Available online at www.sciencedirect.

com

ScienceDirect
The Journal of Finance and Data Science 5 (2019) 1e11
http://www.keaipublishing.com/en/journals/jfds/

Improving trading technical analysis with TensorFlow Long Short-

Term Memory (LSTM) Neural Network
Chenjie Sang, Massimo Di Pierro*
School of Computing, DePaul University, 243 S Wabash Ave, Chicago, IL, 60604, USA
Received 19 May 2018; revised 17 August 2018; accepted 29 October 2018
Available online 14 November 2018

Abstract

In this paper we utilize a Long Short-Term Memory Neural Network to learn from and improve upon traditional trading al-
gorithms used in technical analysis. The rationale behind our study is that the network can learn market behavior and be able to
predict when a given strategy is more likely to succeed. We implemented our algorithm in Python pursuing Google's TensorFlow.
We show that our strategy, based on a combination of neural network prediction, and traditional technical analysis, performs better
than the latter alone.
© 2018 The Authors. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co., Ltd. This is an open access
article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Keywords: LSTM; Neural network; Trading; Tensorflow; Stock market

1. Introduction

Artificial Neural Networks (ANN) are a class of computational models used in Machine Learning (ML). They have
been around for more than 70 years.7 An ANN can be described as a non-linear fitting algorithm whether the fit is
performed by adjusting the weights of information propagation between stacked layers of elementary functions, called
neurons, in analogy with the simplest computational units of the human brains. ANNs have proven to be extremely
good at feature detection compared with other models. The reason is that it is extremely difficult to describe features in
traditional algorithms and, even more so, to create an abstraction paradigm for a generic feature. ANNs have found
many applications in finance12 ranging from modeling stock performance10 to modeling bankruptcies.8 Their
effectiveness in predicting the stock market has been studied and compared with traditional models3 and, as one
naively would expect, results depends on the stock and on market conditions.
There is a revived interest in ANNs because of the confluence of three factors: 1) extremely promising results from
Google in applying Recurrent Neural Networks (RNN) to image and speech recognition; 2) the availability of many
open source libraries that implement different types of networks (for example TensorFlow from Google); and 3) the
increased availability of cheap computing power, thanks to multi-core GPU, such as those produced by NVIDIA.

* Corresponding author.
E-mail address: [email protected] (M. Di Pierro).
Peer review under responsibility of China Science Publishing & Media Ltd.

https://doi.org/10.1016/j.jfds.2018.10.003
2405-9188/© 2018 The Authors. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co., Ltd. This is an open access article
under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
2 C. Sang, M. Di Pierro / The Journal of Finance and Data Science 5 (2019) 1e11

In this paper we propose utilizing an ANN for financial time series forecasting but we take a slightly different spin
in respect to previous work. In particular we ask whether we can use the ANN to learn when traditional and established
Technical Analysis methods work best.
The term Technical Analysis usually refers to the analysis methodology for forecasting the direction of prices
through the study of patterns in past market data, primarily price and volume. Our interest is not so much whether we
can detect those patterns using Neural Networks but whether we can use Neural Networks to augment traditional
algorithms that are already popular and widely used. We assume that there is no fundamental theoretical reason why
most technical analysis algorithms should work other than people use them, thus affecting the stock prices, in a self-
fulfilling prophecy. Yet, it is reasonable to assume that those traders do not always follow the algorithms blindly and
mechanically. Other considerations factor into their trading decisions. The rationale behind our study is that investors
who use technical analysis also apply their own subjective judgment and a neural network can learn that behavior thus
predicting when a given strategy is more likely to succeed.
Specifically we consider the following technical analysis trading strategies:

Simple Moving Average (SMA)

Relative Strength Index (RSI)

Moving average convergence divergence (MACD)

For each of those trading strategies we train a neural network to determine whether the strategy produces a profit or
a loss at a specific moment in time.
We then compare the performance based on annual cumulative profit per 1 share stock investment from using the
naive trading strategy or using the strategy only when the trained neural network recommends doing so.
In our analysis we only consider daily adjusted closing prices and therefore assume that all our trades are executed
at the end of the day without any lag or cost.
For each asset we use two distinct data sets: a training data set and a testing data set. Training data set was selected
from daily stock market data during 2014 and testing data set is being selected from year as of 2015.
We consider different assets: individual stocks in sectors with top five weights that do not have any major M&A
activities and for which we have complete records in the period 2009e2018; and nine indexes representing entire
sectors under Standard & Poor 500. Specifically, the performance of the indexes will be obtained from their ETFs
identified as follows: XLB, XLE, XLF, XLI, XLK, XLP, XLU, XLV, XLV. The stocks, grouped by sector are the
following (see Table 1):

