Time Series Analysis of Cross-Listed Stocks

Sanketh Kokkodu Balakrishna and Lopamudra Pal

Department of Computer Science
University of Massachusetts Amherst, MA, 01003
{skokkodubala,lpal}@cs.umass.edu

1 Introduction
Stock Market prediction is a well-known problem in the Financial and Statistical research areas.
This time series forecasting problem is of tremendous interest to researchers because of the profit
potential it presents. The Efficient-Market Hypothesis (EMH) suggests that stock prices reflect all
currently available information and any price change that is not based on newly revealed information
is inherently unpredictable. The various factors which influence the stock market are political
events, concurrent economic conditions, trader’s expectations, changing interest rates, varying foreign
exchange rates and many more.
However, researchers in technical analysis believe that stock prices can be predicted from historical
analysis because of the assumption that stock prices move in trends and that the information which
affect prices enters the market over a finite period, not instantaneously. A lot of interesting work has
been carried out in applying Machine Learning algorithms for analyzing price patterns and predicting
stock prices and index changes.
In this paper, we present applications of various Machine Learning(ML) Regression algorithms along
with their performance for different stocks. Unlike previous work done on predicting stock prices in
a single exchange, the focus of this project is to predict the price of a stock listed on the National
Stock Exchange(NSE) of India based on historical data collected from NSE and the New York Stock
Exchange(NYSE). We hypothesize here that the prices in emerging stock markets like India follow
the prices of the developed economies due to the huge impact of businesses in the US. The reason for
this hypothesis is the observation of the movement of cross-listed stocks on both the markets.

2.5 Normalized Vedanta prices on NYSE and NSE 2.2 Normalized Tata Motors prices on NYSE and NSE
2.0
2.0
1.8
1.5 1.6
1.4
1.0 1.2
1.0
0.5 Vedanta NYSE prices Tata Motors NYSE prices
0.8
Vedanta NSE prices Tata Motors NSE prices
0.0 0.6
2 2 3 3 4 4 5 5 6 6 2 2 3 3 4 4 5 5 6 6
201 201 201 201 201 201 201 201 201 201 201 201 201 201 201 201 201 201 201 201
Jun Dec Jun Dec Jun Dec Jun Dec Jun Dec Jun Dec Jun Dec Jun Dec Jun Dec Jun Dec

Figure 1: Prices of ’Vedanta’ and ’Tata Motors’ on NYSE and NSE

Figure.1 shows the normalized prices of the cross-listed stocks ’Vedanta’ and ’Tata Motors’ on NSE
and NYSE. We observe similar patterns and trends on both the markets. Hence, we picked only
cross-listed stocks which have a large market share in the US and India for this analysis. For this
problem, we collected data from Yahoo Finance for cross listed stocks in different industries like
Banking (ICICI),Automotive (TATA MOTORS) and Energy (VEDANTA). Apart from the inherent
features (Date, Open, High, Low, Volume, Adjusted Close, Close) in the datasets, we have added
additional features which are described in the Data Sets section. We observed differences in the
predicted results, which are described in detail in Section 5.

2 Related Work
Although a lot of work has been done in using Machine Learning to predict stock market prices based
on historical data, there is very little literature about the problem we’re trying to solve. Some related
work done on analyzing stock markets is briefly described below.

2.1 Machine Learning in Prediction

The application of different Machine Learning techniques to predict the stock market is briefly
described in [1]. The paper discusses the application of Support Vector Machines, Linear Regression,
Prediction using Decision Stumps, Expert Weighting, and Online Learning on two stocks – Google
and Yahoo (data collected from Yahoo Finance). It also describes useful parameters and indicators
to recognize patterns in stock prices for successful prediction. The author has added an indicator
function “Exponential Moving Average(EMA)” to the dataset and observed the behavior of the
stocks. From the experiments, the author concludes that using Support Vector Machine combined
with Boosting gives satisfactory results.
A detailed study of how textual analysis from newspapers can be used to predict the stock prices with
a twenty minutes lagging time is shown in [7]. Using a support vector machine (SVM) derivative
specially tailored for discrete numeric prediction and models containing different stock-specific
variables, they show that the model containing both article terms and stock price at the time of article
release had the best performance in closeness to the actual future stock price (MSE 0.04261), the
same direction of price movement as the future price (57.1% directional accuracy) and the highest
return using a simulated trading engine (2.06% return).
[4] does a survey and evaluation of the recent developments in stock market prediction models and the
effects of global events on stock prices. Comparing various prediction models, they found that Neural
Networks(NNs) offer the ability to predict market directions more accurately than other existing
techniques. The ability of NNs to learn nonlinear relationships from the training input/output pairs
enables them to model non-linear dynamic systems such as stock markets more precisely. Using a
good web-mining technique to identify qualitative factors can help improve the prediction accuracy
using Neural Networks.

