Applied Soft Computing Journal 163 (2024) 111847

Contents lists available at ScienceDirect

Applied Soft Computing

journal homepage: www.elsevier.com/locate/asoc

Algorithmic stock trading based on ensemble deep neural networks trained

with time graph
Muhammed Yilmaz, Mustafa Mert Keskin, Ahmet Murat Ozbayoglu ∗
TOBB University of Economics and Technology, Ankara, 06560, Turkey

ARTICLE INFO ABSTRACT

Keywords: Financial forecasting is generally implemented by analyzing the time series data related to the stock. This is
Financial forecasting accomplished widely with deep neural networks (DNNs) since DNNs can directly extract the related information
Stock market that is otherwise hard to obtain. Time series is the core data representation of financial forecasting problem
Graphs
since it comes naturally. However recent studies show that even if time series representation is necessary, it still
Deep learning
lacks certain aspects related to the problem. One of them is the relationship between the stocks of the market
Deep neural networks
Convolutional neural networks
which can be captured through graph representation. Therefore, DNNs might solve the financial forecasting
Ensemble models problem better when graph and time series representations are combined. In this study, we present different
graph representations that can be used for this purpose. We also present an ensemble network that gives an
investment strategy related to the stock market from stock predictions. Our proposed model returns an average
of 20.09% annual profit on DOW30 dataset through daily buy–sell decisions based on close prices. Therefore,
it can serve as a daily financial investment strategy, offering higher annual returns than conventional heuristic
approaches.

1. Introduction a higher average annual profit than some heuristic strategies, indicating
the need for further improvement to make them suitable for actual
DNNs are used widely for financial forecasting since they can extract trading. This study provides additional graph representations to achieve
the information related to the problem from raw data [1]. However, this goal. DNNs trained with each representation perform similarly,
data representation still plays a crucial role in deep learning since
showing that graph representation methods can vary. This study also
information that can be extracted is strictly correlated to the way
introduces an ensemble model that outperforms the market average in
the information is represented. Different representations can capture
different aspects of the problem even if they belong to the same data. annual return and introduces some heuristic strategies.
Graphs are best suited to represent the relational information of the The key contributions of this paper are:
data which can be significant in a dataset composed of dependent • To the best of our knowledge, for the first time in the litera-
entities. Time series representation can capture the behavior of the
ture, various graph representations, in total, six types of graphs
entity with respect to time but it cannot capture its relation to other
(undirected Pearson, Spearman, Euclidean, and directed Pear-
entities. Therefore, combining two representations can give information
about entities with respect to both time and each other. son, Spearman, Euclidean) are generated and trained with CNNs
Graph representations are applied in several financial forecasting and DNNs in a study, with their corresponding performance
studies since they can capture the relationship between stocks. Analysis comparison.
of graph representation of the market can lead to several inferences • Combining various graph representations (defined in the previous
such as correlation coefficient estimation between stocks, critical stock six graphs) and training a single CNN model, named Graph Com-
analysis or general market behavior prediction. Combining these fea- bined, is one of the other contributions, and we have also shown
tures with time series data leads to more successful and robust financial that it enhances the performance for all six graph models.
forecasting. In a previous study, it was demonstrated that DNNs trained • We employ an ensemble model with graphs in order to outper-
with both graphs and time series return a higher average annual profit form graph-based models and provide more profit. To best of
than some baseline models and the average market growth rate [2].
our knowledge, ensemble models were not used by graph based
However, this study’s further analysis indicates that DNNs do not yield

∗ Corresponding author.
E-mail addresses: [email protected] (M. Yilmaz), [email protected] (M.M. Keskin), [email protected] (A.M. Ozbayoglu).

