line of
SIGHT
The Equity Imperative
CHOOSING A
‘SMART BETA’ FACTOR
NOT WHICH, BUT WHEN
Investors interested in reaping the benefits of factor-based investing in their portfolios have
July 2015
long believed the ultimate question to be, “Which factor should I choose?” Our most recent
research shows, however, that investors should instead ask, “When should I favor each factor?”
While equity factor tilts have historically outperformed cap-weighted alternatives, there
is less evidence to suggest any one factor will consistently outperform another. Factors
tend to follow unique cycles, outperforming cap-weighted indices during some periods
and underperforming during others. Our research has shown that taking advantage of
time diversification – varying factor exposures across the investment horizon – can help
address the problem of factor cycles and improve portfolio outcomes.
To receive our latest equity solutions research papers when they are published,
register at northerntrust.com/imperative.
northerntrust.com | Line of Sight: Choosing a ‘Smart Beta’ Factor | 1 of 12
BUILDING ON PRIOR RESEARCH
Through our ongoing research into Engineered Equity, we continue to explore ways
to apply factor-based investing to enhance investment results. Previously, in our papers,
Understanding Factor Tilts, What is Quality, and Global Low Volatility Anomaly:
Benefiting From an Actively Designed Approach, we examined the empirical evidence
and theoretical rationale behind factor-based equity investment strategies. We concluded
that tilts toward compensated factors such as size, value, low volatility, dividend yield
and quality have the potential to generate higher risk-adjusted returns than capitalizationweighted benchmarks. In Combining Risk Factors for Superior Returns, we showed
that an inherent dilution effect among factors makes the natural temptation to combine all
factors in a single strategy faulty. With this research, we are exploring the importance of
timing as it relates to the investor’s objectives in choosing a factor.
FINDING THE FACTOR CYCLE LENGTHS
We found these factor cycle lengths by analyzing factor cycles in a novel way. This
showed us that the length of factor cycles, i.e., the period of relative outperformance and
underperformance, is not consistent across factors. Although the length of factor cycles is
not consistent, some have consistently longer cycles than others. Capitalizing on these cycle
differences and aligning them to your portfolio’s investment horizon allows you to reap the
benefits of time diversification and compensate for the cyclical nature of factor-based returns.
Next, using a unique multi-period portfolio optimization algorithm, we determined
which factors had more favorable results at different points in the investment holding period:
Factors with longer cycles, such as size and value, are preferred for portfolios with long
investment horizons while other shorter-length factors such as low volatility are preferred in
portfolios with shorter investment horizon. The intuitive rule of thumb that emerges from
these findings is that your factor cycle length should be less than your holding period.
If you consider time diversification in your portfolio construction, any compensated
factor can have a place in the portfolio; the choice is governed by the portfolio’s investment
horizon. In other words, a portfolio with a 30-year investment horizon will benefit from a
different set of factor exposures than one with a three-year holding period. As the portfolio
matures and approaches liquidation, the optimal set of factor exposures will change
(see Exhibit 1).
EXHIBIT 1: OPTIMAL FACTOR EXPOSURES ACROSS THE PORTFOLIO TIME HORIZON
Time to Liquidation
Optimal Factor Exposures
Examples
More than 8 years
Long cycle factors
Size, value, momentum
4 to 8 years
Intermediate cycle factors
Dividend yield
Less than 4 years
Short cycle factors
Low volatility
Source: Northern Trust Quantitative Research
2 of 12 | Line of Sight: Choosing a ‘Smart Beta’ Factor | northerntrust.com
UNDERSTANDING FACTOR CYCLES
One drawback of factor-based investing is that factors can underperform for long periods.
To illustrate, Exhibit 2 shows cumulative return for the value factor within the Russell 3000
universe between January 1979 and June 2014.1 We calculate these value-factor returns using
the MSCI Barra definition of value. Each month we rank all stocks in the universe based on
this value definition and put them into equally weighted quintiles. We then calculate the
subsequent return for each quintile, defining the resultant value-factor return as the first
quintile (highest value) return minus the benchmark return. Doing this provides us the
value-factor return net of market beta.
EXHIBIT 2: VALUE FACTOR CUMULATIVE RETURNS
For first-quintile returns of the Russell 3000
3.00
2.50
2.00
1.50
2013
2011
2009
2007
2005
2003
2001
1999
1997
1995
1993
1991
1989
1987
1985
1983
1981
.50
1979
1.00
Note: Dollar values are calculated as the value of $1 invested in the top quintile of value less the Russell 3000
performance. Shaded bars represent periods of underperformance.