2.2 Technical Indicators in stock market prediction

[9] and [10] present a variety of technical indicators for stock market prediction and propose a new
modular genetic programming for finding attractive and statistically sound technical patterns for stock
trading. The paper mentions that if technical patterns are complex, they generally match a few cases
which degrades the generality if they are highly profitable. Hence it proposes a genetic algorithm to
determine the most attractive technical indicators for a dataset which focuses on three criteria: high
profitability, simplicity, and frequency within the dataset. [10] gives a brief description of the different
classes of technical indicators studied in stock market prediction problem. Technical indicators are
classified in two wide groups: oscillators and trend followers. The oscillators are designed to vary at
the same time the price varies. They represent a form of the price over a fixed past period, which is
the interval used to calculate the indicator. Trend followers work better when markets are immersed
in a clear trend. They are not efficient in moments of side market, without a defined trend. This paper
centers on the optimization of two indicators, an oscillator- Relative Strength Index (RSI), and a trend
follower- Moving Average of Convergence-Divergence (MACD).

2.3 Analysis of Emerging stock markets

Work done on using Artificial Neural Networks (ANN) to predict Istanbul Stock Exchange (ISE)
market index value is presented in [6]. The features considered to build the model are previous
day’s index value, previous day’s TL/USD exchange rate, previous day’s overnight interest rate and
5 dummy variables each representing the working days of the week. Research showed that fewer
the number of hidden layers in the ANN model, the more accurate the predicted are. The ANN
performance was compared to the moving averages over a specified past period (5 and 10 days in this

2
study and it was found that the former outperformed the latter in moving averages).
[2] analyses the results of applying hybrid ML models based on Genetic Algorithms(GA) and Support
Vector Machines(SVM) for stock market prediction. The research work done in [2] is on Indian
Stocks.
[5] and [8] gives us an overview of planning and visualizing the data and results for our experiments.
[5] shows visualizations of how sensitive the Indian Stock Market is on external factors like Mumbai
Train bombings, economic factors like the GDP growth rate, Capital inflow in sectors (eg. Construc-
tion, Health care, IT). [8] describes a very innovative approach of designing a multisensory human
perceptual tool for the real-world task domain of stock market trading. We hypothesize here that the
Indian market(NSE) is heavily dependent on some of the developed markets such as the New York
Stock Exchange(NYSE).

2.4 Study of co-related stocks

The main idea of the research in [2] is to compute the correlation between stocks and determine
the most correlated stocks. Different technical indicators are calculated for these correlated stocks.
Using the proposed Genetic Algorithm, feature selection is carried out to find the most important
features, which in turn are then given as inputs to Support Vector Machine and the predicted results
are analyzed. The original input features are scaled into the range of [1,1] to independently normalize
each feature component to the specified range. It ensures the larger value input attributes do not
overwhelm smaller value inputs, and thus helps to reduce prediction errors. Neuro-genetic stock
prediction system based on financial correlation between companies is explored in [3].

