Asset Management:

2. Portfolio Choice (Part II)

Felix Wilke
Nova School of Business and Economics
Spring 2024
Dynamic portfolio choice
The dynamic portfolio-choice problem

• Horizon of one period is unreasonable for most investors.

• Pension funds (horizon of liabilities ≈ 20 years).
• Saving for house, college, new car etc.

• Investor can change portfolio weights every period over this horizon.
• What is a period? Year (individuals), quarter (institutions), … , millisecond (high-frequency
traders)?
• Why change?
• Time-varying investment opportunities (e.g., predictable returns and volatilities),
• Changes in income and liabilities,
• Time-varying risk preferences (e.g., when approaching the horizon (bequest)).

• Main insight from dynamic portfolio choice: optimal weight will vary with time t or
horizon, or both.

1/26
Setup of the dynamic portfolio-choice problem (1/2)

• Consider the portfolio choice at time t for an investor with investment horizon
T > t + 1.
• N risky asset assets with returns ri , for i = 1, . . . , N. At the beginning of each period
investor chooses portfolio weights xt = (x1,t , . . . , xN,t ) on the the risky assets.
• Wt is wealth at time t. Wealth varies from t to t + 1 due to portfolio return rp,t+1 (xt ),
which is a function of the portfolio choice xt . Wealth dynamics follow:

Wt+1 = Wt (1 + rp,t+1 (xt ))

• Objective is to maximize expected utility of wealth at horizon T by choosing a

dynamic trading strategy:
max Et [U(WT )]
{xt }
1− γ
Wt
• For tractability, assume CRRA preferences U(Wt ) = 1− γ .

2/26
Setup of the dynamic portfolio-choice problem (2/2)

• Suppose T = 5:

• The sequence of weights over time {xt } is called a dynamic trading strategy.
• Note, although chosen at t, weights xt+τ are not implemented immediately. Strategy
is a portfolio rule that optimally responds to changes in asset returns, volatilities,
utility etc.
• Examples of dynamic weights: time-varying return distribution. Invest
• more in risky assets when prices are low / exp. returns high after crash (value investing),
• less in high volatility states (volatility timing).
• {xt } may be constrained in practice: short-selling constraints (xi,t > 0), leverage
constrains (0 < xi,t < 1), turnover constraints,... 3/26
Dynamic programming (1/2)

• Solution to this problem can be found working backwards; we will illustrate this logic.
• Final wealth is the product of current wealth and uncertain one-period returns:
Wt+5 = Wt (1 + rp,t+1 )(1 + rp,t+2 ) . . . (1 + rp,t+5 )
• Standardize current wealth to Wt = 1 (weights do not depend on initial wealth).
• Given uncertain wealth at t + 4, choose portfolio weights that maximize expected
utility at T = t + 5:
max E [U(Wt+5 )] = max U(Wt+4 )E [U(1 + rp,t+5 (xt+4 ))]
xt + 4 xt+4

• Solution to this single-period problem (assuming CRRA utility):

1 µt+4 − rf,t+4
xt∗+4 =
γ σt2+4
• Conditional moments, as the state of the economy at t + 4 is unknown.
• Indirect utility: for xt∗+4 , the maximum expected utility obtained at t + 4,
Vt+4 = E [U(1 + rp,t+5 (xt∗+4 ))] 4/26
Dynamic programming (2/2)

• Now, solve the two-period problem: Given wealth at t + 3, choose xt+3 and xt+4 to
maximize expected utility at T = t + 5:

max E [U(Wt+5 )]
xt+3 ,xt+4

• However, optimal strategy conditional on any outcome at t + 4 already known: xt∗+4 !

• Thus, re-write problem as a one-period problem:

max E [U(1 + rp,t+4 (xt+3 )) · Vt+4 ]

xt+3

• The first part is identical to what a single-period investor would do at t + 3.

• The Vt+4 -term captures the advantage of being a long-term investor.
• The utility derived from this Vt+4 -term depends on uncertain wealth and investment
opportunities at t + 4, to which the investor responds optimally with xt∗+4 .
• The Vt+4 -term may interact with the investment decision at t + 3.

5/26
Graphical representation of dynamic programming

• Working back to t, we are solving such a one-period problem at each point in time.
• Dynamic portfolio choice over long horizons is first and foremost about solving
one-period portfolio choice problems!

6/26
Comparison to myopic portfolio choice

• How does this differ from a static problem over T = t + 5 periods?

