ABSTRACT

TREND WITHOUT HICCUPS - A KALMAN FILTER APPROACH

By

ERIC BENHAMOU
DATE: April 2016

Have you ever felt miserable because of a sudden whipsaw in the price that triggered an
unfortunate trade? In an attempt to remove this noise, technical analysts have used various
types of moving averages (simple, exponential, adaptive one or using Nyquist criterion).
These tools may have performed decently but we show in this paper that this can be
improved dramatically thanks to the optimal filtering theory of Kalman filters (KF). We
explain the basic concepts of KF and its optimum criterion. We provide a pseudo code for
this new technical indicator that demystifies its complexity. We show that this new
smoothing device can be used to better forecast price moves as lag is reduced. We provide
4 Kalman filter models and their performance on the SP500 mini-future contract. Results
are quite illustrative of the efficiency of KF models with better net performance achieved
by the KF model combining smoothing and extremum position.
 ACKNOWLEDGEMENTS

The author would like to acknowledge support from Thomson Reuters. Data are from

Thomson Reuters Eikon while screenshots and source codes are done with TR Eikon

Trading Robot. The author also thanks Denis Dollfus for fruitful conversations.

Disclaimer: The views and opinions expressed in this article are those of the author and do

not necessarily reflect the official position or policy of Thomson Reuters. Materials is

provided “AS IS” with no representations or warranties. Although the author has taken

great care to provide accurate information, neither the author, nor Thomson Reuters, nor

anyone else associated with this publication, shall be liable for any loss, damage, or liability

directly caused or alleged to be caused by this paper. Materials contained herein are not

intended to provide any specific advice or investment recommendations.

ii
 INTRODUCTION

1 INTRODUCTION

Have you ever felt angry because a sudden price whipsaw triggered an unfortunate signal

and a resulting bad trade? Prices have inherent blips and jerks that are not easy to control.

Moreover, prices are inputs for technical analysis indicators. This can result in corrupted

or non-efficient indicators. In an ideal world, one would like prices heading to a clear

direction. Remember the old adage: “trade with the trend”. But in real life, prices hiccups

create noise and perturb the signal.

A first attempt to remove these yanks and jolts is to smoothen prices with moving averages.

However, moving averages suffer from two flaws: lags and no dynamics. The first

drawback, delay in moving average response, is widely known as moving averages used

past data. Adaptations to moving averages have been suggested (exponential, adaptive, zero

lag or Nyquist criterion based moving averages). Durchner in [Dürschner-2012] suggested

the use of Nyquist criterion to create moving average 3.0 with no lag.

This is intellectually very enticing as the lag is completely removed. This improves moving

averages from Patrick Mulloy [Mulloy-1994] with zero lag or the attempts by John Ehlers

to provide sophisticated moving averages ([Ehlers-2001a] or [Ehlers-2001b]). But this

does not address the second problem of capturing price dynamics. What we mean by price

dynamics is the price movement. If we can identify that prices are moving upwards

(respectively downwards), then a good guess for next price observation should be higher

(respectively lower) than the current price. Let us pause for a moment and imagine instead

of prices, we were looking at car position using a GPS. We measure the car position with a
1
GPS but with some noise as the signal is not perfectly accurate. Could we capture the car

dynamics to compute the best guess at next time step and hence reduce noise in car

position? The answer is yes! And guess what, this is what your car GPS is doing. This

theory simply explained is referred to as Kalman filter (from its inventor [Kalman-1960])

shortened to KF in this paper. It was created for the spatial industry to remove noise and

capture shuttle movements. In a scientific way, the Kalman filter is an efficient recursive

filter that estimates the state of a dynamic system from a series of incomplete and noisy

measurements to estimate the best forecast according to an assumed distribution. In the

original paper, Kalman assumes a Gaussian distribution of noise but extended version can

now cope with more advanced distribution (see [Kalman-Filter-Wikipedia]). In this

article, we first revisit moving averages and then present different Kalman filter models

and their implementation to create trading strategies. We then provide performance results

for our 4 KF models on one year of data of the E-mini-SP continuation future.

1.1 Motivation for smoothing

Smoothing prices is natural. The basic idea is to remove noise from prices to better identify

important patterns or trends. Remember, when we trade, we want the big picture. So

smoothing enables us removing bumps, bangs, bounces, and shocks and getting an average

clean signal. If we believe that prices do not follow a random walk model, the smoothened

signal provides us a clear directional signal.

2
1.2 Impact for trading strategies

Conversely, if we do not smoothen prices, we could act on tugs, wrenches, or snatches that

are against the trend and result in bad trades. Smoothing is the right way! But we need to

be careful. If we smoothen with lag (one of the major drawbacks of moving averages), we

act with delay and enter trades too late, potentially facing reverse direction markets. In an

ideal world, we would like the smoothing technique to have zero lag and to provide a fist

move advantage.

3
 MATERIAL AND METHODS

2 REVIEW OF MOVING AVERAGES

2.1 The usual moving averages

The usual way to remove noise in prices is with moving averages. Let us denote weights

by 𝑤𝑖 for the time𝑇𝑖 , where 𝑖 goes between 0 to 𝑁. Then, the moving average is given by

𝑀𝐴(𝑤0 , … , 𝑤𝑁 ) = ∑𝑁
𝑖=0 𝑤𝑖 𝑃𝑖 (EQ 2.1)

Whose lag is

∑𝑁
𝑖=0 𝑤𝑖 𝑇𝑖
𝐿𝐴𝐺 𝑀𝐴(𝑤0 , … , 𝑤𝑁 ) = (EQ 2.2)
∑𝑁
𝑖=0 𝑤𝑖

If we do a moving average of a moving average, equation (2.1) becomes

𝑀𝐴⨀𝑀𝐴(𝑤0 , … , 𝑤𝑁 ) = ∑𝑁 𝑁
𝑖=0 𝑤𝑖 (∑𝑗=0 𝑤𝑗 𝑃𝑖+𝑗 ) (EQ 2.3)