3 Data Sets

3.1 Data Description

Stocks considered in this paper are restricted to firms listed on both the NYSE and NSE only. The
data is collected from Yahoo finance (https://finance.yahoo.com/) and each dataset is pre-processed
individually. Pre-processing includes removing rows corresponding to days when the stock did not
trade, merging features from NSE and NYSE listings, creating new features and appending the gold
standard values. We consider three stocks of different sectors- ICICI Bank from the Banking sector,
Tata Motors from the Automotive sector and Vedanta Limited from the Energy sector. The date range
considered for the training data is March 23 2012 to January 23 2017. For testing data, the dates
considered are from January 26 2017 to April 26 2017. The total number of training examples after
pre-processing reduces to 1210 for the 5 year period and the testing data to 57 data cases for the three
month period. We believe that this is sufficient test data to accurately measure prediction accuracy, as
the deviation in the data is large enough.

3.2 Feature Description

From the data downloaded directly, each listed stock has has these six features:

1. Open price: The price at which the stock opened on a particular day. This may be dif-
ferent than the close price of the previous day due to after hours trading and difference between
supply and demand.
2. High price: The highest price the stock reached for a particular day. This is typically higher than
the opening and the closing price.
3. Low price: The lowest price the stock reached for a particular day. This is typically lower than
the opening and the closing price.
4. Close price: This is the last traded price for the day. The price may change later to account for
post-trading activities.
‘ 5. Volume: The number of shares traded on a particular day. Indicates the activity and also may
indicate the rise or fall of the price.
6. Adjusted Close: The stock’s closing price that has been amended to include any distributions and
corporate actions that occurred at any time prior to next day’s open.
These features are considered for both NSE and NYSE to yield 12 features. Two additional features
are defined to help model price movement:

3
7. Rolling Mean: Rolling Mean or Simple Moving Average(SMA) is the arithmetic moving average
calculated over k previous days. For the current day, SMA is given as
PN + PN −1 + PN −2 + PN −3 + ... + PN −k
k
where PN ...PN −k are the prices for the previous k days. We have considered k = 5 in our problem.
8. Daily Return: The amount by which a stock has moved from the previous day. It is given by
PN
PN −1 − 1, where PN is the current day price and PN −1 is the previous day closing price.
These additional features are added to both the stock listings to give a total of 16 features. The ypred
value is taken as the following day’s closing price of the stock.

4 Methodology
In this section, the complete pipeline which we have built is described. The pipeline is divided into
sections which are detailed as separate subsections.

4.1 Tools Used

We used a variety of tools and software to build different parts. The complete pre-processing was
done in the ’pandas’ library, which offered large number of methods and flexibility in handling the
data. ’Python’ was used for the complete pipeline and sci-kit learn provided regression methods and
hyper parameter and feature optimization.

4.2 Pre-processing

The data obtained from Yahoo Finance had a lot of missing values and ’nan’ values when the stock
did not trade. We dropped those rows which had incomplete information. The features for the stocks
on the NSE and NYSE were combined to get the data set for one cross-listed stock. On some days, the
stock traded on one exchange but did not on the other exchange. Those rows are dropped for getting
consistent results. The Ypred values i.e the closing price of the following day is also appended to each
feature vector to create a proper format for sk-learn fit and predict methods. Two additional features
are added as described in the Data Sets section. Since only 16 features are used which are core to the
stock’s price movement, no feature engineering or selection is required in the pre-processing stage.

4.3 Machine Learning models

Some of the models which we tried are detailed below. The results of these models are described in
the Results section.
• Linear Regression : Linear Regression is a parametric regression method that assumes the
relationship between y and x is a linear function with parameters w = [w1 , ......wD ]T and b.The
PD
regression function is given as fLin (X) = d=1 wd xd + b. This is our baseline method.
• Decision Tree Regressor : Decision Tree Regressor is a parametric regressor that outputs data
cases using a conjunction of rules organized into a binary tree structure. Each node in the tree
consists of a rule of the form (xd < t) or (xd = t)t. The simplicity of this model is the primary
reason we picked this model.

