Trend followers lose more often than they gain

Marc Potters , Jean-Philippe Bouchaud,+

arXiv:physics/0508104v1 [physics.data-an] 16 Aug 2005

Science & Finance, Capital Fund Management, 6-8 Bd Haussmann, 75009 Paris, France.
+

Service de Physique de lEtat

We solve exactly a simple model of trend following strategy, and obtain the analytical shape of
the profit per trade distribution. This distribution is non trivial and has an option like, asymmetric
structure. The degree of asymmetry depends continuously on the parameters of the strategy and
on the volatility of the traded asset. While the average gain per trade is always exactly zero, the
fraction f of winning trades decreases from f = 1/2 for small volatility to f = 0 for high volatility,
showing that this winning probability does not give any information on the reliability of the strategy
but is indicative of the trading style.

A question often asked by investors to fund managers, besides the average return of their strategies, is: What is
your fraction of winning trades? Implicitly, they expect the answer to be larger than 50% , as this would indicate
that the fund manager is more frequently right than wrong, and therefore trustworthy. We want to show in this
paper that this fraction is in fact meaningless. It depends entirely on the trading style of the manager, and tells very
little about the consistency of his returns. It is clear that one can make money on average even if the fraction f of
winning trades is low, provided the average gain per winning trade G exceeds the average loss per losing trade, L.
The condition is, clearly:
f G > (1 f )L.

(1)

Since asset prices are very close to being pure random walks, any statistical signal exploited by systematic traders has
an extremely small signal to noise ratio. The average profit per trade for any hedge fund is bound to be very small,
which means that the above inequality is close to an equality. If the typical holding period of winning trades is TG
and that of losing trades TL , one expects, for a random walk of volatility :
p
p
G TG ,
L TL .
(2)

Therefore, the fraction of winning trades is in fact a measure of the ratio of the holding periods of winning trades to
that of losing trades:
TG

TL

1f
f

2

(3)

For example, a 40% fraction of winning trades merely indicates that (unless the manager is really lousy) the typical
holding period of winning trades is 2 times that of losing trades. This, in turn, means that the manager is probably
mostly trend following, since by definition a trend following strategy stays in position when the move is favorable, but
closes it in case of adverse moves. Hence, conditioned to a winning trade, the holding period is clearly longer. The
opposite would be true for a contrarian strategy. Let us illustrate this general idea by two simple models. The first
one is completely trivial and not very interesting besides driving our point home. The second model is much richer; it
can be solved exactly using quite interesting methods from the theory of random walks and leads to a very non trivial
distribution of profits and losses. Besides its intrinsic interest, the model elegantly illustrates various useful methods
in quantitative finance, and could easily be used as a basis for a series of introductory lectures in mathematical finance.
The outcome of our calculations is that although the daily P&L of the strategy is trivial and reflects the statistics of
the underlying, the P&L of a given trade has an asymmetric, option-like structure!
The first model is the following: suppose that the price, at each time step, can only move up +1 or down 1 with
probability 1/2. The trend following strategy is to buy (sell) whenever the last move was +1 (1) and the previous
position was flat, stay in position if the last move is in the same direction, and close the position as soon as the move
is adverse. Conditioned to an initial buy signal, the probability that the position is closed a time n later is clearly:
pn = (1/2)n .

Electronic

address: [email protected],[email protected]

(4)

2
If n = 1, the trade loses 1; whereas if n > 1, the trade gains n 2. Therefore, the average gain is, obviously:

(n 2) (1/2)n = 0,

(5)

n=1

as it should, whereas the probability to win is 1/4, the average gain per winning trade G is 2 and the average loss per
losing trade L is 1. The average holding period for winning trades is:
TG = 4

n(1/2)n =

n=3

9
.
2

(6)

Let us now turn to our second, arguably more interesting model, where the log-price P (t) is assumed to be a
continuous time Brownian motion. This model has well known deficiencies: the main drawbacks of the model are
the absence of jumps, volatility fluctuations, etc. that make real prices strongly non-Gaussian and distributions fattailed [1]. However, for the purpose of illustration, and also because the continuous time Brownian motion is still
the standard model in theoretical finance, we will work with this model, which turns out to be exactly soluble. We
assume that on the time scale of interest, the price is driftless, and write:
dP
= dW (t),
P

(7)

where dW (t) is the Brownian noise and the volatility. From these log-returns, trend followers form the following
trend indicator (t), obtained as an exponential moving average of past returns:
Z t

e(tt )/ dW (t ),
(8)
(t) =

where is the time scale over which they estimate the trend. Large positive (t) means a solid trend up over the last
days. Alternatively, (t) can be written as the solution of the following stochastic differential equation:
1
d = dt + dW (t).