Source: Northern Trust Quantitative Research
As you can see, the cumulative return line slopes upward suggesting that over time the
value factor does indeed have a positive return. However, value also underperformed
during several protracted periods. For example, the cumulative value-factor return peaks
in February 1989, and then begins to ebb. It is not until October 1993 that the value factor
recovers to this 1989 peak level. In other words, $1 exposed to the value factor in February
1989 would have been worth less than $1 for 56 months. Similarly, during the 34-month
period between May 1998 and March 2001, and the 32-month period between December
2006 and August 2009, value experienced severe cyclical downturns. These downturns are
shaded in Exhibit 2, but they are not the only periods in which value underperformed.
Based on this graph, we could find the average durations of the value cycle and calculate
the cycle length to be around 41 months2. Of course, identifying cycle peaks and troughs
visually produces a very crude estimate of cycle length. Using a statistical process known
as spectral analysis produces more accurate cycle-length estimates, as shown in Exhibit 3.
(We describe the spectral analysis process in Appendix A.) Note that we computed all
factor returns using the same universe, period and process as we used for the value factor.
1 We chose the Russell 3000 because it has a long history, and calculated the value-factor returns using the MSCI
Barra definition of value. We found similar results for non-U.S. developed and emerging market indices.
2 (56+32+34)/3 = 40.67 months.
northerntrust.com | Line of Sight: Choosing a ‘Smart Beta’ Factor | 3 of 12
EXHIBIT 3: VALUE FACTOR CUMULATIVE RETURNS
The spectral analysis results produced a value-factor
cycle length of 47 months – somewhat longer than our
crude estimate of 41 months. As you can see in Exhibit 3,
cycle lengths vary considerably across factors. The size factor
cycle extends for more than 100 months while the low volatility
factor cycle spans just 12 months. These cycle lengths make
sense in light of our previous research. For example, in our
paper Understanding Factor Tilts we note the size factor is
subject to relatively long periods of sustained outperformance
and underperformance while our paper Global Low Volatility
Anomaly shows the low volatility factor cycle is defined by
short, but potentially severe, “junk rallies.”
Cycle lengths as determined using spectral analysis
Factor
Major Cycle Length
Low volatility
12 months
High dividend yield
22 months
High momentum
39 months
High value
47 months
Small size
106 months
Source: Northern Trust Quantitative Research
USING FACTOR CYCLE LENGTHS IN PORTFOLIO CONSTRUCTION
Applying this knowledge of factor cycle lengths to portfolio construction, it seems to
make sense that investors would want to focus on factors with cycle lengths less than their
investment horizon. Tilting toward factors with a longer cycle length exposes the portfolio
to the risk of gaining exposure at or near a peak in a factor cycle and being forced to
liquidate the portfolio before the cycle completes and the factor can fully recover.
To test this, we developed a timeline of factor allocations. We did this using a unique
multi-period portfolio optimization algorithm (see Appendix B for details) that seeks the
factor allocation glide path that maximizes an investor’s utility across the entire investment
horizon. This utility encompasses both risk and returns, and reflects an investor’s well-being
leading to and at portfolio liquidation. The multi-period structure allows the model to
accurately reflect time diversification, a feature that is absent in traditional mean-variance
optimization modeling.
Using this tool, we tested whether the optimal factor allocation changes along the
glidepath. Specifically, does the optimal choice of factor or factors indeed depend on
holding period? If so, should investors favor certain factors further from liquidation and
others closer to liquidation? At every year in the investment horizon, the model determines
the optimal holdings of the MSCI Size, Value, Momentum, Dividend Yield and Low
Volatility equity indices, our proxies for this calculation. The results of this optimization
appear in Exhibit 4.
EXHIBIT 4: OPTIMAL FACTOR ALLOCATION
100%
Size
90%
Factor Allocation
80%
Value
70%
60%
Momentum
50%
40%
Dividend Yield
30%
20%
Low Volatility
10%
0%
40
38
36
34
32
30
28
26
Source: Northern Trust Quantitative Research
4 of 12 | Line of Sight: Choosing a ‘Smart Beta’ Factor | northerntrust.com
24
22
20
18
Time to Liquidation (Years)
16
14
12
10
8
6
4
2
We can draw several interesting observations from Exhibit 4. First, there appears to be a
natural ordering or cascade of factor exposures over the investment horizon. Long cycle
factors such as size, value and momentum appear early in the allocation and are gradually
reduced as the portfolio matures. Approximately eight years from liquidation, the portfolio
ceases to have any long-cycle factor exposures. With an intermediate cycle length, dividend
yield first appears about 30 years from liquidation and ends about four years from
liquidation, at which time the short-cycle low volatility factor becomes the sole allocation.
