Stock Price Prediction Using Dynamic Mode

Decomposition
Deepthi Praveenlal Kuttichira,Gopalakrishnan E.A,Vijay Krishna Menon, Soman K.P
Centre for Computational Engineering and Networking (CEN),
Amrita School of Engineering,
Amrita Vishwa Vidyapeetham, Amrita University,Coimbatore, India
Email:deepthikuttichira93@[Link]

Abstract—Stock price prediction is a challenging problem as support technical analysis [3] [4] [5] [6]. Technical analysis
the market is quite unpredictable. We propose a method for price uses statistical analysis to predict the future price. With the
prediction using Dynamic Mode Decomposition assuming stock advent of automated trading, technical analysis is gaining
market as a dynamic system. DMD is an equation free, data-
driven, spatio-temporal algorithm which decomposes a system to further popularity. In technical analysis time series analysis
modes that have predetermined temporal behaviour associated and neural networks are most widely used [Link] (Au-
with them. These modes help us determine how the system toregressive),MA (Moving Average),ARMA (Autoregressive
evolves and the future state of the system can be predicted. We Moving Average),ARIMA(Autoregressive Integrated Moving
have used these modes for the predictive assessment of the stock Average) are some of the models used for time series analysis
market. We worked with the time series data of the companies
listed in National Stock Exchange. The granularity of time was [7] [8].In AR the future values is determined using past
minute. We have sampled a few companies across sectors listed values,MA uses past error values to determine future values
in National Stock Exchange and used the minute-wise stock price and ARMA uses both for [Link] dynamics of
to predict their price in next few minutes. The obtained price the financial system adds to the non-stationarity of the data
prediction results were compared with actual stock prices. We which makes long term predictions inaccurate. Also time series
used Mean Absolute Percentage Error to calculate the deviation
of predicted price from actual price for each company. Price analysis is computationally costly. As neural networks are
prediction for each company was made in three different ways. efficient in pattern learning, these models can be used to learn
In the first, we sampled companies belonging to the same sector patterns in time series data. Predictions using neural networks
to predict the future price. In the latter, we considered sampled are also widely done in stock market [9].One major drawback
companies from all sectors for prediction. In the first and second of neural network is its requirement of large amount of training
method, the sampling as well as the prediction window size were
fixed. In the third method the sampling of companies was done data. Each time there is a pattern change in financial market,
from all sectors considered. The sampling window was kept the model has to be trained for new parameters [10] [11]. This
fixed, but predictions were made until it crossed a threshold pattern change can happen quite often in financial market.
error. Prediction was found to be more accurate when samples In our proposed method we assume stock market to be a
were taken from all the sectors, than from a single sector. When dynamical system and use DMD (Dynamic Mode Decom-
sampling window alone was fixed; the predictions could be made
for longer period for certain instances of sampling. position), a data driven, spatial-temporal coherent algorithm
to identify the evolutionary patterns of this system [12] [13].
Index Terms—Dynamic Mode Decomposition,Proper Orthog- DMD is computationally very efficient as it exploits the low-
onal Decomposition, Mean Absolute Percentage Error dimensional structure of the data. Due to this advantageous
property of DMD, it finds its application in many fields. A
I. I NTRODUCTION dynamical system can be decomposed to modes and these
Stock price prediction is a challenging problem as the modes can be used to identify the evolutionary patterns of the
market is quite unpredictable. Arguments prevail,regarding system [14] [15]. Few decomposition methods are, wavelet
the possibility of stock price prediction. Efficient market decomposition, fourier decomposition,proper orthogonal de-
hypothesis states that the current price reflects the current state composition. Wavelet modes has been used to identify business
of the market and nothing further can be inferred from it [1] cycles that China shared with the rest of the world economy
[2]. Fundamental analysts believe that stock price predictions [16]. However the system had to be decomposed into large
can be made by analyzing the current financial situations and number of wavelet modes. Using DFT we cannot estimate
by taking into account the financial statements of a company. the growth and decay of the system, as DFT eigen values
Technical analysis involves studying the past data to predict are always 1. The difference between POD and DMD is that
future prices. Though there are a number of studies supporting POD modes are orthogonal in space with multi-frequency time
efficient market hypothesis, recently there are studies that signals, whereas DMD modes are non-orthogonal in space
with single frequency time signal [14].DMD decomposes a
system to a set of modes that have prescribed time dynamics