(9)

The strategy of our trend-follower is then as follows: from being initially flat, he buys +1 when reaches the value
(assuming this is the first thing that happens) and stays long until hits the value , at which point he sells
back and takes the opposite position 1, and so on. An alternative model is to close the position when reaches
0 and remain flat until a new trend signal occurs. The profit G associated with a trade is the total return during
the period between the opening of the trade, at time to ((to ) = ) and the closing of the same trade, at time tc
((tc ) = ). More precisely, assuming that he always keeps a constant investment level of 1 dollar and neglecting
transaction costs,
Z tc
dW (t).
(10)
G=
to

We are primarily interested in the profit and loss distribution of these trend following trades, that we will denote
Q(G). A more complete characterization of the trading strategy would require the joint distribution Q(G, T ) of G on
the one hand, and of the time to complete a trade T = tc to on the other; this quantity is discussed in the Appendix.
Obviously, G and evolve in a correlated way, since both are driven by the noise term dW . For definiteness, we will
consider below the case of a buy trade initiated when = +; since we assume the price process to be symmetric,
the profit distribution of a sell trade is identical. Now, the trick to solve the problem at hand is to introduce the
conditional distribution P (g|, t) that at time t, knowing that the trend indicator in , the profit still to be earned
between t and tc is g. This distribution is found to obey the following backward Fokker-Planck equation:


P (g|, t)
2 P (g|, t) 2 P (g|, t)
P (g|, t) 2 2 P (g|, t)
.
(11)
2
=
+
+
t

2
2
g
g 2
However, since is a Markovian process, it is clear that the history is irrelevant and at any time t, the distribution
of profit still to be made only depends on how far we are from reaching = , that is, P (g|, t) depends on but
not on t. Therefore, one finds the following partial differential equation:


2 P (g|) 2 P (g|)
P (g|) 2 2 P (g|)
= 0.
(12)
2
+
+

2
2
g
g 2

3
Eq. (12) has to be supplemented with boundary conditions: obviously when = the yet to be made profit must
be zero, imposing:
P (g| = ) = (g).

(13)

The final quantity of interest is the profit to be made when entering the trade, i.e:
Q(G) = P (g = G| = +).

(14)

We now proceed to solve Eq. (12). First, it is clear that all the results can only depend on the ratio / , i.e. on
the width of the trend following channel measured in units of the typical price changes over the memory time ,
that is, the orderof magnitude of the expected gains of the trend following strategy. One expects in particular that
in the limit / , the distribution of gains will become Gaussian, since the time needed to reach the edge
of the channel
is then much larger than the memory time of the process. We will from now on measure and G in
units of , and therefore set 2 = 1 hereafter. Now, Fourier transforming P (g|) with respect to g:
Z
d ig
e (),
(15)
P (g|) =
2
one obtains an ordinary differential equation for ():
2 ()
()
2( + i)
2 () = 0.
2

(16)

This is known as the Kummer equation (or, after a simple transformation, as the Weber equation) [2, 3]. The general
solution is the sum of two confluent Hypergeometric functions 1 F1 , with coefficients that are determined by two
boundary conditions. We already know that the boundary condition at should be () = 1, . The second
boundary condition turns out to be that () should be well behaved for , i.e. not grow exponentially with
. A way to be convinced and get some intuition on the solution is to expand () for small as:
() = 0 () + i1 ()

2
2 () + ...
2

(17)

2 22 = 2.

(18)

Plugging this into Eq. (16), one finds:

1 () 0;

The first two results are expected and simply mean that P (g|) is normalized for all , and that the average gain
is identically zero, as must indeed be the case of any strategy betting on a random walk. The last equation is more
interesting; the only reasonable solution of this equation is:
Z
Z
2
u2
2 () = 2
du e
dv ev ,
(19)

which for large behaves as ln . This is indeed expected: if the trade did open when hits a very large value instead
of at +, the time needed for to come back to values of order can be obtained by solving Eq. (8) without the
noise term, giving T ln . The total gain is the sum of T / random contributions, its variance is thus expected
to be ln . In fact, in the limit , the distribution of gains indeed becomes exactly Gaussian, as can be seen
by writing:
() = e

2
2

Z ()