Myopic (static) problem:

• In a one-period problem, the investor chooses portfolio weights only once at the
beginning of the period, which results in end-of-period wealth. This is also called a
buy-and-hold problem.
• Optimal dynamic strategies must always do at least as well as buy-and-hold
portfolios. Why?
7/26
Lesson 1: A long-run investor should not buy and hold

• Buy and hold solves a single, long-horizon problem.

• Special case of dynamic investing where the investor’s optimal choice is to do nothing.

• Thus, dynamic portfolios can do everything buy and hold portfolios do, but also
much more!
• In practice, optimal long-horizon investing is not to buy and hold; long-horizon investing
is a continual process of buying and selling.

• Basic idea: Suppose you believe the optimal weight in stocks for the next ten years is
50%.
• You need to rebalance each period!
• If not, over a sufficiently long period of time, you will have 100% in risky assets. That is
not what you wanted, right?

8/26
Lesson 2: Long-term investing is not so different from short-term investing

• Dynamic programming shows that long-run investors do everything that short-run

investors do!

• However, long-run investors can do more, because they have the advantage of a long
horizon.
• The horizon effect enters through the indirect utility (Vt+τ ) in each one-period problem.
• For instance, suppose at t you predict that stock market returns will be high from t + 1 to
t + 2. How might this affect the optimal portfolio choice xt∗ ?
• Some will invest more risky, because future returns can compensate if returns are low from t to
t + 1.
• Some will invest less risky, to ensure that they have sufficient money to invest when it is most
attractive at t + 1.
• Exact solution depends, among other things, on Covt (rt+1 , rt+2 ) and intertemporal
smoothing preferences with other utility functions.

9/26
Rebalancing
Rebalancing when returns are not predictable

• Suppose returns are i.i.d. (as coin tosses). That is, returns every period have the
same µ and σ. Assume also that rf is fixed. Then, the dynamic strategy is a series of
identical one-period strategies:
Long-Run Weight (t) = Short-Run Weight (t)

• Under i.i.d. returns, investors rebalance back to the same optimal xt∗ each period.
10/26
The case for rebalancing

• Rebalancing is a counter-cyclical strategy:

• If equity has increased over the last period, the equity proportion is too high relative to
optimal => sell equity.
• If equity has decreased over the last period, the equity proportion is too low relative to
optimal => buy equity.
• This counter-cyclicality also has advantages if returns exhibit mean reversion or are
predictable (but mean reversion is not necessary to capture a rebalancing premium).
• Rebalancing buys assets that have declined in price (high future expected returns) and
sells assets that have risen in price (low future expected returns).
• Intuition from dividend discount model (price P when constant future dividends are
discounted at rate R):
D
P=
R
• Buying low and selling high is how any investor, long or short horizon, makes money.
However, being counter-cyclical goes against investors’ behavioral tendencies.

11/26
Counter-cyclical rebalancing in the Great Depression

• With quarterly rebalancing: sell some stocks before the peak in August 1929; buy
some stocks before they rebound after May 1932.
• Relative to not rebalanced portfolio (0.6 * pure stock + 0.4* pure bond):
• reduces variance …
• and increases returns.
• Picture looks similar for other periods, e.g., Global Financial Crisis 2008/2009.
12/26
Where does the increase in returns come from?

• Mean reversion/predictability AND diversification/rebalancing premium:

1. Portfolio diversification decreases portfolio variance without reducing arithmetic
portfolio return.
• This benefit is obtained in a single period but dies out if you do not rebalance (weight in
risky assets → 100%).
2. However, diversification also increases geometric portfolio return.
• This “diversification return” exists for a long-term investor and is collected by rebalancing
(hence, “rebalancing premium” is the better name for the long-term effect).
• Also known as variance reduction return1 :
1
geometric mean ≈ arithmetic mean − × variance
2
• The greater the volatility reduction, the greater the rebalancing premium!
• Intuition: Which two consecutive returns do you prefer, (90%, −50%) or (10%, −10%)?
1 The geometric return g and the arithmetric return r are related: (1 + r) = exp(g). Relationship between their
means is approximately E (g) ≈ E (r) − 21 σ2 (where σ2 is volatility of r) and this relation holds exactly for
log-normal distributions.
13/26
Erb and Harvey (2006)