2. Technical analysis trading

All of the technical analysis trading algorithms that we consider here are based on looking for short time trends and
using them to make a forecast. As mentioned above, we decided to adopt trading strategies based on three naive
trading signals: Simple Moving Average (SMA), Relative Strength Index (RSI), Moving average convergence
divergence (MACD). We selected them because they are easy to implement and are commonly used as trading signals.
The SMA signal is defined as follows:

Table 1
Individual stock listing.
S&P 500 Index ETF by Sector
XLY XLP XLF XLK XLI XLB XLE XLU XLV
AMZN PG BRK/B AAPL BA PX XOM NEE JNJ
HD KO JPM MSFT MMM ECL CVX DUK UNH
NFLX PEP BAC FB GE LYB SLB SO PFE
DIS PM WFC GOOGL UNP APD EOG D MRK
CMCSA WMT C INTC HON SHW OXY EXC AMGN
 C. Sang, M. Di Pierro / The Journal of Finance and Data Science 5 (2019) 1e11 3

1X N
SMAt ¼ Pt ð1Þ
N i¼1 i

where SMAt is the moving average at time t, Pti is the price i days before, and N is an arbitrary parameter which we set
equal to 20, based on Praekhaow's research9 which suggests that using a time length equal to 20 generates the most
profitable results in the case of short-term trading.
The SMA trading strategy consists of buying when Pt > SMAt and selling when Pt < SMAt . As previously stated, we
assume all positions will be flat at the end of the day with closing price.
The RSI signal is defined as:
100
RSIt ¼ 100  ð2Þ
1 þ RSt
where
!, !
1 X
N X
N
RSt ¼ Gainti Lossti ð3Þ
N i¼1 i¼1

and
Gaint ¼ maxð0; Pt  Pt1 Þ ð4Þ

Losst ¼ maxð0; Pt1  Pt Þ ð5Þ

We choose the value N ¼ 14 as our trading strategy following the recommendation of Ref. 11 in which the authors
suggest that both N ¼ 14 and N ¼ 21 generate favorable stock trading results and recommend N ¼ 14 for short term
results.
Our RSI strategy consists of buying when RSIt > 50 and selling when RSIt < 50.
The Moving Average Convergence Divergence (MACD) is a momentum indicator that shows the relationship
between two moving averages of prices. It is calculated by subtracting the M-days Exponential Moving Average
(EMA) from the N-days EMA at an earlier time L.
MACDt ¼ EMAM
t  EMAtL
N
ð6Þ

Here EMAt is defined recursively:

2 
EMANt ¼ Ot  EMANt1 þ EMANt1 ð7Þ
N
In our case we choose M ¼ 12, N ¼ 26, and L ¼ 9 as recommended in Ref. 11. We buy when MACDt > 0 and sell
when MACDt < 0. Again we assume the position will be cleared at the end of the day with daily closing price.

3. Neural network and TensorFlow LSTM model

When using a Neural Network we need to distinguish between five distinct concepts: the computer program that
simulates the neural network, the network being simulated (the model), the weights of the network, the training data,
and the test data.
To some extent it does not matter which computer program is used. This only affects the speed, the file formats, and
the hardware we can run the program on. We choose to use TensorFlow1 from Google because its ease of installation,
its speed, and its easy customization using the Python language.
The model is the single most important choice we must make. It is somewhat arbitrary although based on some
empirical facts. The choice of the model corresponds to a choice of neurons, their numbers, their layers, and their
connection topology. This is equivalent to choosing the overall shape of the fitting function. Different choices yield
different results. In this empirical study we decided to utilize the Long Short-Term Memory (LSTM) neural network
instead of other neural networks.
4 C. Sang, M. Di Pierro / The Journal of Finance and Data Science 5 (2019) 1e11