Second, factors seem to leave the portfolio at roughly twice their cycle length. While
we measure the dividend-yield cycle at 22 months, it leaves the portfolio about 48 months
prior to liquidation. Likewise, value has a measured cycle length of 47 months but leaves
the portfolio about 96 months prior to liquidation. We therefore modify our rule of thumb
to account for some margin of error: your factor cycle length should be less than half of
your holding period.3 After all, our measured cycle lengths are only averages; actual
cycle lengths will fall above or below this average.
The model suggests value is the dominant factor over much of the investment horizon.
In this optimization, the allocation magnitudes are driven by assumptions about factor
return expectations (i.e., historical performance) and the investor’s risk preferences, but the
ordering is largely independent of these assumptions. In other words, the total allocation
to any factor will change as return and preference assumptions change, but the ordering
will tend to stay the same.
OPTIMAL FACTOR CHOICE DEPENDS ON INVESTMENT HORIZON
As our new reasearch has shown, investors interested in or currently using factor-based
investing in their portfolios may need to take another dimension into account in their
investment decision-making process. Rather than focusing just on which factors, when
is an important component of the decision as well.
All factors can have a place in an investment portfolio. Our research shows the choice
of which to favor hinges on investment horizon. Investors with short horizons should avoid
long-cycle factors that may underperform for extended periods, instead opting for shortercycle factors like dividend yield and low volatility. Longer horizon investors are better
positioned to hold longer-cycle factors like value, size and momentum until their portfolio
liquidation approaches. This allows investors to take advantage of time diversification
and even out the effects of factor cycles on returns.
LEARN MORE
If you would like to learn more about adjusting your factor-based investment strategy to
take factor cycle lengths into account, and how this may benefit your portfolio, contact your
Northern Trust representative, or visit us at northerntrust.com/engineeredequity.
3 Conversely, the holding period should be double the length of the factor cycle.
northerntrust.com | Line of Sight: Choosing a ‘Smart Beta’ Factor | 5 of 12
A P P E N D I X A : S P E C T R A L A N A LYS I S
Spectral analysis4 is a method of decomposing a time series, in our case factor returns,
into periodic components or cycles. Most time series are comprised of multiple cycles
with lengths varying from very long (super-cycles) to very short (noise). Our objective is
to analyze factor return series and determine the length of the most significant long, i.e.,
low-frequency, cycle. This low-frequency cycle reflects the period over which a factor
may experience sustained underperformance. As an example, consider the time series
shown in Exhibit 5. Here we graph the daily value of the time series yt over a full year.
EXHIBIT 5: SAMPLE TIME SERIES
1.5
1
353
321
337
305
273
289
257
241
225
209
193
177
161
129
145
97
113
81
65
49
33
17
0
1
y(t)
0.5
-0.5
-1
Day
-1.5
Source: Northern Trust Quantitative Research
Spectral analysis reveals this time series can be decomposed into two major periodic
components. The first cycle is relatively low frequency with a cycle length of about six
months. The second is relatively high frequency with a cycle length of about one month.
Using spectral analysis we find the following equation for this time series:
yt= 0.8sin(2t)+0.2 cos(12t)+εt [A1]
Where =2 /N with t representing a frequency level and t is an error term that is
interpreted as very high frequency noise.
Exhibit 5 overlays equation A1 on the time series and you can see that the equation can
explain more than 97% of the total volatility of the time series.
4 Spectral analysis is a very well defined mathematical procedure. Readers interested in learning more should
consult Greene (2003), Hamilton (1994) or Warner (1998).
6 of 12 | Line of Sight: Choosing a ‘Smart Beta’ Factor | northerntrust.com
EXHIBIT 6: SAMPLE TIME SERIES WITH EQUATION A1 OVERLAY
1.5
1
353
337
321
305
289
273
257
241
225
209
193
161
177
129
145
113
97
81
65
49
17
33
0
1
y(t)
0.5
-0.5
-1
Day
-1.5
Source: Northern Trust Quantitative Research
The cycle lengths are obtained from the coefficients of the frequency terms t. The first
coefficient is 2, which suggests, in months, the cycle length is six months (12/2). The second
coefficient is 12, indicating a one-month cycle length (12/12). If this was factor return data
we would note the lowest frequency cycle of importance was six months, indicating that
our factor could sustain periods of underperformance for periods of up to six months.