• Ridge Regression: In regression methods such as least squares linear regression, the parameters
are susceptible to very high variance.To control variance, we might need to regularize the coefficients.
Ridge Regression is a form of regularized least squares when the weights are penalized using the l2
D
2
norm ||w||2 = wT w = wd 2 .
P
d=1
The regularization of weight parameters during learning is achieved by setting w∗ as
N
1 X 2
w∗ = argminw (yi − xi w)2 + λ||w||2
N i=1
N
1 X 2
= argminw (yi − xi w)2 ...st||w||2 <= c
N i=1

4
The optimized regularized weights which is the ridge regression estimator is w∗ = (X T X +
λI)−1 X T Y . The regularization of Linear Regression may help Ridge to model the non-linearity of
the stock prices. Hence, we chose this model.
• K Neighbors Regressor: The KNN regressor is a non-parametric regressor that stores the training
data, and outputs each new instance according to a majority vote over itsPset of K nearest neighbors
1
in the training set. The regression function is given as fKN N (x) = K i∈Nk (x) yi . K[the number
of neighbors considered] is the major hyper-parameter to be set in this classifier. We considered this
model because the training time of this model is very less.

• Random Forest Regressor : Random forests regressors are ensemble learning methods for
regression that operate by constructing a multitude of decision trees at training time and outputting
the mean prediction of the individual trees. Random decision forests correct for decision trees’ habit
of over-fitting to their training set. This ensemble method may give better results than decision trees.

• Elastic Nets : The elastic net method overcomes

Pp the limitations
Pp of the Ridge and Lasso method
which uses a penalty function based on β1 = j=1 |βj |.β1 = j=1 βj ..
The estimates from the elastic net method are defined by β̂ = argmin(ky − Xβk2 + λ2 kβk2 +
β
λ1 kβk1 ).β̂ = argmin(ky − Xβk2 + λ2 kβk2 + λ1 kβk1 . In this, both l1 and l2 norm is regularized.
β
This variation may help Elastic Nets model the prices accurately. Hence, we chose this model.
• MLPRegressor : A multilayer perceptron (MLP) is a feed forward artificial neural network model
that maps sets of input data onto a set of appropriate outputs. The activation functions are described
by y(vi ) = tanh(vi ) and y(vi ) = (1 + e−vi )−1 . Almost all the projects which are related report
their best results using Neural Networks. The lack of data in our case may not necessarily give us
the best results. However, the ability of Neural Networks to learn non-linear relationships made us
choose this model.

4.4 Pipeline

The flow chart of our pipeline is shown in Figure.2. The data is collected and pre-processed as
described above. The Feature Selection step helps us filter out unnecessary features. We then create
a pipeline combining the Regression and the Feature Selection models. This pipeline is then used
in a Hyper Parameter Selection method called GridSearchCV, which performs an exhaustive search
of the different hyper-parameters and selects the best results by cross-validation. Finally the model
is trained on the training data, after which we predict on the testing data. We evaluate the model
using the Root Mean Squared Error. Evaluation is also done on how well the predictions fit the gold
standard values. The model having the least RMSE and the best fit is identified as the best model.
The RMSE is given as
v
u
u1 X N
RM SE(f, D) = t (yi − f (xi ))2
N i=1

5 Experiments and Results

We experimented with all the models described in the methodology section. Some of the models
clearly outperformed others.

5.1 Baseline

The baseline model we used is Linear Regression. The deviation of predictions from the actuals and
the line plot of the prediction vs actual closing price for our testing data is shown in Fig.3.

5
 Figure 2: Block Diagram of Pipeline Implementation

ICICI_LinearRegression_Baseline_scatterPlot Closing price vs day count for Linear Regression for ICICI Bank
500
500
450
400 400
350
Predicted Values

300
Close price

300
250
200 200
150
100 Actuals
100
Predictions
50
0 0 10 20 30 40 50 60
265 270 275 280 285 290 295 Time points
Actual Value
Figure 3: Predicted vs actual price for Linear Regression

The scatterplot of the Y and Ŷ values is shown in the first sub-plot. The deviation from the linear line
indicates how far the predicted value is from the actual value. The RMSE obtained from the baseline
method without any feature selection and hyperparameter optimization is 99.97.
Analysis: The predicted values always have a deviation from the gold standard due to high prediction
variance of the model. Linear Regression fits a straight line through the data points, and fails to
consider the non-linear nature of the stock prices. Hence, Linear Regression only finds the best fitting
straight line, giving rise to such errors.

5.2 Model Evaluation

Different models are evaluated based on accuracy and speed on all the three data sets. The results on
the ICICI stock data are reported below.