Long Short-Term Memory (LSTM) neural network belongs to Recurrent Neural Network (RNN), which is
specialized in training time-series/sequential data. LSTM neural networks is structured in a similar way as RNN in a
chain structure as follows (see Fig. 1):
A LSTM unit is consisted of four gates: Input Gate; Output Gate; Forget Gate; Update Gate. They are connected in
such way that:
xt : Input Tensor ð8Þ

ht : Output Tensor ð9Þ

W; b : Weights and Biases functions ð10Þ

where
ft is the Forget Gate defined by:
 
ft ¼ sf Wfx xt þ Wfh ht1 þ bf ð11Þ

it is Input Gate defined by:

it ¼ si ðWix xt þ Wih ht1 þ bi Þ ð12Þ

and ot is the Output Gate defined by:

ot ¼ so ðWox xt þ Woh ht1 þ bo Þ ð13Þ

and finally, ct is the State Update Tensor defined as:

ct ¼ ft ct1 þ it tanhðWcx xt þ Wch ht1 þ bc Þ ð14Þ

and
ht ¼ ot tanhðct Þ ð15Þ

The TensorFlow LSTM is able to update the current state with information in the past and compute the gradient
using truncated Backpropogation Through Time (BPTT).4 There is an advantage of Truncated BPTT over full BPTT
when the length of the input sequential data is large as the former is able to effectively prevent gradient vanishing
problem which is common in many neural networks especially Recurrent Neural Networks.
Our sample data consists of 1764 training samples and 504 testing samples. These are relatively small samples,
therefore we used a shallow LSTM models consisting of one single hidden layer in additional to the input and output
layers. In TensorFlow, the hidden unit size is simply equal to the cell size times 4 since one unit consists of four parts
(input gate, output gate, forget gate, hidden state).

Fig. 1. LSTM chain structure.6

 C. Sang, M. Di Pierro / The Journal of Finance and Data Science 5 (2019) 1e11 5

Fig. 2. Cumulative profit differentials (SMA).

We experimented with different learning rates. A large learning rate causes the gradient to bounce and explode. A
small learning rate prevents the cost function from converging to its minimum. Also a small learning rate may cause
the gradient to remain stuck in local minimum. In our experiments, we found that a learning rate as of 0.0001 works
best. The choice of a learning rate will be discussed below.
We also experimented the dropout function in TensorFlow to minimize over-fitting problems but given small data
sample size we found it not to be useful and it resulted in under-fitting problems. We found that setting the batch size to
15 works well for our training process.
Another technique which we utilized is the Batch Normalization (BN)5 method. The general idea of BN is to
normalize the input data around sample mean and variance before the activation function. This allows faster learning
rate. The batch normalization requires scaling and shifting parameters and they were both trained during the back-
propagation process.
The training data set consists of the data we want to fit. The data contains both an output, what we want to predict,
and an input information to be used for the prediction. For each trading strategy, every day, we want to predict the
success of the strategy (Pt  Pt1 ) given the knowledge of the previous N-day prices Pti for i ¼ 1:::7.
From the training data we adjusted the weights (the parameters) of the model to minimize the difference between
predicted data and actual data. The testing data is not used for training but it is used exclusively for benchmarks of the

Fig. 3. Cumulative profit differentials (RSI).

6 C. Sang, M. Di Pierro / The Journal of Finance and Data Science 5 (2019) 1e11

trained network. Throughout the training and testing processes, we assume there are no trading related cost (no
lending or borrowing cost, no margin interest cost etc). The testing result is based on annual cumulative profit in dollar
amount per 1 share of the stock starting at the beginning of the year til the year end.

4. Results

(See Figs. 2e4).

Fig. 4. Cumulative Profit with Differentials (MACD).

5. Sector trading performance before and after training summary in dollars

SMA SMA eTF SMA-Alpha RSI RSI-TF RSI-Alpha MACD MACD eTF MACD-Alpha
XLY 9.11 4.23 4.88 6.69 6.37 0.32 12.85 3.71 16.56
XLP 14.83 3.27 18.1 7.88 9.47 17.35 1.72 6.1 7.82
XLF 1.29 3.83 2.54 0.87 2.76 3.63 0.9 3.97 3.07
XLK 1.47 12.96 14.43 1.09 11.17 10.08 3.2 12.97 9.77
XLI 2.35 10.78 8.43 2.65 13.17 10.52 3.81 3.8 7.61
XLB 2.15 6.65 8.8 6.11 8.21 14.32 0.41 4.93 5.34
XLE 7.12 10.81 17.93 0.91 5.4 4.49 19.89 5.17 14.72
XLU 5.18 13.21 8.03 8.21 5.71 2.5 11.43 8.11 19.54
XLV 4.24 12.69 16.93 6.17 14.59 20.76 4.97 8.06 3.09
TOTAL 32.02 50.27 18.25 14.28 57.91 43.63 36.32 19.74 16.58