And the corresponding lag is

∑𝑁
𝑖=0,𝑗=0 𝑤𝑖 𝑤𝑗 𝑇𝑖+𝑗
𝐿𝐴𝐺 𝑀𝐴⨀𝑀𝐴(𝑤0 , … , 𝑤𝑁 ) = (EQ 2.4)
∑𝑁
𝑖=0,𝑗=0 𝑤𝑖 𝑤𝑗

We can easily derive similar formula for a recursive moving average at the order kth:

𝑀𝐴⨀𝑘 (𝑤0 , … , 𝑤𝑁 ) = ∑𝑁
𝑖1 =0,…,𝑖𝑘 =0 𝑤𝑖1 … 𝑤𝑖𝑘 𝑃𝑖1 +⋯+𝑖𝑘 (EQ 2.5)

The resulting lag is

∑𝑁𝑖1 =0,…,𝑖𝑘 =0 𝑤𝑖1 …𝑤𝑖𝑘 𝑇𝑖1 +⋯+𝑖𝑘

𝐿𝐴𝐺 𝑀𝐴⨀𝑘 (𝑤0 , … , 𝑤𝑁 ) = (EQ 2.6)
∑𝑁𝑖1 =0,…,𝑖𝑘=0 𝑤𝑖1 …𝑤𝑖𝑘

2.2 Explicit Lag Computation

4
 𝑖
Prices are sampled with equidistant time steps 𝑇𝑖 = 𝑁 𝑇. Formulae (EQ.2.6) can be easily

computed in terms of its first order value as follows:

𝐿𝐴𝐺 𝑀𝐴⨀𝑘 (𝑤0 , … , 𝑤𝑁 ) = 𝑘 𝐿𝐴𝐺 𝑀𝐴(𝑤0 , … , 𝑤𝑁 ) (EQ 2.7)

(See Proof A.1:)

Furthermore, if we combine recursive moving averages, it is easy to find back the results

of Mulloy. In the case of a moving average of moving average, the only possible choice

with zero lag whose coefficient sum is equal to 1 is the double moving average:

𝐷𝐸𝑀𝐴 = 2𝑀𝐴- 𝑀𝐴⨀2 (EQ 2.8)

(See Proof A.2)

And for the triple moving average (if we impose the additional constraint that the 3rd order

recursive moving average coefficient is 1), we have

𝑇𝐸𝑀𝐴 = 3𝑀𝐴- 3. 𝑀𝐴⨀2 + 𝑀𝐴⨀3 (EQ 2.9)

(See Proof A.3)

3 INTRODUCTION TO KALMAN FILTER

3.1 Basic concepts

Kalman filter is a recursive algorithm that has been invented in the 1960s to track a moving

target, remove any noisy measurements of its position and predict its future position. In

finance, KF has been used by the asset management industry for various purposes. KF is

an optimal choice in many cases and do at least better than a moving average smoothing.

Dao et al [Bruder-Dao-Richard-Roncalli-2011] and [Dao-2011] showed that for price

following random walk with noise, KF is equivalent to the optimal exponential moving

average with parameter equal to Kalman gain. However, for more sophisticated dynamics,

5
like a linear Gaussian model, KF is the optimal choice and the most efficient computational

solution for finding the model parameters.

KF has also been used over the last decade by different authors. Martinelli and Rhoads in

[Martinelli-2006] and [Martinelli-Rhoads-2010] used Kalman filter to find optimal guess

for trading strategies on stocks. Haleh-et-al in [Haleh-et-al-2011] used Extended Kalman

filter for forecasting stock prices, combining technical and fundamental data. They showed

it outperformed regression and neural networks. Ernie Chan in [Chan-2013] suggested

using KF for pair correlation trading, while Cazalet and Zheng in [Cazalet-Zheng-2014]

used KF for hedge fund replication.

In a general way, Kalman filter consider a linear dynamic system given by

𝑋𝑡+1 = 𝜙𝑋𝑡 + 𝑐𝑡 + 𝑤𝑡 (EQ 3.1)

𝑌𝑡 = 𝐻𝑋𝑡 + 𝑑𝑡 + 𝑣𝑡 (EQ 3.2)

Where 𝜙 is the state transition matrix, 𝐻 the measurement matrix, 𝑤𝑡 the model noise, 𝑋𝑡

the state vector, 𝑌𝑡 the measurement vector, 𝑣𝑡 the measurement noise, 𝑤𝑡 and 𝑣𝑡 the

independent white noises with zero mean and their variance matrices given by 𝑄 and 𝑅

respectively. 𝑐𝑡 , respectively 𝑑𝑡 , is the drift of the state vector, respectively the

measurement vector. The corresponding Kalman filter is:

Prediction step: 𝑋𝑡+1|𝑡 = 𝜙𝑋𝑡|𝑡 + 𝑐𝑡 (EQ.3.3)

With 𝑃𝑡+1|𝑡 = 𝜙𝑃𝑡|𝑡 𝜙 𝑇 + 𝑄 (EQ.3.4)

Correction step: 𝑋𝑡+1 = 𝑋𝑡+1|𝑡 + 𝐾𝑡+1 (𝑌𝑡+1 − 𝑌𝑡+1|𝑡 ) (EQ.3.5)
With 𝑌𝑡+1|𝑡 = 𝐻𝑋𝑡 + 𝑑𝑡
−1
With Kalman gain 𝐾𝑡+1 = 𝑃𝑡+1|𝑡 𝐻 𝑇 [𝐻𝑃𝑡+1|𝑡 𝐻 𝑇 + 𝑅] (EQ.3.6)
With 𝑃𝑡+1|𝑡+1 = [𝐼 − 𝐾𝑡+1 𝐻]𝑃𝑡+1|𝑡 (EQ.3.7)

6
KF works in a two-step process (prediction and correction steps). The algorithm is

recursive, can run in real time, using only the present input measurements, the previously

calculated state and its uncertainty matrix.