(20)

The corresponding ode for Z () reads:

(21)

from which one immediately finds that for , Z ln independently of . Therefore, in that limit, the
characteristic function () indeed becomes Gaussian (in and thus in g). The above equation on Z will be useful
below to extract the large behaviour of . The conclusion of this analysis is that the large behaviour of ()
is a decreasing power-law:
2

()

/2

(22)

4
=1

0.5

0.4

Q(G)

0.3

0.2

0.1

10

FIG. 1: Shape of the distribution of gains, Q(G), for a rescaled channel width = 1. Note the strong asymmetry of the
distribution, which peaks at G 1.88, with a total probability of loss of 0.635. When the strategy makes money, the average
gain is G = 2.164.

This gives us our second boundary condition. The correct solution of our problem can then be written as:
() =

W ()
,
W ()

(23)

with W the following combination of hypergeometric functions:[4]

2
2
, , ( + i) 2
+ , , ( + i) ,
W () = 1 F1
( + i) 1 F1
2
4 2
4
2 2
( 4 )

(24)

related to the so-called the Weber function [3]. One can check, using the known asymptotic behaviour of the hyper2
geometric functions, that this particular combination indeed decays as /2 for large .
From these expressions, one can reconstruct the whole distribution Q(G), that we now describe. As already
mentioned above, in the limit the distribution becomes Gaussian. As decreases, the distribution becomes
more an more positively skewed: the fraction of winning trades decreases, but the average gain per winning trade
becomes larger. This is illustrated in Fig. 1 where we plot Q(G) for the intermediate case = 1. It is clear that
the most likely profit is negative; the probability to lose is in that case 1 f 0.635. The distribution can be
characterized further by studying its asymptotic tails for G . This can be done by observing that () has
poles for imaginary, corresponding to zeros of W (). For = 1, we find that the zeros closest to = 0 are
+ = 0.432i and = 5.058i, translating into the following large |G| behaviour:
Q(G) e0.432G

(G +);

Q(G) e5.058|G|

(G ),

(25)

showing again the strong asymmetry in the profit and loss distribution.
The large behaviour of () is important to control, in particular to determine accurately the numerical Fourier
transform that gives Q(G). Using Eq. (21), we find:



p
4
() exp 2i 3/2 || .
(26)
3
In the limit 0, the distribution becomes maximally skewed. Since the sell threshold is so close to the buy
threshold, most events correspond to a small (O(2)) immediate loss. Only with a small probability, also of order ,
is the strategy leading to an order 1 profit. In the small limit, one finds that the small expansion of () reads:

ln () 2 2i 2.38..4 + ... ,
(27)

5
=1

0.5

0.4

Q(G)

0.3

0.2

0.1

0
-5

10

FIG. 2: Simulated distribution of gains for a trend following strategy on the Swiss Franc/Dollar, compared to our theoretical
prediction based on a Gaussian model for the returns. As expected, the empirical distribution is indeed asymmetric, but also
fatter than predicted.

which translates into a diverging skewness, given by hG3 i/hG2 i3/2 1.798/ and a diverging kurtosis hG4 i/hG2 i2
4.545/. In that limit, Q(G) becomes a peak at G = 2 of width and weight 1 , plus a regular function of
total weight . The distribution Q(G) decaying exponentially for G 1, with a rate + ( 0) 0.810i, whereas
for the G negative region, we find that ( 0) i/, in agreement with our statement that Q(G) becomes
sharply peaked around G = 2.
We have performed a numerical simulation of the above simple trend following strategy on the Swiss Franc against
Dollar, using 5358 days between 1985 and 2005, with = 5 days, and choosing the channel width = 2 , where
is the historical volatility over the whole time period. The result for Q(G) is given in Fig. 2, and compared with the
theoretical prediction. The agreement is only qualitative, mostly due to the fact that the trading is in discrete time
(daily) and to non Gaussian character of the returns which makes the distribution Q(G) fatter than predicted by the
above model (see Fig. 3). However, the strong asymmetry is indeed observed; in particular, the loss probability is
found to be 0.605, not far from the theoretical prediction of 0.635.
Conclusion. We have therefore solved exactly a simple model for the profit and losses of a trend following strategy,
and obtained the analytical shape of the profit distribution per trade. This distribution turns out to be highly non
trivial and, most importantly, asymmetric, resembling the distribution of an option pay-off. The degree of asymmetry
depends continuously on the parameters of the strategy and on the volatility of the traded asset; while the average
gain per trade is always exactly zero, the fraction of winning trades decreases from f = 1/2 for small volatility to
f = 0 for high volatility, showing that this probability does not give any information on the reliability of the strategy
but is indicative of the trading style. In fact, we could repeat the same calculations as above for a mean-reverting
strategy, where the position of the trade is to sell when the trend indicator is high, and vice-versa. It is clear that
the distribution of gains in that case is the mirror image of that computed above; for a mean reverting strategy,
gains are more frequent than losses, but of a smaller amplitude. Note that the non trivial structure of the gain
distribution entirely comes from the conditioning on being associated to a given trade. If one asks, for example, for
the unconditional distribution of the daily returns of the strategy, then it is perfectly symmetrical and reproduces
exactly the return distribution of the underlying asset (see Fig. 3)!