As shown in the plots and table above, TensorFlow generated consistent positive results in the S&P 500 sectors
using three trading strategies. Under SMA trading method, TensorFlow outperformed 6 out of 9 sectors; Under RSI
trading method, TensorFlow outperformed 6 out of 9 sectors and under MACD trading method, TensorFlow out-
performed 5 out of 9 sectors as well. Noticing that under each trading method, the total profit after TensorFlow all
outperforms that without TensorFlow training. In numerical analysis, we also found out the annual cumulative profit
per 1 share underlying ETF difference under three trading strategies are þ$18.25, þ$43.63, þ$16.58, respectively. In
individual stock analysis (the tables are appended in appendix), Out of 45 stocks, for each of the three trading stra-
tegies we find 24, 22, and 24 stock, respectively. The annual cumulative profit per share underlying ETF differences
are þ$198.66, þ$10.81, þ$368.96.
 C. Sang, M. Di Pierro / The Journal of Finance and Data Science 5 (2019) 1e11 7

In this study, we discovered that when changing the parameters, including the learning rate and cell size, affects the
performance of the neural network drastically. Consequently we had to carefully choose the parameters setup to
ensure TensorFlow's convergence.

6. Time consumption and average cost for different optimizers

avg time for each sector (seconds) avg cross entropy cost for each batch training
Gradient Descent 0.305 1.03
Adam 0.321 0.91
Adadelta 0.336 1.187
Adagrad 0.315 0.831
RMSProp 0.331 0.945
Momentum 0.316 1.01

Currently there are few optimizers that are commonly used in deep learning training. In this paper, because of the
relatively small sample size, we selected optimizers with algorithm that converge faster. One of such optimizers was
discussed by Adam and Adagrad. They have shown it to be relatively fast using data of comparable size. In our
empirical study, Adam optimizer not only achieves decent performance in terms of speed and cost but also gives the
best performance in computing the 1-yr return of each trading algorithm after training. The table also indicates that no
matter what optimizer we choose, it takes about around 0.3 s for TensorFlow training using only one CPU (excluding
initialization time and data loading time). Using a GPU would further improve the speed.

7. Learning rate and its related cost comparison

Learning Rate Training Cost Under SMA TF Training Cost Under RSI TF Training Cost Under MACD TF
0.00001 0.99 1.2 1.05
0.00005 0.91 0.95 0.9
0.0001 0.89 0.81 0.85
0.0005 0.89 0.95 0.88
0.001 1.05 1.04 0.94
0.005 1.63 1.77 1.65
0.01 2.05 2.23 2.03
0.05 2.21 2.42 2.39

The table above shows the cross entropy cost with different learning rate for all TensorFlow trading methods. And
as we can see learning rate between 0.00005 and 0.001 usually generates the best result with lowest training cost. Thus
we are picking learning rate to be 0.0001 since it generates the lowest training cost among all others.

8. Conclusions

Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) have been widely applied in image
recognition with remarkable success. In this empirical study we explored whether they the potential to enhance the
performance of the trading algorithms. We found positive results and, most importantly, we showed that TensorFlow,
and deep learning in general, can be useful to the Financial Industry. Most financial data is time-serially correlated,
consequently, Long Short-Term Memory (LSTM) and Recurrent Neural Network perform best and, in our studies,
better than traditional trading algorithms.
8 C. Sang, M. Di Pierro / The Journal of Finance and Data Science 5 (2019) 1e11

Acknowledgements

We thank Dr. Mark Shore for useful comments about this manuscript and Prof. Luca De Alfaro for interesting
discussions and advice about LSTM.