Obviously, one needs to specify the state and measurement vector. A logic choice is to use

a physical system with concepts similar to speed and acceleration:

𝑥𝑡+1 = 𝑥𝑡 + 𝑥̇ 𝑡 δt + 1/2 𝑎𝑡 δt 2 (EQ.3.8)

𝑥̇ 𝑡+1 = 𝑥̇ 𝑡 + 𝑎𝑡 δt (EQ.3.9)

𝑦𝑡 = 𝑥𝑡 + 𝑣𝑡 (EQ.3.10)

Where 𝑥𝑡 and 𝑥̇ 𝑡 are price and rate of change of stock price at time 𝑡 (similar to position

and speed). 𝑎𝑡 can be seen as the acceleration of price at time 𝑡. It is considered to be a

model noise. 𝑇 is the sampling period, 𝑦𝑡 the measurement, 𝑣𝑡 the measurement noise. This

can be analyzed as a KF system with

1
𝑥𝑡 1 δt δt 2
𝑋𝑡 = [𝑥̇ ] , 𝜙 = [ ] , 𝑤𝑡 = [2 ] 𝑎𝑡 , 𝐻 = [1 0], 𝑐𝑡 = 0 (EQ.3.11)
𝑡 0 1 δt
This is named model One. This model has the advantage to take into account a certain

dynamic compared to the simple Random Walk model that is often used in the KF literature

where there is no speed term. In our model One, The speed is initially estimated as the

difference between two consecutive prices. The parameters to estimate are the following (4

in total)

𝑝1 𝑝1 𝑝2 𝑝1 𝑝4 0
𝑄 = [𝑝 𝑝
1 2 𝑝2 𝑝2 ] , 𝑅 = [𝑝3 ], 𝑃𝑡=0 = [ 0 𝑝4
] (EQ.3.12)

It is interesting to note that this model is very closed to a local linear trend model. Indeed,

the local linear trend model writes as

xt+1 = xt + βt + w1,t (EQ.3.13)

βt+1 = βt + w2,t (EQ.3.14)

7
 yt = xt + vt (EQ.3.15)
We can notice that in this specific case, the KF parameters are the following:

𝑥𝑡 1 1 𝑤1,𝑡
𝑋𝑡 = [𝛽 ] , 𝜙 = [ ] , 𝑤𝑡 = [𝑤 ] , 𝐻 = [1 0], 𝑐𝑡 = 0 (EQ.3.16)
𝑡 0 1 2,𝑡

The parameters to estimate are the following (5 in total)

𝑝1 𝑝1 𝑝2 𝑝1 𝑝4 0
𝑄 = [𝑝 𝑝
1 2 𝑝2 𝑝2 ] , 𝑅 = [𝑝3 ], 𝑃𝑡=0 = [ 0 𝑝5
] (EQ.3.17)

This model has almost the same parameters as model one. This is named model Two.

Comparing equation 3.12 and 3.17, we know that model One and Two should have very

similar behavior.

We can create a more general two factors model with contribution to price split between a

short term 𝑥𝑡1 and a long term 𝑥𝑡2 . This leads to:

x1t+1 = a11 x1t + a12 xt2 + w1,t (EQ.3.18)

2
xt+1 = a22 xt2 + w2,t (EQ.3.19)
yt = h1 x1t + h2 xt2 + vt (EQ.3.20)
In this specific model, we have the following parameters

x1 a11 a12 w1,t h1

Xt = [ t2 ] , ϕ = [ 0 a22 ] , wt = [ w2,t ] , H = [h2 ] , ct = 0 (EQ.3.21)
xt
We call this model Three. Because of its generality, this model encompasses models 1, 2.

The parameters to estimate are the following (10 in total)

p1 p2 p4 p6 p6 p7 p6 p9 0
ϕ= [0 p3 ] , H = [p5 ] , Q = [ p6 p7 p7 p7 ] , R = [p 8 ], Pt=0 = [ ] (EQ.3.22)
0 p10

The last model we use is a model inspired by a combination of oscillators and the previous

model. In this model, we use the price position with respect to its extremums as in the fast

stochastic oscillator. We denote the variable 𝐾𝑡 over a d period given by

Current Close−Lowest Low(d)

K dt = Highest High(d)−Lowest Low(d) × 100 (EQ.3.23)

8
We denote by 𝐿𝑑𝑡 = 𝐿𝑜𝑤𝑒𝑠𝑡 𝐿𝑜𝑤(𝑑) and 𝐻𝑡𝑑 = 𝐻𝑖𝑔ℎ𝑒𝑠𝑡 𝐻𝑖𝑔ℎ(𝑑) the lowest low and
highest high over d period. We use in our example a 14 days period. As in model 3, we also
split the contribution of the price due to short term 𝑥𝑡1 and long term 𝑥𝑡2 . This leads to:
1
𝑥𝑡+1 = 𝑎11 𝑥𝑡1 + 𝑎12 𝑥𝑡2 + (𝑀1 − 𝑁1 𝐾𝑡𝑑 ) + 𝑤1,𝑡 (EQ.3.24)
2
xt+1 = a22 xt2 + (M2 − N2 K dt ) + w2,t (EQ.3.25)
h0 xt −Ld
With K dt = t
(EQ.3.26)
Hd d
t −Lt

yt = h1 x1t + h2 xt2 + vt (EQ.3.27)