978-1-5090-6367-3/17/$31.00 ©2017 IEEE 55

[12] [14] [15] [13]. Since DMD modes are dynamic modes, Hereψk andωk are the eigenvectors and eigenvalues of matrix
they capture the trend of market in them. Dynamic systems A. If real part of eigenvalues are positive and greater than
are usually described using equation that describe how their one, then it means a growing mode and growing money. If
trajectories evolve in finite dimensional space. An alternative eigenvalues are negative, then it means decaying modes and
description of this can be made using an operator that acts losing money. The data we used was the minute-wise data
on infinite dimensional space of [Link] operator of transactions from 1 July 2014 to 30 June [Link] data
is one such infinite dimensional,linear operator [17] [18].The was structured as date,transaction id,time,company name,price
evolution of a dynamic system from time t to t+1 can be of the stock and volume of the stock. For our analy-
captured using koopman operator. This koopman operator sis,we selected 57 companies across sectors IT,financial ser-
captures the dynamics of the system. DMD decomposes this vices,pharma,automobiles. These 4 sectors combined hold
matrix into modes. These are the dynamic modes of the 63.94 %market [Link] sampling interval is one [Link]
system. DMD eigen values are not always 1 unlike that of all companies have transactions in every minute. In such cases
DFT. This means that DMD can estimate the rate of growth we substitute the previous stock price in the current minute.
or decay of the system [15]. In the context of analyzing stock The reasoning behind this is that, even though no transaction
market, DMD modes can be thought of as coherent structures was done in that minute,the stock price at that time is the
in the financial activity. DMD has been used in finance to previously transacted [Link] data matrix given for DMD
extract cyclic nature in market as well as for price prediction algorithm is snapshots of the system taken at equispaced
[19]. In our work, we use DMD for short term prediction of time [Link] snapshot consists of the stock price of all
price. the companies considered at that particular time [Link]
DMD framework and proposed method is detailed in sec- decomposes data matrix to modes that spans spatially and
tion[II].Section[III] contains Results. In Section [IV] the ob- has temporal frequency associated with them [12].The data
tained results are discussed in detail. Section[V] draws out the matrix will be an n×m matrix where n=number of companies
conclusions and discusses the scope for future work. considered m=number of snapshots taken The matrix will be
of the form,
II. M ETHODOLOGY
X = [x1 , x2 , x3 , . . . xm ] (6)
A dynamic system is defined using a set of governing
[Link] consider financial system as a non-linear dy- From this matrix we need to capture the underlying dynamics
namic system whose governing equations are not known to of the [Link] this we splits the data matrix into,
[Link] snapshots we take corresponds to a state of the system.
X1 = [x1 , x2 , x3 , . . . xm−1 ] (7)
Snapshot of a system consists of observed measurements of
that system at time t. In our case the observed measurements X2 = [x2 , x3 , x4 , . . . xm ] (8)
are the stock price of each company at the time t. A dynamic
system is described using a governing set of differential X2 is one time slot shifted from X1 .Let A be a linear
equations. operator that maps Xi to xi+1 .Then X2 can be expressed as
dx
= F (x, t) (1) X2 = [Ax1 , Ax2 , Ax3 , . . . Axm−1 ] (9)
dt
At each state we can make different kinds of measurements X2 = AX1 (10)
of the [Link] measurement function can be denoted
DMD expects X1 to have low-rank structure. If not DMD fails
as
immediately. SinceN ≫ M ,the matrix X1 is sure to have a
G(x, tk ) = 0 (2)
low rank [Link] some point addition of a new snapshot
where k = 1, 2, . . . , M where M is measurement [Link] to matrix X1 will not add on to the vector space spanned by
initial condition is stated as X1 . We can express xm as combination of previous columns
of X1 .
x(0) = x0 (3) m−1
X
xm = ai x i + r (11)
The non-linear function F that defines the set of governing
i=1
equations is unknown. All we have is the initial conditions
and the measurements [Link] our case the measurement is the Here r is the residual [Link] algorithm does the mini-
stock price. In DMD procedure, we construct an approximate mization of r.
linear evolution of the system. X2 = X1 S + re∗m−1 (12)
dx̃ Here S is the companion [Link]−1 is (m-1)th unit [Link]
= Ax̃ (4)
dt can be seen that
The well-known solution for equation(4) is, AX1 = X2 ≈ X1 S Eigen values of A are approximately
K
that of [Link] utilize the low rank structure of X1 we apply
X SVD to [Link] Value Decomposition decomposes X1 to
x̃(t) = bk ψk exp(ωk t) (5)
k=1
U ΣV ∗ where U ∈ C n×r ,