GENERATING THE SMOOTHED SPECTRUM
The smoothed spectrums for each of the value, momentum, volatility, dividend yield
and size factors are shown in Exhibit 75. The materiality threshold is shown as a dotted
line. Red lines indicate the lowest frequency spectral peak judged to material using the
following criteria:
1. We prepared the factor return series using the method described in the paper.
2. To avoid exaggerating the materiality of longer cycles, we “detrended” the factor series
(see Warner [1998])
3. We used a Fast Fourier Transform, as described by Bloomfield (2013), to generate a
periodogram (a decomposition of the percentage of variance of the time series that
is accounted for by sinusoids of different frequencies). In general, computing a
periodogram is done by solving the equation:
π

π

yt= µ+ α(ω)cos(ωt)dω+ δ(ω)sin(ωt)dω
[A2]
00
4. We then smoothed the periodograms were smoothed using a Blackman-Tukey window
as described in Blackman and Tukey (1958) to form spectrums.
5. We identified the lowest frequency spectral peak that explained more than 3% of total
variance as the defining factor cycle. Note that since the spectrum is smoothed, the actual
variance explained in the neighborhood of this frequency is typically well over 10%.
5 Spectrums were rescaled from their traditional format to show percentage of total variation. They were also
truncated at 10 months to eliminate spurious high frequency peaks.
northerntrust.com | Line of Sight: Choosing a ‘Smart Beta’ Factor | 7 of 12
EXHIBIT 7: SMOOTH FACTOR SPECTRUMS
High Momentum
7%
6%
6%
5%
5%
Months
7%
6%
5%
4%
3%
2%
1%
10.4
10.9
11.5
12.2
12.9
13.7
14.7
15.8
17.0
18.5
20.3
22.4
25.1
28.4
32.8
38.7
47.3
60.9
0%
Months
With a cycle length centering on 106 months we can see small size has the longest cycle
length of all factors. High value and high momentum also have relatively long cycle
lengths of around 40 to 50 months. At just 12 months, the shortest cycle is low volatility;
dividend yield had a 22-month cycle.
8 of 12 | Line of Sight: Choosing a ‘Smart Beta’ Factor | northerntrust.com
11.5
10.4
10.4
12.2
10.9
12.9
12.2
10.9
13.7
12.9
11.5
14.7
18.5
20.3
25.1
28.4
32.8
22.4
18.5
20.3
22.4
25.1
28.4
32.8
38.7
47.3
10.4
10.9
11.5
12.2
12.9
13.7
14.7
15.8
17.0
18.5
20.3
22.4
25.1
28.4
32.8
0%
38.7
0%
47.3
1%
60.9
1%
60.9
2%
85.2
2%
3%
142.0
3%
4%
426.0
4%
85.2
13.7
5%
85.2
14.7
5%
142.0
17.0
6%
142.0
15.8
6%
426.0
15.8
7%
% of Total Variation
7%
426.0
Months
High Dividend Yield
Small Size
% of Total Variation
17.0
Months
47.3
426.0
10.4
11.5
10.9
12.2
12.9
13.7
14.7
17.0
15.8
18.5
20.3
25.1
22.4
28.4
32.8
47.3
38.7
0%
60.9
0%
85.2
1%
142.0
1%
Low Volatility
% of Total Variation
2%
38.7
2%
3%
60.9
3%
4%
85.2
4%
142.0
% of Total Variation
7%
426.0
% of Total Variation
High Value
Months
APPENDIX B: MULTI-PERIOD
FACTOR OPTIMIZATION
Portfolio optimization tools can be useful in determining appropriate allocations to
asset classes or risk factors. However, traditional optimization techniques such as the
mean-variance (MV) selection of Markowitz (1952) and Tobin (1958) can carry very strong
assumptions about the underlying asset return generating process. For example, using
historical returns to construct covariance matrices and expected return vectors assumes,
perhaps unknowingly, that returns follow a Wiener process with drift so that the expected
return of any asset is independent of past returns6.