In this specific model, we have the following parameters

x1 a11 a12 w1,t h1 M1 − N1 K dt

Xt = [ t2 ] , ϕ = [ 0 ] , wt = [ w2,t ] , H = [ ] , c = ( ) (EQ.3.28)
xt a22 h2 t M2 − N2 K dt
We call this model Four. Because of its generality, this model encompasses models 1, 2

and 3. It capture short and long term effect as well as position with regards to extrema like

what oscillators do. This is by far the most realistic model. Short term factor 𝑥𝑡1 models

extreme market reactions that last for a few days. Long term factor 𝑥𝑡2 is only influenced

by itself and not by the short term 𝑥𝑡1 . The parameters to estimate are the following (15 in

total) (the same set as model 3 and 5 additional parameters)

p1 p2 p4 p 6 p6 p7 p6 p9 0
ϕ= [0 p3 ] , H = [p5 ] , Q = [ p 6 p7 p7 p7 ] . R = [p8 ], Pt=0 = [ 0 p10
] (EQ.3.29)

p11 − p12 K dt
ct = ( ) , p15 = d (EQ.3.30)
p13 − p14 K dt
3.2 Pseudo code

/// Initialization phases: parameters contains

/// - initial value for model state + measurement of model
/// - measurement of state and model variance
Kalman2D k = new Kalman2D(parameters);
k.Setup( parameters );
int length = timeSeries.Length;

Point2D[] kalmanResult = new Point2D[length];

9
/// the loop to update in real time
for( int i = 0; i<length; ++i )
{
if( i<Period )
{
k.Predict();
k.Update(timeSeries[i]);

kalmanResult.Set(0, timeSeries[i]);
kalmanResult.Set(1, timeSeries[i]);
}
else
{
k.Predict();
kalmanResult.Set(0, k.X.Get(0,0) );
k.Update(timeSeries[i]);
kalmanResult.Set(1, k.X.Get(0,0) );
}
}

Where the pseudo code for Predict and Update value is given as follows:

// <summary>
// Predict the state
// </summary>
public void Predict()
{
// Predict to now, then update.
// Predict:
// X = Phi*X + C
// P = Phi*P*Phi^T + Q
m_x = Phi*m_x+C;
m_p = Phi*P*Transpose(Phi) + Q;
}

// <summary>
// Update the state thanks to the realized measurement Y
// </summary>
public void Update( Point1D Y )
{
// Update:
// I = Y-(HX+D) Called the innovation= measurement – state transformed by H.
m_I= Y(t)-(H*m_X+D);
// S = H*P*H^T + R S= Residual covariance = covariance transformed by H + R
m_S= H*m_P*Transpose(T) + R;
// K = P * H^T *S^-1 K = Kalman gain = variance / residual covariance.
m_K= m_P * Transpose(H) * Inverse(S);
// X = X + K*I Update with gain the new measurement
m_X += m_K*Y;
// P = (I – K * H) * P Update covariance to this time.
m_P= (I – K * H) * P;
}

10
4 TRADING STRATEGIES WITH KALMAN FILTER

4.1 Basic concepts

KF model enables us various things:

- It smoothens any data. Hence the data produced by the KF can be used instead of prices

to remove any spike. This opens multiple options as these inputs can be used in cross over

moving averages strategies, MACD indicator, oscillators and combination of these. We do

not explore this as the paper goal is to study the predictive power of KF models.

-it can be used as a predictive tool to help deciding when entering long or short strategies.

We compare the prediction with the current. This is precisely the subject of this paper.

4.2 Pseudo code

/// <summary>
/// Called on each new bar event
/// </summary>
protected override void OnNewBar()
{
if (KalmanFilter(Param1,..,ParamN).Predict[0] > Close[1]+Offset)
EnterLong();
else if (KalmanFilter(Param1,..,ParamN).Predict[0] < Close [1]-Offset)
EnterShort();}
}

11
 RESULTS

5 NUMERICAL RESULTS

5.1 Description of the sample set

To test the efficiency of KF models 1, 2, 3 and 4, we use the E-mini-S&P-500 continuation

Future, whose RIC is Esc1. We use the Eikon App “Trading Robot” that has been

developed by the author. We look at daily data between 28 Feb 2015 and 28 Feb 2016.

5.2 Comparison of Kalman filters with standard technical indicators

We provide graphics of various indicators to measure how KFs best fit price information.

We display:

- some standard technical analysis indicators:

 Moving averages with lag: standard and exponential moving average with

12 days period

 Moving averages with zero lag: double exponential moving average with 12

days period as (EQ.2.9) and triple exponential moving average with 12 days

period as (EQ.2.10)

- the different KF indicators, KF model 1, 2, 3 and 4.

In Figure 1, we see that the KF model 1 sticks much better to price data than any of the two

moving averages. This is normal as KF model has 0 to 1 period lag. We do not show in this

graphic the other KF models as they would be barely distinguishable. In Figure 2 and Figure

3, we compare KF model with zero-lag moving averages like DEMA or TEMA. We

emphasize area of difference with orange circles and see that KF models stick much better

12
to price data. In Figure 4, we compare the different KF models and see that KF model 1&2

are similar while models 3&4 are also similar, with an advantage to the latter ones.

Figure 1: Comparison of Kalman filter with classical moving averages.

The red line representing the KF model 1 sticks much better to the price data than any of the two moving averages
(standard and exponential ones with both 12 periods).

Figure 2: Comparison of Kalman filter with double and triple exponential moving averages
Red lines representing the KF model 1 stick much better to the price data than DEMA or TEMA as displayed in orange
circles.