6
0.6

Histogram

0.4

0.2

0
Daily returns

FIG. 3: Simulated distribution of the daily returns (in units of the daily volatility) of the trend following strategy. This
distribution is nearly symmetric, but clearly displays non Gaussian tails (a Gaussian distribution is shown in dotted line).

Appendix: Duration of the trades

In this appendix, we give an alternative derivation of the gain distribution Q(G) which also allows to gather some
information on their duration. The method presented in the main text is elegant precisely because it gets rids of all
temporal aspects. Suppose that at t = 0 a buy trade is opened, with = +. We will now focus on g , the profit
accumulated up to time t. Let us introduce the quantity R(, g , t) as the probability that the trade is still open at
time t, has accumulated a profit g and such that the trend indicator is . After Fourier transforming on g , this
quantity admits the following path integral representation:
(t)=i




Z t
1
2
d
D(u) exp 2
du ( )2 + 2 + 2 4 + V () ,
2 0
du

(t=0)=i
(28)
where V () enforces the constraint that never touched the lower edge of the channel , i.e. V () = 0 if >
and V () = + if < . Using standard techniques, one sees that the path integral is the Feynman-Kac
representation of the imaginary time Green function of the quantum harmonic oscillator with an impenetrable wall
at = . Setting again 2 = 1 and using a wave function representation, one can therefore write:
X
, t) = e2 /2+(i)2 /2(i)2 /2
R(,
m ( i)m ( i)eEm t/ ,
(29)
, t) = e2 /2+(i)2 /2(i)2 /2t/
R(,

where m and Em are the eigenvectors and eigenvalues of a quantum harmonic oscillator, obeying:


1 2
2
2 1

m () = Em m (),
+
+
2 2
2
2

(30)

with the following boundary conditions: m () = 0 (hard wall condition) and m ( ) 0. These two
conditions lead to a quantized spectrum of eigenvalues, indexed by an integer number m; as expected, the m () can
again be written in terms of Weber functions [3].
, t) one can compute the flux of fictitious particles just hitting the wall at time t = T and leaving the
From R(,
system, given by:


1
R(, , t)
,
(31)
J(, , t = T ) =
2
=

7
which, for = is precisely the joint probability that the profit of the trade is G and its duration is T (Fourier
transformed over G.) The integral over all Gs, corresponding to = 0, gives the unconditional distribution of trade
times. The result can be written in a fully explicit way if the trade closing condition is at = 0, in which case
the eigenvectors m are simply the odd levels of the harmonic oscillator and can be expressed in terms of Hermite
polynomials. More generally, the distribution of duration decays at large times as exp(E0 T / ), where E0 is the
ground state energy of the constrained Harmonic oscillator. One can also check that the integral over all T s of
J( = , , T ) obeys the same ode (with respect to the initial condition = + and up to a sign change of ) as
in the main text , as it should since the former quantity then becomes the Fourier transform of Q(G), with the
same boundary condition (g 0 when = ).

[1]
[2]
[3]
[4]

see e.g. J.P. Bouchaud, M. Potters, Theory of Financial Risks and Derivative Pricing, Cambridge University Press (2004).
see e.g. I. Gradshteyn, I. Ryzhik, Tables of Integrals, Series and Products, Academic Press, (1980) p. 1057-1059
see: W. N. Mei, Y. C. Lee, Harmonic Oscillator with potential barriers, J. Phys. A 16, 1623 (1983)
Note that we cannot write W as the Kummer function of the second kind U because as we follow the solution from to
+ we run into a branch cut of U at = 0 when the third argument falls on the negative real axis. Eq (24), on the other
hand, does not have a branch cut at = 0.