14/26
The implementation of rebalancing

• Most investors are pro-cyclical, and chase returns. They invest precisely at the wrong
time: when prices are high, investors are drawn by past high returns, but future
expected returns are low.
• Investor Policy Statements (IPS) might help investors by committing ex-ante.
• Do investors rebalance?
• Calvet, Campbell and Sodini (2009) find some inertia (i.e. stick to buy and hold) for
Swedish households but many rebalance. Wealthy, educated investors hold more
diversified portfolios and rebalance more actively.
• Brunnermeier and Nagel (2008) find inertia to be the dominant factor for US households.
• Also institutional investors often fail to rebalance (e.g., CalPERS sold equities at lowest
point in 2008/2009 when expected returns were hightest!)
• If equity portfolio weight is 61% at the end of a quarter you, should you rebalance
back to the 60% target?
• In practice, instead of calendar time, consider rebalancing bands.
• Rebalancing bands are functions of transaction costs, liquidity, asset volatility, and
minimum transaction sizes.
15/26
Strategic asset allocation
Opportunistic strategies when returns are predictable (1/2)

• If returns are not i.i.d. (say, high volatility this period is followed by high volatility
next period), the static portfolio weight changes every period.
• Myopic weight = one-period solution (tactical asset allocation)
• Changing myopic allocations are optimal each period for a single horizon investor.
µ −r
The time-varying short-run weight is γ1 t σ2 f,t , where µt = Et [rm,t+1 ], σt2 = Vart [rm,t+1 ].
t
• With a long horizon, the optimal portfolio strategy over risky assets can be written as

Long-run weight (t) = Short-run weight (t) + Opportunistic weight (t)

| {z } | {z }
myopic portfolio hedging demand

which changes over time as the means and standard deviations of returns change.
• Opportunistic weight is called hedging demand because the portfolio hedges against
changes in the investment opportunity set.
• That is, long-run investors can opportunistically take advantage of time-varying
returns (strategic asset allocation).

16/26
Opportunistic strategies when returns are predictable (2/2)

• Campbell and Viceira (2002) decomposition:

Long-run weight (t) = Long-run myopic weight (t) (1)
+ [Short-run weight (t) - Long-run myopic weight (t)] (2)
+ Opportunistic weight (t) (3)
µ̄−r¯
1. Long-run fixed weight determined by long-run average return and volatility: γ1 ¯2 f
σ
This is the constant rebalancing weight in the i.i.d. case!
2. Tactical asset allocation: response of both short- and long-run investors to changing
means and volatilities (constant rebalancing weight plus temporary deviation)
3. Captures how long-term investor can take advantage of time-varying, predictable
returns in ways short-run investors cannot. (Long-term returns are more predictable
(through mean reversion) than single period returns).
• Strategic asset allocation is the long-run weight and is the sum of the long-run fixed
weight, tactical asset allocation, and the opportunistic weight.
17/26
Characterizing the opportunistic weight

• Difficult, but two broad determinants:

1. Investor-specific: risk tolerance (like in one-period portfolio), smoothing preferences, and
horizon.
2. Asset-specific: how do returns vary over time?
• Interaction between horizon and time-variation is crucial:
• If expected returns are high in the future (say from t + 1 to t + 2), a long-term investor
might want to adjust his portfolio already at t. A short-term investor cannot use this
information at all.
• A strategy can be noisy in the short run, but over the long run volatility mean-reverts, and
long-term investors can afford to invest.
• Rebalancing rule: conservative lower bound for “buying low and selling high.”
• Both tactical and opportunistic weights require estimates of future expected returns
and volatilities.
• How good are predictions of µt and σt empirically?

18/26
Predictability
On stock return predictability (1/2)

• Predicting stock returns (in the time series) is TERRIBLY difficult! In the cross-section,
it is considerably easier (long-short equity).
• Predictive regression approach: regress returns between t and t + h (h = 1m, ..., 5y)
on predictor variables Xt
rt,t+h = a + bXt + ε t,t+h

• Long history of evidence consistent with in-sample predictability:

Predictors: Cash-flow based (dividend/price, earnings/price); Business cycle
indicators (term spread); Technical (past volatility, momentum)
• Consider the Gordon (1963) Dividend Discount Model:
• P = D ⇒ E(r) = D/P + g
E(r)−g
• If cashflows are largely unpredictable (or are constant), then expected returns should
move around with valuation ratios.
• High E/P or D/P (low prices) ⇒ high future returns.

19/26
On stock return predictability (2/2)

• Example: regressing monthly U.S. equity returns on lagged dividend/price ratio from
1962 to 2014:
rm
t,t+1 = a + bDPt + et,t+1
b = 0.33 with t-stat ≈ 2, R2 < 0.006
• Conditionally expected (or forward-looking) return µt is hard to estimate:
• Over-lapping observations,
• In-sample analysis with low R2 (out-of-sample R2 even smaller),
• Coefficient b unstable in sub-samples.
• Goyal and Welch (2008) and others question extent and exploitability of this
predictability out-of-sample, i.e., when making investment decisions in real-time.