https://doi.org/10.1016/j.asoc.2024.111847
Received 10 January 2024; Received in revised form 30 April 2024; Accepted 28 May 2024
Available online 21 June 2024
1568-4946/© 2024 Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
M. Yilmaz et al. Applied Soft Computing 163 (2024) 111847

models in finance. Therefore, using this ensemble model with representation of the market. At the end, the previous study calculated
graphs is a novel approach. the average annual return for daily buy/sell decisions based on the
• We exhibit the effectiveness of our graph-based and ensemble stock price prediction of DNNs. This way, it was shown that combining
graph-based models in analyzing the DOW30 stocks. Our ap- undirected or directed graph representation with time series increases
proach involves the use of objective and equitable measurements, the average annual return of daily buy/sell decisions.
which include financial evaluation metrics, baseline techniques, In a recent study the S&P 500 dataset was represented as an
and both MLP and MLP-Multi models. undirected graph, treating each stock as a vertex. Spearman correlation
• Our ensemble model offers a strategy that reduces risk while coefficient was used as the weight of the edges between two stocks.
maximizing returns over the long haul when implementing the The authors state that each vertex’s degree represented how central the
algorithmic trading system. corresponding stock was. Therefore, they were able to analyze the most
reliable stocks and sectors [10]. In our study, we applied Spearman
The rest of the paper is organized as follows. Section 2 explains correlation metric to generate graphs [10] and trained them like in the
other studies that use graph representations for financial forecasting. previous study [2]. The important takeaway of this study [10] is that
Section 3 provides several graph representations of the stock market it gave us an idea of how to generate new graph types, train them, and
and introduces ensemble neural network for financial forecasting. Sec- outperform the performance of the previous [2] model.
tion 4 compares the performance of DNNs trained with different graph
representations. It also evaluates the performance of ensemble model. 3. Method
Finally, Section 5 summarizes and concludes the study.
In this study, we modified and enhanced the model proposed in [2].
2. Related works Our main contribution and enhancement to that former study can be
summarized in two folds: Firstly, we provide several graph representa-
In the literature, there are lots of studies that analyze the graph tions for the same problem showing that graph data provide additional
representation of the market in order to make inference about stocks or robust insights independent of the way the graph is created. Secondly,
the stock market [3–10]. There are significant number of studies that we show that an ensemble network trained with the outputs of earlier
combine the graph representation and DNN since DNNs are generally DNNs achieves better average annual return than heuristic strategies.
very successful to extract features from the data. For example, some We use the same dataset, but extended period of data to 2024, and
studies train support vector machine (SVM), multi-layer perceptron experimental setup with the earlier study [2].
(MLP), convolutional neural network (CNN) or long-short term memory
(LSTM) with graph representation of the stock market [11–15]. In the A. Dataset
literature, there are some models that specialize for graphs such as In this study, we used DOW30 stocks, as in the previous study [2].
graph neural network (GNN) and graph convolutional network (GCN). We gathered the daily stock prices of DOW30 companies from 2012
They are also used for analysis of graph representation of the stock to 2024 using finance.yahoo.com. We trained several DNNs with
market [4,14,16–24]. GNN and GCN are feed with graphs directly. graphs and their ensemble versions to identify the optimal daily
Therefore, they are more successful compared to earlier mentioned buy–sell strategy that maximizes the annual return. We selected the
DNNs for financial analysis. However, they also require more data for daily closing prices of DOW30 as labels in this dataset. However,
training. we calculated the percentage increase in stock prices between two
Table 1 provides an overview of some recently investigated graph- consecutive days to make a stock price time series stationary. During
based models found in the literature. It details characteristics such the training of the models, the label values were normalized inside
as dataset, period, time resolution, method, and performance criteria. the range of [−1, 1].
These studies have enhanced traditional graph models like GCN and B. Baseline Models
GNN by altering their design or graph representation. They have been In the previous study [2], four models were implemented: MLP,
tested across various datasets using a range of evaluation metrics, MLP-multi, Graph, Graph-directed. For each stock, a model was
including accuracy, precision, and recall for classification, as well trained to predict the daily price rate of change of that stock.
as MAE, MSE, and RMSE for regression. Other metrics like Sharpe Then buy/sell strategy was applied according to a heuristic strategy:
ratio, maximum drawdown, cumulative return, and annual return are The stock that was predicted to increase the most was bought.
also examined. These models have demonstrated superior performance This process was repeated 30 times and average annual return was
compared to traditional machine learning, deep learning, and classical calculated. At the end, it was observed that MLP based models bring
graph models. less average annual profit than the average market growth rate. On
On a separate note, ensemble models enhance the performance of the other hand, custom developed graph based models resulted in
individual weak learners trained separately by using techniques such more annual return than the average market growth rate [2]. This
as bagging, boosting, or stacking. These models have been extensively was expected since graph based models also learn from the graph
researched in the context of financial issues, particularly in addressing representation and relation between the stocks in addition to the
challenges in portfolio optimization. They analyze multiple stock prices corresponding time series.
and develop distinct models for each, making the use of ensemble In Graph model, at a given time step, the stock market is rep-
models in finance highly beneficial. Table 2 highlights recent successful resented as an undirected graph such that vertices of the graphs
applications of ensemble models in finance. The research includes correspond to stocks and there is a weighted edge between each
traditional machine learning models like Support Vector Regression pair of vertices. The weight of an edge is calculated by Pearson
(SVR), ensemble approaches such as SVR, LightGBM, and XGBoost, correlation coefficient of the prices of two stocks for the last N
as well as advanced deep learning models like the Transformer and days. N is the parameter that provides the window length for the
Convolutional Neural Networks (CNN), along with other ensemble and correlation calculation. Therefore, the resolution of the history of
Reinforcement Learning (RL) models. stocks can be determined by changing N. A graph is represented
In the previous study, DOW30 was represented as 30 × 30 undi- with its adjacency matrix. It is preferred since CNN is specialized to
rected and directed graph [2]. Baseline DNN is trained with time extract features from 2-Dimensional data.
series data and some extracted features to predict the daily stock price Graph-directed model also uses the adjacency matrix. However,
increase. Graph based models were trained with combining the same the adjacency matrix of Graph-directed model is not symmetric.
data and features with the output of CNN which is feed with graph Therefore, an edge from stock A to stock B is different than an edge

