. This was not evident either which
investment horizon (short or long) gives higher results. The simulator, which is available on GitHub
as open-source, may also be used to continue research on MPT – a few contradictions to other
research were found. Most importantly - sharing the source code on GitHub allows for cooperative
research and easy reproducibility and verification of results.

8
References

Affleck-Graves, John F. and Arthur H. Money. "A comparison of two portfolio selection models."
The Investment Analysts Journal 7, no. 4 (1976), 35-40.

Arnott, Robert D. (Editor). "Dividends and The Three Dwarfs" Financial Analysts Journal 59, no. 2
(2003), 4-6.

Basu, Susanto. "Investment performance of common stocks in relation to their price-earnings

ratios: A test of the efficient market hypothesis." Journal of Finance 32, no. 3 (1977), 663–682.

Bekaert, Geert & Andrew Ang. "Stock Return Predictability: Is it There?" The Review of
Financial Studies 20, no. 3 (2006), 651-707.

Bhandari, Laxmi C. "Debt/equity ratio and expected common stock returns: Empirical evidence."
Journal of Finance 43, no. 2 (1988), 507–528.

Bradford, Jordan & Thomas Miller. Fundamentals of investments (5th ed.). New York: McGraw-
Hill, 2009.

DeMiguel, Victor, Lorenzo Garlappi & Raman Uppal. "Optimal versus Naive Diversification: How
Inefficient Is the 1/N Portfolio Strategy?" Review of Financial Studies 22, no. 5 (2009), 1915-1953.

Dickson, Henry C. & Charles L. Reinhard. "Dividend Yield Screens" Portfolio Strategy, Lehman
Brothers Equity Research, September 26, 2005.

Fabozzi, Frank, Francis Gupta & Harry M. Markowitz, H. "The legacy of modern portfolio theory"
Journal of Investing, 11, no. 3 (2002, Fall), 7-22.

Fama, Eugene F. & Kenneth R. French. "The Cross-Section of Expected Stock Returns" Journal of
Finance 67, no. 2 (1992), 427-465.

Fama, Eugene F. "Market efficiency, long-term returns, and behavioral finance" Journal of
Financial Economics 49 (1998), 283-306.

Graham, Benjamin & David Dodd. Security Analysis. New York: McGraw-Hill, 1934.

Haugen, Robert A. & A. James Heins. "Risk and the Rate of Return on Financial Assets: Some Old
Wine in New Bottles" The Journal of Financial and Quantitative Analysis 10, no. 5 (1975, Dec.),
775-784.

Kahneman, Daniel, Paul Slovic & Amos Tversky. Judgment Under Uncertainty: Heuristics and
Biases. New York: Cambridge University Press, 1982.

Keppler, Michael A. "The Importance of Dividend Yields in Country Selection" Journal of

Portfolio Management 17, no. 2 (1991, Winter).

Lakonishok, Josef, Andrei Shleifer & Robert W. Vishny. "Contrarian Investment, Extrapolation, and
Risk" The Journal of Finance 49, no. 5 (1994, Dec.).

Markowitz, Harry. Portfolio Selection. New York: John Wiley & Sons, 1959.

Murphy, J. Michael. "Efficient Markets, Index Funds, Illusion, and Reality" Journal of Portfolio
Management 4, no. 1 (1977, Fall), 5-20.
9
Rosenberg, Barr, Kenneth Reid & Ronald Lanstein. "Persuasive evidence of market inefficiency"
Journal of Portfolio Management 11, no. 3 (1985), 9–16.

Rubinstein, Mark. "Markowitz’s Portfolio Selection: A Fifty-Year Retrospective" The Journal of

Finance 57, no. 3 (2002, June).

Sklar, Abe. "Random variables, joint distribution functions, and copulas" Kybernetika 9, no. 6
(1973), 449–460.

10