13
 Figure 3: Zoom on differences between Kalman filter and zero lag moving averages.
KF model 1 reacts faster to price changes as emphasized by orange circles.

Figure 4: Comparison between the different Kalman filters.

Within the KF model family, model 3 and 4 are even better than model 1 and 2. Model 1 and 2 (respectively model 3
and 4) have similar behaviors.

14
5.3 Kalman filter trading strategies performance

We look at the same one-year of data and compute the optimal parameters for the 4 KF
models. For each model, we use no leverage, trade only one future contract regardless of
the current trading account. We also assume a 4 dollar round trip commission, which is the
observed price at retailed brokers like Interactive-Brokers. For large trader with more than
20,000 contracts per months and CME membership, round trip commission lowers to 1.4
USD.
Table 1 shows that the best model is model 4 with an annual net profit of 39.5 k USD,
followed by model 3 with 29 k USD and the last two being model 2 and 1 with net profit
of 22 and 19 k USD.
We can make various remarks:
 The final model ranking makes senses as model 4 is richer than 3 that is itself richer
than 2 that is richer than 1.
 The best model, KF 4, provides a nice net profit, 39 k with a max drawdown of -
2,600, hence representing a ratio of net profit over drawdown (also called recovery
ratio) of 15. This is excellent!
 E-mini-S&P daily margin is about 5 to 6 k USD, hence 40k USD net profit is an
amazing statistic. In addition, model 4 incurs only positive monthly PnL (Figure 5).
 KF model 3 has a nice and steady cumulative profit curve (Figure 6), while model
4 outperforms because it captures a few large additional trades (Figure 5 and Figure
9).
 KF model 1 & 2 are Kalman filter models already explored in literature. We find
some negative monthly PnL and large drawdown (see Figure 7 and Figure 8). This
is a known feature as these models have a poor dynamic. This may explain why
these standard KF models have been disregarded.
 The difference between KF model 3 and 4 is the oscillator factor. This confirms the
well-known fact that oscillators capture others features than trending indicators and
catch any mean reverting market (in trading range environment). The combination
of trend following factors (like in model 3) with the new extra term inspired from
oscillators yields a powerful model called 4. We can notice that parameter 14, 𝑁2 ,
is null. It indicates that the oscillator factor plays a role only on short term factor.
This can be interpreted as an empirical evidence that range trading has only
influence on short term while trend dominates in long term.
 The parameters 11 and 13 in KF model 4 represent the neutrality level at which the
oscillator factors changes from bullish to bearish. It is amazing that its optimal value
turns out to be 50%, which is also a well-known feature of oscillators where the
level of neutrality is 50%
We provide optimal parameters in Table 2. We also provide various statistics for KF model
4, 3, 2 and 1 (starting with the best model and going to the worst) in Table 3, Table 4, Table
5, Table 6 and the list of all trades in Table 7

15
We provide in Figure 5, Figure 6,Figure 7, Figure 8 the cumulative profit and loss curve
for trading strategy of model 4, 3, 2 and 1, starting by the best one. Figure 9 zooms on the
period where model 4 locks in large profit due to accurate prediction of turning points.

Model Net Profit Gross Profit Gross Loss Drawdown Trades Commission Recovery ratio Sharpe Ratio
Kalman Filter 1 18,755 39,151 -20,396 -7,348 55 220 2.55 0.72
Kalman Filter 2 22,380 40,747 -18,367 -7,348 55 220 3.05 0.76
Kalman Filter 3 29,022 47,548 -18,526 -3,800 57 228 7.64 1.22
Kalman Filter 4 39,558 50,243 -10,685 -2,600 48 192 15.21 0.73
Table 1: Trading performance of Kalman filter model 1, 2, 3 and 4.

Model Kalman Filter 1 Kalman Filter 2 Kalman Filter 3 Kalman Filter 4

Parameter 1 5.00 5.00 1.00 1.00
Parameter 2 5.00 5.00 0.40 0.40
Parameter 3 45.00 41.00 1.20 1.20
Parameter 4 10.00 1.00 1.00 1.00
Parameter 5 1.00 1.00 1.00
Parameter 6 0.80 0.80
Parameter 7 0.40 0.40
Parameter 8 0.70 0.70
Parameter 9 1.00 1.00
Parameter 10 0.40 0.40
Parameter 11 0.50
Parameter 12 0.90
Parameter 13 0.50
Parameter 14 -
Parameter 15 5.00

Table 2: Model parameters for Kalman filter model 1, 2, 3 and 4

16
 Kalman Filer Model 4
Field All Long Short
Net Profit (A+B) 39,558 17,279 22,279
Gross Profit (A) 50,243 20,957 29,286
Gross Loss (B) (10,685) (3,678) (7,007)
Total Commission 192 96 96
Drawdown (2,600) (2,137) (2,520)
Sharpe Ratio 0.73 0.84 0.55
Profit Factor (A/B) 4.70 5.70 4.18

Number of Trades 48 24 24
Winning Trades 30 17 13
Average Trade Profit 824 720 928
Average Winning Trade 1,675 1,233 2,253
Largest Winning Trade 11,309 5,434 11,309
Max. conseq. Winners 6 5 3

Losing Trades 18 7 11
Average Losing Trade (594) (525) (637)
Largest Losing Trade (1,729) (1,729) (1,267)
Max. conseq. Losers 4 2 3

Ratio avg. Win / avg. Loss 2.82 2.35 3.54

Winning/Total 0.63 0.71 0.54

Avg. Time in Market 6.92 days 3.88 days 9.96 days

Profit per Month 3,623 1,583 2,047
Max. Time to Recover 58 days 56 days 92 days
Table 3: Trading strategy statistics for Kalman filter model 4