Appendix

Annual cumulative Profit per 1 Share Stock Before & After Training Using SMA.

stock name 1-yr return before tf in dollars 1-yr return after tf in dollars 1-yr return alpha in dollars
AMZN 100.9 332.11 231.21
HD 2.375 25.115 22.74
NFLX 49.7205 31.9855 17.735
DIS 5.388 9.308 3.92
CMCSA 14.235 22.785 8.55
PG 11.5596 15.1364 26.696
KO 2.881 10.781 7.9
PEP 11.212 18.48 29.692
PM 33.955 17.325 51.28
WMT 6.665 1.335 5.33
BRK/B 5.0395 23.6505 28.69
JPM 0.2382 0.3018 0.0636
BAC 0.67 5.87 6.54
WFC 9.085 11.985 2.9
C 13.54 29.86 16.32
AAPL 48.7525 20.6825 28.07
MSFT 41.575 14.325 55.9
GOOGL 202.525 211.005 413.53
INTC 9.73 3.81 5.92
V 14.403 13.487 0.916
BA 59.6079 62.7445 3.1366
MMM 47.2381 81.2621 34.024
GE 7.815 0.565 8.38
UNP 6.811 48.179 54.99
HON 8.49 3.05 11.54
PX 23.785 51.845 28.06
ECL 43.775 18.115 25.66
LYB 0.82 7.9 7.08
APD 8.21 42.97 34.76
SHW 26.42 52.28 25.86
XOM 2.68 3.85 6.53
CVX 17.126 14.954 2.172
SLB 12.63 18.93 6.3
EOG 20.74 2.35 23.09
OXY 3.14 3.15 6.29
NEE 7.365 18.955 11.59
DUK 16.2334 18.7134 2.48
SO 0.249 14.611 14.362
D 13.685 3.505 17.19
EXC 8.25 18.49 10.24
JNJ 28.855 25.795 3.06
UNH 29.01 1.09 30.1
PFE 2.325 1.243 1.082
MRK 10.019 5.239 15.258
AMGN 76.3951 20.9349 55.4602
TOTAL 53.0552 145.6018 198.657
 C. Sang, M. Di Pierro / The Journal of Finance and Data Science 5 (2019) 1e11 9

Annual cumulative Profit per 1 Share Stock Before & After Training Using RSI.

stock name 1-yr return before tf in dollars 1-yr return after tf in dollars 1-yr return alpha in dollars
AMZN 145.88 24.55 121.33
HD 12.005 25.965 37.97
NFLX 53.4205 59.5195 6.099
DIS 5.352 18.828 13.476
CMCSA 8.675 18.275 26.95
PG 14.8976 11.9764 26.874
KO 5.199 6.231 1.032
PEP 9.482 20.22 10.738
PM 29.925 29.805 59.73
WMT 10.265 10.875 0.61
BRK/B 21.1095 37.6125 16.503
JPM 1.4382 25.5658 24.1276
BAC 0.63 1.21 1.84
WFC 1.785 10.365 8.58
C 8.62 3.17 5.45
AAPL 15.2625 1.4375 16.7
MSFT 4.875 7.965 12.84
GOOGL 3.335 123.775 127.11
INTC 1.5 10.42 8.92
V 6.257 15.247 21.504
BA 67.4879 55.0499 12.438
MMM 46.1181 81.9221 35.804
GE 10.095 1.455 11.55
UNP 20.547 23.157 43.704
HON 24.35 1.43 22.92
PX 12.565 40.945 28.38
ECL 25.185 7.555 17.63
LYB 20.26 5.48 14.78
APD 11.67 17.23 28.9
SHW 44.94 136.2 91.26
XOM 6.64 20.27 26.91
CVX 18.386 9.686 28.072
SLB 24.13 8.03 32.16
EOG 7.47 32.47 25
OXY 2.92 4.6 1.68
NEE 5.305 4.845 10.15
DUK 10.1534 15.7434 5.59
SO 7.551 11.771 19.322
D 6.405 21.055 27.46
EXC 3.3 7.37 4.07
JNJ 8.375 41.355 32.98
UNH 1.87 32.89 34.76
PFE 8.185 0.453 7.732
MRK 7.679 8.919 16.598
AMGN 31.9549 10.0351 57.99
TOTAL 449.0206 459.8262 10.8056

Annual cumulative Profit per 1 Share Stock Before & After Training Using MACD.

stock name 1-yr return before tf in dollars 1-yr return after tf in dollars 1-yr return alpha in dollars
AMZN 134.83 9.45 144.28
HD 35.795 16.845 18.95
NFLX 23.7595 49.0455 72.805
DIS 6.278 19.298 25.576
(continued on next page)
10 C. Sang, M. Di Pierro / The Journal of Finance and Data Science 5 (2019) 1e11