56
 Σ ∈ C r×r and that the predicted result using second approach was more
V ∈ C (m−1)×r .Here Σ is diagonal matrix,whereas U accurate. In the third method sampling window size was kept
and V are unitary [Link] X1 using SVD and fixed at 6,but prediction was made until it crossed a threshold
substituting it,we get error of 3.5%. It was observed that for certain time instances,
the prediction could be made for longer time periods. But for
X2 ≈ U ΣV ∗ S (13) other time instances, the prediction could be made only for
S ≈ V Σ−1 U ∗ X2 (14) shorter time period than that of fixed window approach. MAPE
was used on the results obtained by all the three approaches.
For robust implementation of the algorithm, instead of finding MAPE values for second and third method was significantly
the eigen values and eigen vectors of S,we find the same for lower than the first method. The comparison between the
a similar matrix S̃ results obtained by DMD and ARIMA for a pharma company
S̃ ≈ U ∗ X2 V Σ−1 (15) CIPLA is shown in table 1. The ARIMA model that fit the
data was ARIMA(1,1,0).
The eigen values of similar matrices are [Link] vector
can be found using similar matrix [Link] eigen values
are,
INDOCO
280
S̃vj = λj vj (16) actual price
predicted price 20
DMD eigen vectors are, 260

Price(Rupee)
predicted price 57
φj = U vj (17)
240
For ease of representation DMD modes can be converted
to fourier modes as ωj = ln(λj ).So now using these eigen
220
vectors and eigen values we can reconstruct the system as
follows
Xr 200
XDM D (t) = bj φj eωj t = φdiag(eωt )b (18)
j=1 180
XDM D (t) defines the state of the system at time [Link] our
0 20 40 60 80 100 120 140 160 180 200
case,as we had taken stock prices to represent the state of Time(Minute)
the system,XDM D (t) gives stock prices at time [Link] (10)is
Fig. 1: Plot between the actual, predicted values using DMD
basically a regression [Link] tries to find a least square
for 20 companies and DMD for 57 companies for pharma
fit for the points considered. In our method we considered
company INDOCO
6 snapsots of the system to make future predictions. In one
approach the prediction window was kept fixed at the size of
sampling window. In another approach the sampling window
was kept same, but prediction was made until it crossed a WIPRO
threshold value. MAPE was used to validate the result. The 600
actual
predicted values by DMD was also compared with ARIMA
dmd
model. We gave an input of 2000 data points, to fix the model.
Price(Rupee)

ada
For this we used forecasting package in tool R. ARIMA model
is of the form ARIMA(p,d,q). Here p is the lag or AR, it is
550
the number of previous data points on which the current data
point is dependent. The number of times the series had to be
differenced to obtain a stationary signal is denoted by d. The 500
lag of MA is denoted by q.
III. R ESULTS
We approached our problem in three different ways using 450
DMD [Link] the first method we used, the snapshots 0 50 100 150 200 250 300 350 400
consisted of the companies sampled from same [Link] the
second method the snapshots consisted of the companies Time(Minute)
sampled across different sectors. In both these approaches,the Fig. 2: Plot between the actual, predicted values using DMD
size of the sampling window and prediction window was kept and prediction using adaptive DMD for IT company WIPRO.
[Link] sampling window size chosen for our experiment
was 6. On comparison with the actual values, it was observed From fig.1,it is clear that,there exists inter sectoral depen-

57
[Link] captures these inter-sector [Link]
measure was used on the predicted output. MAPE Pharma
n
2
100 X At − Ft
M= (19)
n t=1 At

mape
1
TABLE I: Price prediction for pharma company CIPLA
Arima(1,1,0) DMD Arima DMD
Sl:no Actual
predicted predicted error percent error percent
1 463.2 462.76 463.15 0.09 0.01 0
2 463.3 462.76 463.12 0.11 0.03 0 5 10 15 20
3 463.5 462.76 463.19 0.15 0.066
4
5
463.45
463.2
462.76
462.76
463.19
463.02
0.14
0.09
0.05
0.03
company
(a) MAPE using DMD