Andrew Ang (2014): ”... the evidence for predictability is weak, so I recommend that both the
tactical and opportunistic portfolio weights be small in practice. Opportunistic demands
become much smaller once investors have to learn about return predictability or when they
take into account estimation error.”
20/26
On the predictability of market volatility

• Tactical asset allocation requires an estimate of conditional expected return (µt ) as

well as volatility (σt ).
• While predicting returns is difficult, predicting volatility is much less so:
No autoregressive component in returns, but clearly in squared returns. Volatility is
persistent and positively auto-correlated.
• Financial econometrics: time-series auto-regressive type models used to filter out
conditional variance (Vart (r+1 ) = σt2 ). GARCH(1,1) is the model most applied:
σt2 = a + bσt2−1 + cε2t−1

21/26
Application: Volatility timing in tactical asset allocation (1/3)

• Portfolio of U.S. stock market (MKT) and 1 month t-bill (RF).

• Summary statistics (monthly):
• Average return (St.Dev.): MKT = 0.93% (5.4%), RF = 0.28%.
• Using γ = 4, what is the optimal myopic weight in the risky asset?
1 µ̄ − R̄RF 1 0.93% − 0.28%
x∗ = = = 0.55 (as before)
γ σ̄2 4 (5.4%)2
• Consider a volatility timing strategy using σ̂MKT,t = St.Dev.(RMKT,t−12:t−1 ) to proxy for
conditional volatility in month t.
• x∗ = Target volatility/σ̂MKT,t .
vol,t
• Invest more (less) when recent volatility was low (high).
• Choose Target volatility so that the average of x∗ equals 0.55.
vol,t
• Then, the tactical weight is: 1 µt −rf,t 1 µ̄−r̄f ∗ ∗
γ σ2 − γ σ̄2 = xvol ,t − x
t
(short-run weight - long-run myopic weight)

22/26
Application: Volatility timing in tactical asset allocation (2/3)

Myopic Volatility timing

Avg. weight 0.55% 0.55%
• Results: Avg. return 7.68% 7.57%
St. Dev. 10.34% 8.23%
Sharpe 0.41 0.51

• Increase in annualized Sharpe of 0.10 is economically large!

• Predictability of volatility (and unpredictability of returns) leads to reduction in risk
(without harming average return) of volatility timing strategy: volatility ↑, tactical
weight ↓.
• Volatility timing attractive for many strategies (Muir and Moreira, 2017).

23/26
Application: Volatility timing in tactical asset allocation (3/3)

• Another example:
Static 60% equities / 40% T-bill strategy vs. volatility timing based on VIX:

• If volatility timing does so well, why doesn’t everyone do it? Portfolio weights move
all over the place, from x∗ = 0 to x∗ = 1.5. Hard to implement (liquidations, funds
mandate, ...)

24/26
Conclusions

• Rebalancing is the foundation of any long-term investment strategy!

• Under i.i.d. returns the optimal policy is to rebalance to constant weights for both
short- and long-term investors.

• When returns are predictable, the optimal short-run portfolio changes over time, and
the long-run investor could add additional opportunistic strategies.

• Although predicting returns is hard, volatility timing has clear benefits.

• Practical strategic asset allocation advice: pick reasonable weights on a few easily
investible asset classes and rebalance.

25/26
References

• Ang, Asset Management: A Systematic Approach to Factor Investing, Ch. 4

• Articles
• Brunnermeier and Nagel, 2008, Do Wealth Fluctuations Generate Time-Varying Risk Aversion?
Micro-Evidence on Individuals’ Asset Allocation, American Economic Review
• Calvet, Campbell and Sodini, 2009, Measuring the Financial Sophistication of Household, American
Economic Review P&P
• Campbell and Viceira, 2002, Strategic Asset Allocation: Portfolio Choice for Long-Term Investors
• Erb and Harvey, 2006, The Tactical and Strategic Value of Commodity Futures, Financial Analysts
Journal
• Goyal and Welch, 2008, A Comprehensive Look at The Empirical Performance of Equity Premium
Prediction, Review of Financial Studies
• Moreira and Muir, 2017, Volatility-Managed Portfolios, Journal of Finance

• AQR podcast: The Curious Investor

• Calculated Risks Rise of the Machines

26/26