This assumption of independence, of course, negates the findings of Appendix A7. If
factors follow discernable cycles, representing them as a Wiener process is flatly incorrect,
especially if the horizon of the static MV model is less than the factor cycle length, which
seems typical. To overcome this difficulty, we devised a multi-period stochastic optimization
model that correctly accounts for factor cycles in the choice of portfolio allocations. This
utility-based approach models both factor autocorrelations and cross-correlations, and is
capable of handling a variety of investment horizons. The basic algorithm is very
straightforward and proceeds in four steps:
1.Simulate I sets of factor return paths Ri(N assets × T time periods) using information
about factor autocorrelations and cross-correlations8. This could be done with a simulated
VAR but we simply applied a Cholesky decomposition of the factor correlation matrix
to a series of AR models for each factor.
2. Use a numerical optimizer to find a matrix of portfolio weights (N assets × T time
periods) that maximizes the expected utility as defined by some utility function. At
each iteration, the candidate weights are run through all R sets of paths to compute
the utility of each path and find the expected value.
3. The optimizer will terminate according to the defined constraint and objective function
tolerances and/or the limit on maximum iterations and function evaluations. If the
optimization is successful, the resultant weight matrix W is stored. Repeat Steps 1 to 3
a total of M times.
4.After M repititions, average all weight matrices to find the final optimal allocation w*.
This reduces estimation error and optimizer “noise” as a resampling/shrinkage technique
in the spirit of Michaud (1998).
6 More specifically, return series have independent increments such that rt–rS~N(d(t–s),t–s) for 0≤s≤t where d is the
drift and t and s are measured in units corresponding to the drift.
7 Interestingly, Wiener (1923) himself gives a spectral representation of a Wiener process as a sine series whose
coefficients are independent N(0,1) obtained via the Karhunen-Loève theorem. In this sense, all cycle lengths of a
Wiener process are random with zero expected value.
8 Autocorrelations are the time domain corollary to factor spectrums.
northerntrust.com | Line of Sight: Choosing a ‘Smart Beta’ Factor | 9 of 12
In Step 1, we used monthly historical data for the MSCI World Value, Minimum Volatility,
Momentum, Dividend Yield and Small Cap indices from January 2001 through June 2014.
We employed 24 models (AR) and calculated the Cholesky decomposition from a correlation
matrix using the full data history. In our case I = 5000 sets of return paths, N= 5 factors
and T = 40 years/periods. Expected return or drift was simply the historical average
return for the index. Volatilities were historical volatilities. Step 2 was accomplished
using the following program:
Maximize:E(θ)
Where:
[B1]
θi=U(geo(wTRi))[B2]
{
11-y
U(x)=
—x
1-yi
∀y≠1
log x
otherwise
[B3]
geo(x)=geometric return of series x=(∏i(1+xi))1/n–1[B4]
Subject to:
wn,twn,t–1≤δ∀n=1, ... ,N
∀t=2, ... ,T
[B5]
wn,t–1wn,t≤δ∀n=1, ... ,N
∀t=2, ... ,T
[B6]
0≤w≤1
[B7]
Equation B3 is a standard constant relative risk-aversion utility function, frequently
employed in these types of models (see Pfau 2009). Azar (2006) suggests the risk aversion
parameter y realistically ranges from 1 to 5 so in our case we set y=5. Of course, final
results are sensitive to y, but in repeating the optimization with different values of y (not
shown) we obtained the same time-preferential ordering of factors in the portfolio but at
somewhat different magnitudes and times. For example, at a higher level of risk aversion,
the portfolio would tend to hold larger allocations of the low-volatility factor further
from liquidation.
Without the constraints of equations B5 and B6, the program would be computationally
difficult to solve and very time consuming. At least conceptually, this is because the solution
space for the problem is very large at 5×40=200 decision variables and 5,000 sets of return
paths to evaluate for each iteration. Using the interior point algorithm of Matlab’s fmincon()
optimization function took more than 48 hours to solve to tolerance on a quad core AMD
A8-5500B APU (not parallelized). Adding constraints B5 and B6 with =3% essentially
‘tied down’ the problem, considerably reducing the solution set and dropping the solve time
to less than two hours. Constraints B5 and B6 are also in the spirit of a true “glidepath” in
the sense that they stabilize the rate of change of factor allocations through time, which
makes intuitive sense.