IV. D ISCUSSION
MAPE IT
1.5
From figure(1) it is clear that prediction by taking into
account, different companies from across sectors gives better
1

mape
prediction. This means that there is an inter-sector dependency
in stock market. So prediction can be made better by observing
the whole market dynamics than exclusively observing a par- 0.5
ticular stock. From figure(2) it is observed that simple DMD
gives predictions close to actual values,but at times shows
sharp variations. Adaptive DMD was conditioned to mitigate
0
these sharp variations by setting a threshold error. From
1 2 3 4 5 6 7 8 9
figure(3) and (4), it can be observed that the MAPE values for company
second and third method are more or less the same. In adaptive (b) MAPE using DMD
DMD at certain instances, the predictions could be made for
longer period of time than simple DMD, but not so in other MAPE Auto
time periods. This is because, at certain times, a particular 3
market dynamics holds for a longer period. Technical analysis
can be most effectively used at this period of time. It can
mape

2
also be seen that for certain companies simple DMD gives
better results than adaptive DMD. This is because when, more
companies are added, the predictions become better. Since a 1
threshold is set for adaptive DMD, these improvements are
not reflected in adaptive DMD. We set a common threshold 0
for all companies. Lowering the threshold when number of 0 5 10 15
companies are more, would mean that prediction duration
will be cut shorter. This is because if error of any single company
company exceeds threshold, the algorithm again takes actual (c) MAPE using DMD
data to make predictions. It can be seen from table 1 that
DMD predictions are better than ARIMA predictions. Also
MAPE Finance
4
for ARIMA we had to give 2000 data points, for it to fix a
model. The model that fit was ARIMA(1,1,0). Here DMD only
mape

required previous 6 minutes data. This signifies the efficiency

of DMD algorithm. Comparing the fitted ARIMA model with 2
the form ARIMA(p,d,q), we can see that p and d are 1.
This means the current values of CIPLA is dependent on its
previous value and also the series was differenced once, to
obtain stationary signal. 0
0 2 4 6 8 10 12 14
V. C ONCLUSION
company
DMD is an efficient,computationally fast algorithm that
can predict the future states of the system,without any prior (d) MAPE using DMD
knowledge of the underlying equations. Also while using Fig. 3: DMD Result without using adaptive (a) shows the
plot for MAPE values for each company taken in pharma
sector (b) shows the plot for MAPE values for each
company taken in IT sector (c) shows the plot for MAPE
values for each company taken in Automobiles sector (d)
58 shows the plot for MAPE values for each company taken in
Financial Services sector
 adaptive DMD, it was observed that at some instances, the
MAPE Pharma predictions could be made for a longer time period but
4 not so in other instances. This happens because for certain
time periods,the market dynamics doesn’t change. So our
mape

study supports technical analysis, but limited to certain times.

2 In our work, we also observed an interdependence among
sectors. Prediction improves when companies from various
sectors were taken into account. We believe this is because
of the existence of inter-sectoral dynamics. So predictions
0 considering the whole of market dynamics is more accurate
0 5 10 15 20 than predictions made by focusing only on a single stock, or a
company small subset of stocks. Our method fails to predict accurately,
when there is a sudden change in the dynamics,like when
(a) MAPE using adaptive DMD stock split is announced or when a new product is launched
MAPE IT in the market. In short our method doesn’t take care of
1.5 the exogenous influences. Input-output DMD can be used to
take care of exogenous factors [20]. Stock split identification
can be done by identifying variation point. Since DMD is
mape

1 computationally efficient we can apply it for large scale

stock data to obtain profitable results. DMD unlike univariate
0.5 methods like ARIMA can make predictions for more number
of companies at a time. These predictions are made taking into
account the interdependencies between the companies. As this
0 approach produced better result than univariate ARIMA, we
1 2 3 4 5 6 7 8 9 can conclude that considering interdependencies can improve
company prediction. We also observed as more number of companies
(b) MAPE using adaptive DMD
were considered for prediction, the predictions improves. This
implies that considering whole of market dynamics to make
MAPE Auto predictions is better than making predictions by monitoring
3 only a particular stock or a subset of stocks.