5.2.1 Accuracy
The accuracy of different models on the ICICI stock data are as follows:
Decision Tree(DTR): Decision tree gives an RMSE of 43.287. With feature selection and hyperpa-
rameter optimization, the RMSE reduces to 6.97.
Ridge Regression(Ridge): Ridge regressor gives an RMSE of 118.84. With feature selection and
hyperparameter optimization, the RMSE reduces to 4.53.
KNN(KNR): KNN gives an RMSE of 318.3. With feature selection and hyperparameter selection,

6
the RMSE reduces to 4.35.
Random Forest(RFR):Random Forest gives an RMSE of 44.22. With feature selection and hyper-
parameter selection, the RMSE reduces to 6.12.
Elastic Nets(EN): Elastic Nets gives an RMSE of 231.2. With feature selection and hyperparameter
selection, the RMSE reduces to 4.4.
Neural Network(MLPR): The MLP regressor model is also evaluated on the data. The RMSE
recorded with and without feature selection and hyperparameter optimization is 127 and 4.9 respec-
tively.
Analysis: Although Elastic Nets gave the least accuracy, the variance of the actual values are
modelled better by Ridge Regressor. The reason for Elastic Nets and Ridge giving similar results is
due to the fact that Ridge is l1 normalization and Elastic Net combines l1 and l2 penalties. All the
models were also tested on the Tata Motors and Vedanta stocks, and similar results were obtained.
The line plot of predicted values vs the actual gold standard values for the Ridge Regressor and the
RMSE values for all the models is shown in Figure.4.

400 RidgeBaselineICICI 450 ICICIBaseline RMSE

350 400
300 350
250 300
Closing Price

200 250

RMSE
150 200
100 150
50 Actuals 100
0 Predictions 50
50 0
0 10 20 30 40 50 60 LR DTR KNR Ridge MLPR RFR EN
Day Count Regressor Models
Actuals vs Predictions Accuracies of different models on ICICI stock data
Figure 4: Model performance before feature selection and optimization

The line plot for Ridge Regressor and the accuracy of all the models with feature selection and
hyperparameter optimization on ICICI stock data is summarized in Figure.5.

295 RidgeICICI 7 ICICI RMSE

290 6
5
285
Closing Price

4
RMSE

280
3
275
2
270 Actuals 1
Predictions
265 0
0 10 20 30 40 50 60 LR DTR KNR Ridge MLPR RFR EN
Day Count Regressor Models
Actuals vs Predictions Accuracies of different models on ICICI stock data
Figure 5: Model performance after feature selection and optimization

5.2.2 Value Optimization

Although we see a huge improvement in the predicted values after feature selection and hyperprameter
optimization, the output of the regressors seem to be shifted by δ on the majority of the predictions.
We propose the following method to account for the shift
Correction: The data is trained on the training data as before. We reduce the testing data by one case
and hold out the case as the correction sample xc . The rest of the testing data is used for testing. xc is
then used to shift the output by δ for all the predictions. The predicted output is now given as
ŷ = y 0 − (xc − yc )
where yc is the gold standard output for the correction sample. After the correction, the Ridge
regressor is found to perform exceedingly well, giving an RMSE of 0.33.

7
 295 RidgeBESTICICI
290
285

Closing Price
280
275
270
265 Actuals
Predictions
260
0 10 20 30 40 50 60
Day Count

Figure 6: Price movement after Value optimization

Analysis: After the correction is performed, the model tends to perform exceedingly well, and the
predicted prices accurately modelled. This method works well only for ICICI data. Different types of
corrections may be needed for different stocks, based on the level of shift. Hence, we conclude that
one model and one method will not work well for all the stocks, and feature engineering and method
selection is needed for individual stocks.