2
M. Yilmaz et al. Applied Soft Computing 163 (2024) 111847

Table 1
Algorithmic stock trading with graph models.
Art. Data set Period Time Method Performance
resolution criteria
[25] CSI 300 2015–2019 Daily Multi- Acc
CSI 500 GCGRU PR
RC
F1
MDD
[26] DJI 2002–2020 Daily GCN, RL MDD
SR
Return
[27] Taiwan Stock – Daily FinGAT MRR
S&P 500 Precision
NASDAQ
[28] NASDAQ 2011–2021 Daily GCNET Accuracy
MRR
[29] S&P 500 2016–2021 Daily THGNN Acc
CSI 300 PR
RC
F1
MDD
CR
[30] Bitcoin 2015–2021 Daily GNN, LSTM MAE
Litecoin RMSE
Ethereum
Dash Coin
[28] S&P 500 2012–2016 Daily GCNET Accuracy
NASDAQ 2011–2021
[31] DowJones2005 2013–2018 Daily GNNs TR
EuroBonds SR
ItBondComodities
WorldMixBonds
[32] WTI 1983–2022 Daily LSTM, GCN MSE
RMSE
R2
MAPE
[33] Chinese market – – GCN Acc, F1, Pr
[34] American & Taiwan – 1 Day CNN-LSTM Acc
market 3 Days
7 Days
[35] China A-shares 2018–2019 Daily MAGNN A return
market D return
Sharpe Ratio
[36] S&P 500 2005–2021 Daily SCRG PnL
NIST-GNN Sharpe Ratio
NIST-GNN-SCRG t-Statistics