R EFERENCES
2
mape

[1] B. G. Malkiel, “The efficient market hypothesis and its critics,” The
Journal of Economic Perspectives, vol. 17, no. 1, pp. 59–82, 2003.
1 [2] A. Timmermann and C. W. Granger, “Efficient market hypothesis and
forecasting,” International Journal of forecasting, vol. 20, no. 1, pp.
15–27, 2004.
0 [3] P. Abinaya, V. S. Kumar, P. Balasubramanian, and V. K. Menon,
0 5 10 15 “Measuring stock price and trading volume causality among nifty50
stocks: The toda yamamoto method,” IEEE, pp. 1886–1890, 2016.
company [4] H. Bessembinder and K. Chan, “The profitability of technical trading
rules in the asian stock markets,” Pacific-Basin Finance Journal, vol. 3,
(c) MAPE using adaptive DMD no. 2, pp. 257–284, 1995.
[5] W. Brock, J. Lakonishok, and B. LeBaron, “Simple technical trading
MAPE Finance rules and the stochastic properties of stock returns,” The Journal of
finance, vol. 47, no. 5, pp. 1731–1764, 1992.
2 [6] K. Matia, Y. Ashkenazy, and H. E. Stanley, “Multifractal properties
of price fluctuations of stocks and commodities,” EPL (Europhysics
Letters), vol. 61, no. 3, p. 422, 2003.
mape

[7] A. A. Ariyo, A. O. Adewumi, and C. K. Ayo, “Stock price prediction

1 using the arima model,” pp. 106–112, 2014.
[8] A. Gunasekarage and D. M. Power, “The profitability of moving average
trading rules in south asian stock markets,” vol. 2, no. 1. Elsevier, 2001,
pp. 17–33.
[9] S. K. Mitra, “Optimal combination of trading rules using neural net-
0 works,” International Business Research, vol. 2, no. 1, p. 86, 2009.
[10] A. Kar, Stock Prediction using Artificial Neural Networks.
0 2 4 6 8 10 12 14 [11] M. Majumder and M. Hussian, “Forecasting of indian stock market
company index using artificial neural network,” Available from: www. nse-india.
com/content/research/FinalPaper206. pdf, 2007.
(d) MAPE using adaptive DMD [12] J. Mann and J. N. Kutz, “Dynamic mode decomposition for financial
trading strategies,” vol. 16, no. 11. Taylor & Francis, 2016, pp. 1643–
Fig. 4: Adaptive DMD Result (a) shows the plot for MAPE 1655.
values for each company taken in pharma sector (b) shows
the plot for MAPE values for each company taken in IT
sector (c) shows the plot for MAPE values for each company
taken in Automobiles sector (d) shows the plot for MAPE
values for each company taken in Financial Services sector 59
[13] L.-x. Cui and W. Long, “Trading strategy based on dynamic mode
decomposition: Tested in chinese stock market,” Physica A: Statistical
Mechanics and its Applications, vol. 461, pp. 498–508, 2016.
[14] P. J. Schmid, “Dynamic mode decomposition of numerical and experi-
mental data,” vol. 656. Cambridge Univ Press, 2010, pp. 5–28.
[15] B. W. Brunton, L. A. Johnson, J. G. Ojemann, and J. N. Kutz, “Extract-
ing spatial–temporal coherent patterns in large-scale neural recordings
using dynamic mode decomposition,” Journal of neuroscience methods,
vol. 258, pp. 1–15, 2016.
[16] J. Poměnková, J. Fidrmuc, and I. Korhonen, “China and the world
economy: Wavelet spectrum analysis of business cycles,” 2014.
[17] A. Mauroy and I. Mezić, “Global stability analysis using the eigen-
functions of the koopman operator,” IEEE Transactions on Automatic
Control, vol. 61, no. 11, pp. 3356–3369, 2016.
[18] M. O. Williams, I. G. Kevrekidis, and C. W. Rowley, “A data–driven
approximation of the koopman operator: Extending dynamic mode
decomposition,” Journal of Nonlinear Science, vol. 25, no. 6, pp. 1307–
1346, 2015.
[19] J.-C. Hua, S. Roy, J. L. McCauley, and G. H. Gunaratne, “Using dynamic
mode decomposition to extract cyclic behavior in the stock market,”
Physica A: Statistical Mechanics and its Applications, vol. 448, pp. 172–
180, 2016.
[20] J. L. Proctor, S. L. Brunton, and J. N. Kutz, “Dynamic mode decom-
position with control,” SIAM Journal on Applied Dynamical Systems,
vol. 15, no. 1, pp. 142–161, 2016.

60