(continued )
stock name 1-yr return before tf in dollars 1-yr return after tf in dollars 1-yr return alpha in dollars
CMCSA 6.665 22.865 16.2
PG 9.3204 14.1416 4.8212
KO 2.779 11.321 8.542
PEP 28.902 9.14 38.042
PM 1.265 13.905 12.64
WMT 5.505 7.775 13.28
BRK/B 25.4495 19.8105 45.26
JPM 20.6718 20.8982 0.2264
BAC 6.74 2.69 9.43
WFC 1.955 20.005 21.96
C 11.85 10.43 22.28
AAPL 59.6175 31.4375 28.18
MSFT 9.035 5.715 14.75
GOOGL 13.215 98.435 111.65
INTC 4.58 4.55 9.13
V 2.627 36.497 33.87
BA 32.4201 95.2919 127.712
MMM 18.9419 8.0521 10.8898
GE 2.585 11.685 9.1
UNP 41.669 71.739 113.408
HON 8.19 1.49 6.7
PX 14.345 58.585 44.24
ECL 1.375 16.835 15.46
LYB 12.58 16.08 28.66
APD 12.53 16.57 29.1
SHW 62.2 16.06 78.26
XOM 12.65 19.33 31.98
CVX 19.384 23.376 42.76
SLB 4.81 1.05 3.76
EOG 46.76 18.56 28.2
OXY 14.25 1.3 15.55
NEE 8.695 19.975 28.67
DUK 6.1766 24.6234 30.8
SO 14.331 10.229 4.102
D 1.955 6.855 8.81
EXC 0.73 7.15 7.88
JNJ 14.255 43.935 29.68
UNH 6.49 10.37 3.88
PFE 3.925 2.285 6.21
MRK 2.681 6.321 9.002
AMGN 33.2551 1.1649 32.0902
TOTAL 290.3548 78.6042 368.959

References

1. Abadi Martn, Barham Paul, Chen Jianmin, et al. TensorFlow: a system for large-scale machine learning. In: Proceedings of the 12th USENIX
Conference on Operating Systems Design and Implementation (OSDI’16). USENIX Association, Berkeley, CA, USA. 2016:265e283.
3. Fernandez-Rodrguez Fernando, Gonzalez-Martel Christian, Sosvilla-Rivero Simon. On the profitability of technical trading rules based on
artificial neural networks: Evidence from the Madrid stock market. Econ Lett. 2000;69(1):89e94 [in absence of trading costs, the technical
trading rule is always superior to a buy-and-hold strategy for both bear market and stable market episodes. On the other hand, we find that the
buy-and-hold strategy generates higher returns than the trading rule based on ANN only for a bull market subperiod.].
4. Hochreiter Sepp, Schmidhuber Jurgen. Long short-term memory. Neural Comput. 1997;9(8):17351780.
5. Ioffe Sergey, Szegedy Christian. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift.”
[1502.03167] Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. 02 Mar. 2015.
6. Olah Christopher. Understanding LSTM networks. Understanding LSTM networks e Colah's blog. Colah's Blog; 27 Aug. 2015, colah.github.
io/posts/2015-08-Understanding-LSTMs/.
 C. Sang, M. Di Pierro / The Journal of Finance and Data Science 5 (2019) 1e11 11

7. McCulloch Warren, Pitts Walter. A logical Calculus of ideas immanent in nervous activity. Bull Math Biophys. 1943;5(4):115133. https://
doi.org/10.1007/BF02478259.
8. Odom Marcus D, Ramesh Sharda. A Neural Network Model for Bankruptcy prediction. 1990. Neural Networks, 1990., 1990 IJCNN Inter-
national Joint Conference on. IEEE.
9. Praekhaow Puchong. Determination of Trading Points Using the Moving Average Methods. 01 June 2010.
10. Refenes Apostolos Nicholas, Zapranis Achileas, Francis Gavin. Stock performance modeling using neural networks: a comparative study with
regression models. Neural Network. 1994;7(2):375e388.
11. Terence Tai-Leung, Wing-Kam Ng, Venus Khim-Sen Liew. Revisiting the Performance of MACD and RSI Oscillators. Germany: MPRA
Paper. University Library of Munich; 02 Feb. 2014.
12. Trippi Robert R, Efraim Turban. Neural Networks in Finance and Investing: Using Artificial Intelligence to Improve Real World Performance.
McGraw-Hill, Inc.; 1992.