Professional Documents
Culture Documents
by PANAGIOTIS AVRAMIDIS
T
he Global Investment Performance Standards the original Dietz method, X i . We denote with 1 + b the
(GIPS) Handbook states that presentation of error factor—that is, X i = X i x(1 + b) . If b > 0, we have an
a portfolio’s performance should be based on overestimation of the return, and the opposite holds if b < 0.
time-weighted rates of return, which adjust for external For simplicity, we assume that during period i, the port-
cash flows. In particular, the guidance on the method- folio receives one large cash flow, CF, in a single period (the
ology cites that for periods after 1 January 2010, firms single-period assumption will be relaxed later).1 Right after
must assess the performance for interim subperiods the cash flow occurs, the portfolio value (including the cash
between all large cash flows and geometrically link the flow) is P1. The beginning and end values are, respectively, P0
performance to calculate periodic returns. For presen- and P2. The actual compounding factor at period i is
tation of historical periods, the Handbook recommends P1 − CF P2
using approximation methods, such as the original Dietz Xi = ×
P0 P1
and the modified Dietz methods, as a good compromise (1)
between accuracy and practicality in computing cash P2 CF
= × 1 − .
flow–adjusted time-weighted rates of return. P0 P1
This study aimed to identify the conditions in which whereas the original Dietz formula generates
the original Dietz method generates an unacceptable error
P − P − CF
and to answer such questions as how large a cash flow X i = 1 + 2 0
should be for the firm to determine that it distorts mean P0 + 0.5 × CF
(2)
performance if the portfolio is not valued at the time of the P2 − 0.5 × CF
= .
external cash flow. Knowing these answers should be help- P0 + 0.5 × CF
ful for investment firms wishing to give a rigorous defini-
Hence, the error factor, 1 + b, is
tion of “large” cash flow, as required by the GIPS standards.
1 − 0.5 × Z 2 1
1+ b = × , (3)
THEORETICAL GAINS FROM EXACT 1 + 0.5 × Z 0 1 − Z1
CALCULATION OF RETURN RATES
To begin, we denote with Xi a sample of n observations that where Zs equals CF/Ps, the weights of the cash flow
represent return rates in the form of a compounding factor relative to the portfolio value at the beginning (s = 0),
(i.e., Xi =1 + Ri over n periods). Assume that at period i, right after the cash flow (s = 1), and at the end (s = 2)
instead of the actual value Xi, we approximate the rate as in of the period.
Table 1: Theoretical Values of G X − G X Based on Equation 7 for a Range of Cash Flow Weights
Z0 Z1 Z2 Approx. Error Z0 Z1 Z2 Approx. Error Z0 Z1 Z2 Approx. Error
1% 1% 1% 0.01% 1% 1% 5% –2.01% 1% 1% 10% –4.52%
5 1 1 –1.95 5 1 5 –3.92 5 1 10 –6.38
10 1 1 –4.28 10 1 5 –6.20 10 1 10 –8.61
1 5 1 4.22 1 5 5 2.12 1 5 10 –0.50
5 5 1 2.18 5 5 5 0.13 5 5 10 –2.44
10 5 1 –0.25 10 5 5 –2.26 10 5 10 –4.76
1 10 1 10.01 1 10 5 7.79 1 10 10 5.03
5 10 1 7.86 5 10 5 5.69 5 10 10 2.98
10 10 1 5.29 10 10 5 3.17 10 10 10 0.53
2 ◆ WWW.CFAINSTITUTE.ORG
So, we can conclude that the original Dietz approxi- of periods increases. If we use an approximation for mul-
mation may produce sufficiently accurate mean estimates tiple periods, k = 5, the error continues to shrink to zero
for performance measurement if cash flows are relatively as we increase n (although at a lower rate than for a single
small (<1%) in relation to portfolio value or if the port- period). If all periods are approximated, k = n, the aver-
folio value is fairly stable (low portfolio volatility), which age approximation rises to 7.65%, the 25% largest errors
will ensure that the relative weights of the cash flow will exceed 11.44%, and the approximation error remains
be close. In either of these circumstances, firms could use almost unaffected by the number of periods.
the approximation instead of evaluating the portfolio at Individual scenarios in Table 3 (for n = 10) are more
the time of each cash flow without jeopardizing mean conclusive as to the effect of cash flow size on the approxi-
performance representation. In contrast, the higher the mation. The most influential factor is the size of the cash
portfolio volatility, the more likely that the approxima- flow relative to the value of the portfolio right after the
tion error will be large because of the deviation of weight time of cash flow occurrence, Z1. Overall, higher approxi-
Z1 from weights Z0 and Z2. mation errors are linked to greater distance between the
cash flow weight, Z1, and the cash flow weights relative to
NUMERICAL APPLICATIONS the beginning and ending portfolio values. Inversely, low
To see how the approximation method might be applied, or moderate but equal cash flow weights (Z0 = Z1 = Z2)
we will let –1 ≤ Ri < ∞ denote daily portfolio returns, and yield lower approximation errors. The rank in Table 3 is
we define Yi = ln(1 + Ri). Then, –∞ < Yi < ∞. Through the robust to the number of approximated periods, k, as well
restriction in Ri, we assume that the maximum amount as the number of periods, n.
that can be lost is the entire capital. Thus, the minimum Next, we consider the impact on the approximation
return is –100%. (This restriction exempts from our error of an increase in the duration of the period from
analysis leveraged portfolios, for which the total loss may daily to annual. Table 4 summarizes the average numeri-
exceed the initial capital.) If we assume that Yi follows a cal approximation errors for μ = 0% and σ = 2% (daily
normal distribution, with mean μ and standard deviation return rates) and μ = 15% and 30% (annual return rates).2
σ, then the transformed Xi = exp(Yi) = 1 + Ri follows a log- As Table 4 shows, the average error for a single period,
normal distribution, with μX = eμ+σ2/2 and σX = μX (eσ2 – 1). k = 1, increases from 0.78% for daily return rates to 0.95%
Initially, we will simulate the daily return rates of an for annual return rates. When approximation is applied
equity portfolio, so according to Brown and Warner (1985), over all periods, k = n, the average error increases from
we assume that μ is 0 and σ is 2%. Because the geometric 7.65% to 9.75% as the duration of the period changes
mean of the lognormal distribution of X is eμ (see Limpert, from daily to annual.
Stahel, and Abbt 2001), we conclude that GX = 1. The weights Consequently, higher expected returns yield larger
of the cash flows Z0, Z1 and Z2 range over five grid points (1%, approximation errors, which is consistent with the
5%, 10%, 15%, and 20%), yielding a total of 125 scenarios. theoretical findings (the approximation error is expo-
Table 2 shows that for a single approximated period, nentially dependent on expected return μ, GX = eμ). The
k = 1, and total periods n = 10, the average approxima- impact of volatility on error is not significant. Overall,
tion error is 0.78% with a standard deviation of 0.529%. the conclusion is that longer durations of the underlying
As expected, the error shrinks to zero when the number periods yield, on average, higher approximation errors.
G X − G X Z0 Z1 Z2 G X − G X Z0 Z1 Z2
0.028% 10% 10% 10% 0.619% 10% 10% 10%
0.013 5 5 5 0.178 5 5 5
0.001 1 10 20 0.059 1 10 20
10–4 1 5 10 0.030 1 5 10
10–6 1 1 1 0.005 1 1 1
4 ◆ WWW.CFAINSTITUTE.ORG