5.2.3 Speed
Speed is one of the factors we need to consider when evaluating stock prediction models in High
Frequency Trading(HFT). While we do not deal with HFT in our project, we do need to consider
speed if we’re applying our method to a large number of listings since we’ll have only a 8 hour
window to compute predictions before the NSE opens the next day. The training and test computation
time for the ICICI stock is summarized in the Table 1. The time for different stocks are comparable
to these times.
Model Names Stock Speed in seconds
DecisionTreeRegressor ICICI 0.3531
ElasticNet ICICI 0.1716
KNeighborsRegressor ICICI 0.1489
MLPRegressor ICICI 2.8148
RandomForestRegressor ICICI 5.5945
Ridge ICICI 0.0494
Table 1: Table of speed listings for ICICI Stock

It is observed that Ridge regression performs best in terms of speed. The entire training and testing of
all the data cases takes 0.04 s to complete, which is sufficient for the scope of the problem.

6 Discussion and Conclusion

We observe that certain trade-offs have to be made in each model in terms of speed or accuracy.
Ridge regression gives the overall best accuracy for all the stocks and also gives the best results in
terms of speed. Since we don’t consider speed to be a big factor in our problem, we conclude that
Ridge regression is the best performing model for our problem. Most of the related work treats stock
prediction as a classification problem. [1] achieves a classification accuracy of 60% using Support
Vector Machines. Most of the papers reported their best results using Neural Network. The reason
why neural network failed for our problem is because of the low number of data samples. After the
correction step, our model predicted near-perfect prediction values for the ICICI data but failed to
generalize for other stocks. We did not address the classification problem of predicting whether the
price increases or decreases for a particular day, but it is an interesting problem to pursue further. Also,
we would like to factor the impact that news has on a movement. This can be achieved by scrapping
data off news websites, predicting sentiment of the news and modelling it in the price movement
prediction. We conclude that the stock regression problem is a hard problem to generalize but with
sufficient data and with the consideration of only technical influencers, the price of a particular stock
can be accurately predicted.

8
References
[1]Vatsal H. Shah. Machine Learning Techniques for Stock Prediction. 2009,
[2]Rohit Choudhry, and Kumkum Garg. A hybrid machine learning system for stock market forecasting. In:
World Academy of Science, Engineering and Technology, p. 39 (2008),
[3]Yung-Keun Kwon, Sung-Soon Choi, Byung-Ro Moon. Stock prediction based on financial correlation,
Proceedings of the 2005 conference on Genetic and evolutionary computation, June 25-29, 2005, Washington
DC, USA [doi>10.1145/1068009.1068351],
[4]Paul D. Yoo, Maria H. Kim, Tony Jan.Machine learning techniques and use of event information for stock
market predictions: A survey and evaluation" in , Sydney, Australia:University of Technology, 2005.,
[5]Suchismita Naik.A Narrative Data Visualization Of The Indian Stock Market, Proceedings of the
8th Indian Conference on Human Computer Interaction, December 07 - 09, 2016, Mumbai, India
[doi>10.1145/3014362.3014382],
[6]Birgul Egeli,Meltem Ozturan,Bertan Badur,Stock market prediction using artificial neural networks, in:
Hawaii International Conference on Business, 2003.,
[7]Robert P. Schumaker, Hsinchun Chen. Textual analysis of stock market prediction using breaking financial
news: The AZFin text system, ACM Transactions on Information Systems (TOIS), v.27 n.2, p.1-19, February
2009 [doi>10.1145/1462198.1462204]
[8]K. V. Nesbitt, S. Barrass, "Finding trading patterns in stock market data", IEEE Comput. Graph. Appl., vol.
24, no. 5, pp. 45-55, Sep. 2004.
[9]Seung-Kyu Lee , Byung-Ro Moon, A new modular genetic programming for finding attractive technical
patterns in stock markets, Proceedings of the 12th annual conference on Genetic and evolutionary computation,
July 07-11, 2010, Portland, Oregon, USA [doi>10.1145/1830483.1830704]
[10]Diego J. Bodas-Sagi , Pablo Fernández , J. Ignacio Hidalgo , Francisco J. Soltero , José L. Risco-Martín,
Multiobjective optimization of technical market indicators, Proceedings of the 11th Annual Conference Compan-
ion on Genetic and Evolutionary Computation Conference: Late Breaking Papers, July 08-12, 2009, Montreal,
Québec, Canada [doi>10.1145/1570256.1570266]