Figure 5: Cumulative profit and monthly PnL distribution for KF model 4

17
 Kalman Filer Model 3

Field All Long Short

Net Profit (A+B) 29,022 12,009 17,013
Gross Profit (A) 47,548 24,403 23,145
Gross Loss (B) (18,526) (12,394) (6,132)
Total Commission 228 116 112
Drawdown (3,820) (3,366) (1,579)
Sharpe Ratio 1.22 0.38 1.65
Profit Factor (A/B) 2.57 1.97 3.77

Number of Trades 57 29 28
Winning Trades 35 18 17
Average Trade Profit 509 414 608
Average Winning Trade 1,359 1,356 1,361
Largest Winning Trade 5,234 5,234 4,271
Max. conseq. Winners 14 7 7

Losing Trades 22 11 11
Average Losing Trade (842) (1,127) (557)
Largest Losing Trade (2,879) (2,879) (1,579)
Max. conseq. Losers 5 3 3

Ratio avg. Win / avg. Loss 1.61 1.20 2.44

Winning/Total 0.61 0.62 0.61

Avg. Time in Market 5.82 days 6.45 days 5.18 days

Profit per Month 2,658 1,100 1,860
Max. Time to Recover 53 days 64 days 71 days
Table 4: Trading strategy statistics for Kalman filter model 3

Figure 6: Cumulative profit and monthly PnL distribution for KF model 3

18
 Kalman Filer Model 2
Field All Long Short
Total Net Profit 22,380 8,692 13,688
Gross Profit 40,747 19,611 21,136
Gross Loss (18,367) (10,919) (7,448)
Commission 220 108 112
Drawdown (7,348) (6,628) (2,283)
Sharpe Ratio 0.76 0.32 0.55
Profit Factor 2.22 1.80 2.84

Number of Trades 55 27 28
Winning Trades 32 16 16
Average Trade Profit 407 322 489
Average Winning Trade 1,273 1,226 1,321
Largest Winning Trade 6,521 3,221 6,521
Max. conseq. Winners 4 6 3

Losing Trades 23 11 12
Average Losing Trade (799) (993) (621)
Largest Losing Trade (3,679) (3,679) (1,717)
Max. conseq. Losers 4 5 2

Ratio avg. Win / avg. Loss 1.59 1.23 2.13

Winning/Total 0.58 0.59 0.57

Avg. Time in Market 6.04 days 6.41 days 5.68 days

Profit per Month 2,050 801 1,254
Max. Time to Recover 132 days 146 days 70 days

Table 5: Trading strategy statistics for Kalman filter model 2

Figure 7: Cumulative profit and monthly PnL distribution for KF model 2

19
 Kalman Filer Model 1
Field All Long Short
Total Net Profit 18,755 6,880 11,876
Gross Profit 39,151 18,861 20,290
Gross Loss (20,396) (11,982) (8,415)
Commission 220 108 112
Drawdown (7,348) (6,628) (2,283)
Sharpe Ratio 0.72 0.26 0.48
Profit Factor 1.92 1.57 2.41

Number of Trades 55 27 28
Winning Trades 31 16 15
Average Trade Profit 341 255 424
Average Winning Trade 1,263 1,179 1,353
Largest Winning Trade 6,521 3,221 6,521
Max. conseq. Winners 4 6 2

Losing Trades 24 11 13
Average Losing Trade (850) (1,089) (647)
Largest Losing Trade (3,679) (3,679) (1,717)
Max. conseq. Losers 4 5 2