from stock B to stock A. The weight of an edge from stock A to stock invested only in the stock that brought the highest annual return
B represents the strength of the correlation between the past price in the previous year. In Strategy2, we bought 3 stocks, which
of stock A and the current price of stock B. In other words, it shows resulted in the highest annual returns in the previous year. While
how the price of stock B is affected by the price of stock A. This strategies 1 and 2 represent good baseline strategies, there were
is accomplished with adding the following additional constraints to issues with the growth rates of the APPLE stock price in 2020
Pearson correlation formula which was given in the previous study: and CAT (Caterpillar Inc.) in 2017. These rates were significantly
higher than the DOW Jones average or their previous growth rate.
Therefore, these values were assumed to be anomalies. As a result,
𝐺𝑡 (𝑖, 𝑗) = max(𝐶((𝑡 − 𝑘)𝑖 , 𝑡𝑗 )) 0 ≤ 𝑘 ≤ 𝑀 − 1 (1) we implemented two additional strategies, Strategy3 and Strategy4,
for this case. These strategies excluded the aforementioned stocks
𝐺𝑡 (𝑖, 𝑗) = 𝐺𝑡 (𝑗, 𝑖) ∃ (𝑖, 𝑗) ∈ 𝐷 (2) when calculating baseline greedy strategies solely for the years that
exhibited anomalies. Therefore, in 2017, CAT stock market was
Sliding window method is used to calculate the weight 𝐺𝑡 (𝑖, 𝑗). For
removed for calculating strategies 3 and 4, and in 2020, APPLE
this, initially the correlation between stock i and j is calculated from
was removed as well. Hence, for Strategies 3 and 4, the calculations
the data with window size N. The position of the window is fixed
for these two years were based on 29 stock prices. While Strategy3
for stock j but each time it is shifted one day backwards for stock
mirrored Strategy1, and Strategy4 was identical to Strategy2, the
i and new correlation is calculated. This process is repeated for M
sole difference lied in the exclusion of stock prices for the years
times and maximum value is chosen as the weight of the edge from
2017 and 2020. We have modified these strategies to ensure a more
stock A to stock B.
comparable evaluation. As a result, we decided to use the following
We selected several basic and greedy strategies to compare the
models and strategies as baselines:
performance of our novel models. First, the DOW30 index is the
average return of 30 Dow stocks for each year. In Strategy1, we • MLP [2]

3
M. Yilmaz et al. Applied Soft Computing 163 (2024) 111847

Table 2
Algorithmic stock trading with ensemble models.
Art. Data set Period Time Method Performance
resolution criteria
[37] A-share market 2013–2019 Daily HGTAN Acc
Pre
Recall
F1
[38] DAX 2018–2023 Daily XGBoost MAE
MASI LSTM MAPE
HKEX LSTM-XGBoost RMSE
CAC 40 R2
NASDAQ
FTSE 250
[39] S&P 500 2012–2019 Daily RL MDD
Return etc.
[40] BTC 2018–2019 Hourly LSTM Acc, AUC, F1
ETH BiLSTM
XRP CNN
[41] DAX 2014–2019 Daily CNN-LSTM MSE, MAE
DOW 2000–2019 GRU-CNN
S&P500 2017–2021 Average
Bagging
Stacking
[42] Cryptocurrency – 30 m CNN, DRL APV
SR
MDD
[43] S&P500 – Daily Transformer Return rate
Graph SR
CNN MDD
Ranking
Ensemble
[44] JSE – – ANN MSE
Bagging Ensemble RMSE
Monte Carlo R2
[45] SP100 2011–2021 Daily Fuzzy Ensemble Return rate
RL SR
Ensemble
[46] EU ETS 2008–2020 Daily ARMA RMSE
SCI 2010–2020 CNN-LSTM MAPE
BTC 2012–2020 MAE
Dstat
[47] DOW30 2000–2021 Daily RL AR, CR
Japanese Stocks SR, MD
British Stocks
[48] Various dataset – Daily BiLSTM DL, RL RMSE, MAE, DA
[49] CSI 300 2020–2022 Daily SVR, Lightgbm Various metrics
XGBoost, Random
Forest
LSTM