10 of 12 | Line of Sight: Choosing a ‘Smart Beta’ Factor | northerntrust.com
The tolerances and parameters of the problem were as described in Exhibit B1:
EXHIBIT 8: TOLERANCES AND PARAMETERS
Parameter
Value
Maximum iterations
10,000
Function tolerance
0.0001
Maximum function evals
10,000
X tolerance
0.0001
Constraint tolerance
0.0001
Optimization method
Interior point
Initial trust region radius
Sqrt(# variables)
Subproblem method
LDL factorization
Finally, Steps 1 to 3 were repeated 100 times as a method of eliminating (some)
estimation error and reducing optimizer jitter. Using this approach we obtained much
smoother factor glidepaths, but at the expense of considerable computation time.
REFERENCES
Appendix A
1. Blackman, R.B., and Tukey, J.W., The measurement of power spectra from the point of view of
communication engineering. Dover Publications, 1958.
2. Bloomfield, Peter, Fourier Analysis of Time Series: An Introduction, Second Edition, John Wiley
and Sons, New York, 2013.
3. Greene, William H., Econometric Analysis, Fifth Edition, Prentice Hall, New Jersey, 2003.
4. Hamilton, James D., Time Series Analysis, Princeton University Press, New Jersey, 1994.
5. Warner, Rebecca M., Spectral Analysis of Time-Series Data, The Guilford Press, New York, 1998.
Appendix B
1. Azar, S. A., Measuring Relative Risk Aversion, Applied Financial Economics Letters, 2, 341-45, 2006.
2. Markowitz, M., Portfolio Selection, The Journal of Finance, Vol. 7, No. 1. pp. 77-91, 1952.
3. Michaud, R., Efficient Asset Management: A Practical Guide to Stock Portfolio Optimization and
Asset Allocation. Boston: Harvard Business School Press. 2nd ed. Oxford: Oxford University Press,
1998. ISBN 978-0-19-533191-2.
4. Pfau, W., An Optimizing Framework for the Glide Paths of Lifecycle Asset Allocation Funds, GRIPS
Policy Information Center Discussion Papers, 09-16, 2009.
5. Tobin, J., Liquidity Preference as Behavior Towards Risk, The Review of Economic Studies,
(67):65-86, 1958.
northerntrust.com | Line of Sight: Choosing a ‘Smart Beta’ Factor | 11 of 12
© 2015 Northern Trust Corporation. Head Office: 50 South La Salle Street, Chicago, Illinois 60603 U.S.A. Incorporated with limited liability in the U.S. Products and services provided by subsidiaries of Northern Trust
Corporation may vary in different markets and are offered in accordance with local regulation.
Past performance is no guarantee of future results. All material has been obtained from sources believed to be reliable, but its accuracy, completeness and interpretation cannot be guaranteed. This information does
not constitute investment advice or a recommendation to buy or sell any security and is subject to change without notice.
Northern Trust Asset Management is composed of Northern Trust Investments, Inc. Northern Trust Global Investments Limited, Northern Trust Global Investments Japan, K.K, NT Global Advisors, Inc., 50 South Capital
Advisors, LLC, and personnel of The Northern Trust Company of Hong Kong Limited and the Northern Trust Company.
This material is directed to professional clients only and is not intended for retail clients. For Asia-Pacific markets, it is directed to expert, institutional, professional and wholesale investors only and should not be relied
upon by retail clients or investors. For legal and regulatory information about our offices and legal entities, visit northerntrust.com/disclosures. The following information is provided to
comply with local disclosure requirements: The Northern Trust Company, London Branch; Northern Trust Global Services Limited; Northern Trust Global Investments Limited. The following information is provided to comply
with Article 9(a) of The Central Bank of the UAE’s Board of Directors Resolution No 57/3/1996 Regarding the Regulation for Representative Office. Northern Trust Global Services Limited Luxembourg Branch, 6 rue
Lou Hemmer, L-1748 Senningerberg, Grand-Duché de Luxembourg, Succursale d’une société de droit étranger RCS B129936. Northern Trust Luxembourg Management Company S.A., 6 rue Lou Hemmer, L-1748
Senningerberg, Grand-Duché de Luxembourg, Société anonyme RCS B99167. Northern Trust (Guernsey) Limited (2651)/Northern Trust Fiduciary Services (Guernsey) Limited (29806)/Northern Trust International
Fund Administration Services (Guernsey) Limited (15532) Registered Office: Trafalgar Court Les Banques, St Peter Port, Guernsey GY1 3DA.
Issued by Northern Trust Global Investments Limited. Issued in the United Kingdom by Northern Trust Global Investments Limited.
northerntrust.com | Line of Sight: Choosing a ‘Smart Beta’ Factor | 12 of 12
Q57753 (7/15)