Ratio avg. Win / avg. Loss 1.49 1.08 2.09

Winning/Total 0.56 0.59 0.54

Avg. Time in Market 6.04 days 6.45 days 5.64 days

Profit per Month 1,718 634 1,088
Max. Time to Recover 132 days 146 days 71 days

Table 6: Trading strategy statistics for Kalman filter model 1

Figure 8: Cumulative profit and monthly PnL distribution for KF model 1

20
Trade Direction Entry date Entry price Exit date Exit price Profit PnL Commission Days in position
1 Long Mar-31-15 2,043 Apr-01-15 2033.5 (454.0) (454.0) 4 2
2 Short Apr-01-15 2,034 Apr-02-15 2026 371.0 (83.0) 4 2
3 Long Apr-02-15 2,026 Apr-08-15 2045.25 958.5 875.5 4 4
4 Short Apr-08-15 2,045 Apr-09-15 2047.5 (116.5) 759.0 4 2
5 Long Apr-09-15 2,048 Apr-10-15 2061.5 696.0 1,455.0 4 2
6 Short Apr-10-15 2,062 Apr-21-15 2076.25 (741.5) 713.5 4 8
7 Long Apr-21-15 2,076 Apr-22-15 2069.75 (329.0) 384.5 4 2
8 Short Apr-22-15 2,070 May-04-15 2081.75 (604.0) (219.5) 4 9
9 Long May-04-15 2,082 May-05-15 2079.75 (104.0) (323.5) 4 2
10 Short May-05-15 2,080 May-07-15 2048.25 1,571.0 1,247.5 4 3
11 Long May-07-15 2,048 May-11-15 2084.75 1,821.0 3,068.5 4 3
12 Short May-11-15 2,085 Jun-10-15 2062.25 1,121.0 4,189.5 4 23
13 Long Jun-10-15 2,062 Jun-11-15 2083.5 1,058.5 5,248.0 4 2
14 Short Jun-11-15 2,084 Jul-01-15 2054.25 1,458.5 6,706.5 4 15
15 Long Jul-01-15 2,054 Jul-03-15 2049.75 (229.0) 6,477.5 4 3
16 Short Jul-03-15 2,050 Jul-10-15 2050.5 (41.5) 6,436.0 4 6
17 Long Jul-10-15 2,051 Jul-14-15 2074.5 1,196.0 7,632.0 4 3
18 Short Jul-14-15 2,075 Jul-29-15 2071 171.0 7,803.0 4 12
19 Long Jul-29-15 2,071 Jul-30-15 2077.5 321.0 8,124.0 4 2
20 Short Jul-30-15 2,078 Aug-24-15 1851.25 11,308.5 19,432.5 4 18
21 Long Aug-24-15 1,851 Aug-28-15 1960 5,433.5 24,866.0 4 5
22 Short Aug-28-15 1,960 Sep-02-15 1921.25 1,933.5 26,799.5 4 4
23 Long Sep-02-15 1,921 Sep-10-15 1920.25 (54.0) 26,745.5 4 7
24 Short Sep-10-15 1,920 Sep-11-15 1928.25 (404.0) 26,341.5 4 2
25 Long Sep-11-15 1,928 Sep-17-15 1974.75 2,321.0 28,662.5 4 5
26 Short Sep-17-15 1,975 Sep-21-15 1948.5 1,308.5 29,971.0 4 3
27 Long Sep-21-15 1,949 Oct-06-15 1967.5 946.0 30,917.0 4 12
28 Short Oct-06-15 1,968 Oct-15-15 1986.25 (941.5) 29,975.5 4 8
29 Long Oct-15-15 1,986 Oct-19-15 2010.25 1,196.0 31,171.5 4 3
30 Short Oct-19-15 2,010 Dec-15-15 2031 (1,041.5) 30,130.0 4 42
31 Long Dec-15-15 2,031 Dec-16-15 2048.25 858.5 30,988.5 4 2
32 Short Dec-16-15 2,048 Dec-22-15 2022.75 1,271.0 32,259.5 4 5
33 Long Dec-22-15 2,023 Dec-23-15 2043 1,008.5 33,268.0 4 2
34 Short Dec-23-15 2,043 Jan-11-16 1924.5 5,921.0 39,189.0 4 12
35 Long Jan-11-16 1,925 Jan-14-16 1890 (1,729.0) 37,460.0 4 4
36 Short Jan-14-16 1,890 Jan-15-16 1862 1,396.0 38,856.0 4 2
37 Long Jan-15-16 1,862 Jan-18-16 1869.5 371.0 39,227.0 4 2
38 Short Jan-18-16 1,870 Jan-19-16 1894 (1,229.0) 37,998.0 4 2
39 Long Jan-19-16 1,894 Jan-26-16 1878.5 (779.0) 37,219.0 4 6
40 Short Jan-26-16 1,879 Jan-27-16 1890.25 (591.5) 36,627.5 4 2
41 Long Jan-27-16 1,890 Jan-28-16 1894 183.5 36,811.0 4 2
42 Short Jan-28-16 1,894 Jan-29-16 1894.5 (29.0) 36,782.0 4 2
43 Long Jan-29-16 1,895 Feb-02-16 1913.5 946.0 37,728.0 4 3
44 Short Feb-02-16 1,914 Feb-04-16 1901.5 596.0 38,324.0 4 3
45 Long Feb-04-16 1,902 Feb-17-16 1906 221.0 38,545.0 4 10
46 Short Feb-17-16 1,906 Feb-25-16 1931.25 (1,266.5) 37,278.5 4 7
47 Long Feb-25-16 1,931 Feb-26-16 1959.75 1,421.0 38,699.5 4 2
48 Short Feb-26-16 1,960 Feb-26-16 1942.5 858.5 39,558.0 4 1

Table 7: Trades list for Kalman model 4

21
Figure 9: Efficiency of Kalman filter model 4 to detect trends

22
 6 DISCUSSION

Parameters for the Kalman filter models are obtained by a general optimization. Hence they
provide the best possible choice of parameters. Results presented here should be analyzed
with this in mind.
We clearly see that model 1 and 2 provides similar results about 20 k of net profit for one
year trading the E-mini contract. When adding the new feature of a short and long term
model factor, we increase net profit to 29 k, which is substantial. We reduce maximum
drawdown from -7,300 USD to -3,800 USD. This is a material gain. Model 4 performs even
better as we generate an additional 10 k with net profit skyrocketing to 40 k USD with an
further reduction of drawdown to -2,600 USD.

23
 CONCLUSION

In this paper, we empirically validate that Kalman filters with meaningful dynamics have
predictive power. After reviewing moving averages and the general equation for its lag at
order n with respect to the one at first order, we examine 4 Kalman filter models: the
common one with speed and acceleration concepts, the traditional statistical one referred
to as the local linear trend, a new model that splits price contribution between short and
long term effect and a last one that encompasses all above with an additional term
corresponding to the position of the price with regards to its extremums. We find
empirically that model 4 performs far better than any other models. We also confirm that
KF models have zero lag and capture price dynamic better than previous combinations of
moving averages like DEMA or TEMA. We confirm on model 4 that oscillators and trend
following indicators are a powerful combination that performs better than any single
indicators.

24
 REFERENCES

[Bruder-Dao-Richard-Roncalli-2011] Benjamin Bruder, Tung-Lam Dao, Jean Charles

Richard, Thierry Roncalli, Trend Filtering Methods for Momentum Strategies; SSRN paper.
http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2289097
[Cazalet-Zheng-2014] Zelia Cazalet, Ban Zheng, Hedge Funds in Strategic Asset Allocation,
White Paper #11 http://www.thehedgefundjournal.com/sites/default/files/hf-strategic-asset-
allocation-lyxor-mar-2014.pdf
[Chan-2013] Ernie Chan, Algorithmic Trading: Winning Strategies and Their Rationale,
Wiley 2013 http://www.amazon.com/Algorithmic-Trading-Winning-Strategies-
Rationale/dp/1118460146
[Dao-2011] Tung-Lam Dao, Momentum Strategies: From Novel Estimation Techniques to
Financial Applications, SSRN White Paper
http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2358988
[Dürschner-2012] Manfred G. Dürschner, Moving Averages 3.0, IFTA Journal 2012,
http://ifta.org/public/files/journal/d_ifta_journal_12.pdf
[Ehlers-2001a] John F. Ehlers: Signal Analysis Concepts (Internet, 2001),
http://www.technicalanalysis.org.uk/moving-averages/Ehle.pdf
[Ehlers-2001b] John F. Ehlers: Zero Lag, MesaSoftware Technical Paper 2001
http://www.mesasoftware.com/papers/ZeroLag.pdf
[Haleh-et-al-2011] H. Haleh, B. Akbari Moghaddam, and S. Ebrahimijam, A New Approach
to Forecasting Stock Price with EKF Data Fusion, International Journal of Trade, Economics
and Finance, Vol. 2, No. 2, April 2011
[Kalman-1960] Kalman, R. E. A New Approach to Linear Filtering and Prediction Problems".
Journal of Basic Engineering 82: 35, 1960.
[Kalman-Filter-Wikipedia] https://en.wikipedia.org/wiki/Kalman_filter
[Martinelli-2006] Martinelli, Rick Harnessing The (Mis)Behavior Of Markets, Technical
Analysis of Stocks & Commodities, Volume 24: June, 2006.
[Martinelli-Rhoads-2010] Martinelli, Rick, and Neil Rhoads, Predicting Market Data Using
The Kalman Filter, part 1, Technical Analysis of Stocks & Commodities, Volume 28: January
2010
[Mulloy-1994] Patrick Mulloy: Smoothing Data With Faster Moving Averages, Stocks &
Commodities Magazine (February 1994)

25
 APPENDIX

Proof A.1:
We want to prove
𝐿𝐴𝐺 𝑀𝐴⨀𝑘 (𝑤0 , … , 𝑤𝑁 ) = 𝑘 𝐿𝐴𝐺 𝑀𝐴(𝑤0 , … , 𝑤𝑁 ) (EQ.2.7)

For k=1, this result is obviously true. If the result is true at order 𝑘 − 1, let us show that it
holds at order 𝑘. We want to compute the following lag

∑𝑁𝑖1 =0,…,𝑖𝑘 =0 𝑤𝑖1 …𝑤𝑖𝑘 𝑇𝑖1 +⋯+𝑖𝑘

𝐿𝐴𝐺 𝑀𝐴⨀𝑘 (𝑤0 , … , 𝑤𝑁 ) = (A.1.1)
∑𝑁𝑖1 =0,…,𝑖𝑘=0 𝑤𝑖1 …𝑤𝑖𝑘

Which can be reformulated as:

∑𝑁𝑖1 =0,…,𝑖𝑘−1 =0 𝑤𝑖1 …𝑤𝑖𝑘−1 ∑𝑁𝑖𝑘 =0 𝑤𝑖𝑘 (𝑇𝑖1 +⋯+𝑖𝑘−1 +𝑇𝑖𝑘 )

(A.1.2)
∑𝑁𝑖1 =0,…,𝑖𝑘−1 =0 𝑤𝑖1 …𝑤𝑖𝑘−1 ∑𝑁𝑖𝑘 =0 𝑤𝑖𝑘

∑𝑁𝑖1 =0,…,𝑖𝑘−1 =0 𝑤𝑖1 …𝑤𝑖𝑘−1 𝑇𝑖1 +⋯+𝑖𝑘−1 ∑𝑁𝑖𝑘 =0 𝑤𝑖 𝑇𝑖𝑘

Or
∑𝑁𝑖1 =0,…,𝑖𝑘−1 =0 𝑤𝑖1 …𝑤𝑖𝑘−1
+ 𝑁
𝑘
∑𝑖𝑘 =0 𝑤𝑖𝑘
(A.1.3)

Or = (𝑘 − 1) 𝐿𝐴𝐺 𝑀𝐴(𝑤0 , … , 𝑤𝑁 ) + 𝐿𝐴𝐺 𝑀𝐴(𝑤0 , … , 𝑤𝑁 ) (A.1.4)

Which proves the result!

Proof A.2:
If we denote by a and b the coefficient of the linear combination between the moving
average and the second order recursive moving average, the resulting formula should be
𝑎𝑀𝐴 + 𝑏𝑀𝐴⨀2 (A.2.1)
With the constraints that first, the sum of the coefficients should be equal to 1, hence
𝑎+𝑏 =1 (A.2.2)
Second, the lag of the resulting combination should be 0, hence
𝑎 + 2𝑏 = 0 (A.2.3)
It is easy to solve this linear system and find

26
 𝑎 = 2, 𝑏 = −1 (A.2.4)
Which is the result!
Proof A.3:
If we denote by a, b and c the coefficient of the linear combination between the moving
average, the second and the third order recursive moving average, the resulting formula
should be
𝑎𝑀𝐴 + 𝑏𝑀𝐴⨀2 + 𝑐𝑏𝑀𝐴⨀3 (A.3.1)
With the following constraints. First, the sum of the coefficients should be equal to 1, hence
𝑎+𝑏+𝑐 =1 (A.3.2)
Second, the lag of the resulting combination should be 0, hence
𝑎 + 2𝑏 + 3𝑐 = 0 (A.3.3)
Third, we impose that the coefficient for the third order recursive moving average is equal
to 1,
𝑐=1 (A.3.4)
It is easy to solve this linear system and find
𝑎 = 3, 𝑏 = −3, 𝑐 = 1 (A.3.5)
Which is the result!

27