• MLP-Multi [2] where 𝜌 denotes the Pearson correlation coefficient applied to rank
• Graph Undirected [2] functions of two variables. Spearman correlation calculates mono-
• Graph Directed [2] tonic relationship between two variables whereas Pearson corre-
• Strategies 1, 2, 3 and 4 lation calculates the linear relationship. In linear relationship, the
change in one variable is proportional to the change in the other
C. Similarity Metrics and Graph Models variable. In monotonic relationship, the variables move together
In the previous study, Pearson correlation coefficient was used as but not at a constant rate. Therefore, unlike Pearson correlation,
the weights of the edges in graph representation of the stock market. Spearman correlation coefficient can calculate the correct weight
However, other methods also can be used as weights as long as they when the correlation relation is nonlinear.
represent a similarity or dissimilarity metric between two stocks. We also used the inverse of Euclidean distance as the weight of the
Therefore, in the first part of this study we implemented other graph edge between two stocks in order to show that a DNN can also learn
representations in order to test whether a DNN could also bring the relationship between stocks from a graph representation based
similar annual profit independent of the graph representation. on dissimilarity metric. Euclidean distance between two points is:
Firstly, graph representation with Spearman correlation coefficient √
√𝑁
as its weights was implemented. Spearman correlation is a similarity √∑
𝑑(𝑋, 𝑌 ) = √ (𝑋𝑛 − 𝑌𝑛 )2 (5)
metric like Pearson correlation. Its formula is:
𝑛=1
𝑐𝑜𝑣(𝑅(𝑥), 𝑅(𝑦))
𝑟𝑠 = 𝜌𝑅(𝑥),𝑅(𝑦) = (3) Then, the weight between stocks i and j becomes:
𝜎𝑅(𝑥) 𝜎𝑅(𝑦)
40
𝐶(𝑡𝑖 , 𝑡𝑗 ) = (6)
𝐶(𝑡𝑖 , 𝑡𝑗 ) = 𝑟𝑠 (4) 𝑑(𝑋𝑡𝑖 , 𝑋𝑡𝑗 )

4
M. Yilmaz et al. Applied Soft Computing 163 (2024) 111847

Fig. 1. CNN model trained by Graph-combined data.

where 40 is the normalization constant. developed. This approach allows for investment in the market with
Spearman and Euclidean models were trained with undirected potentially greater success. The main motivation of using ensemble
graph which was created with Spearman correlation coefficient and networks is to formulate such strategy.
Euclidean distance respectively for a window size of N similar to In the previous study [2] a heuristic investment strategy was applied
the earlier study. Also, directed graphs were created with the same for the buy/sell decision: Each day, the money was invested to the
sliding window process of the previous study which were used stock which was predicted to increase the most. In this study, we
to train Spearman-directed and Euclidean-directed models. Graph trained a DNN to predict the best stock for investment. The DNN was
representations were used to train CNN the same way with the further trained with the outputs of the 30 models which were each
previous study except the hyper-parameter tuning stage. previously trained to predict the percent profit of a specific stock.
The weights of the edges were constructed differently in each graph In addition to the outputs of 30 models, other features were also fed
model. Therefore, it can be claimed that they can capture the dif-
to the ensemble model in order to provide a better understanding
ferent aspects of the relationship between stocks. If this hypothesis
of the problem. These features are:
is true, a DNN trained by the combination of all graph representa-
tions (Graph-combined) can have better inference performance and – R2 score between last predictions and target (real) values
bring more annual profit. In order to test it, graph representations – Correlation coefficient between last predictions and target
with window size 5, 10, 20 were combined as single data. There values
were 6 graph representations in total: Graph, Graph-directed, Spear- – Ratio of correct sell/buy transactions between last predictions
man, Spearman-directed, Euclidean, Euclidean-directed. Therefore, and target values.
the input of Graph-combined model becomes a tensor with size – Target values of the last day
18 × 30 × 30. Fig. 1 illustrates the CNN trained with combined
graph data in details. The only difference in its structure from the These features were calculated with a window size of 10 which is
previous study’s was that the input is 18 × 30 × 30 tensor instead optimized through exhaustive search. Each feature was represented
of 3 × 30 × 30 tensor. as a 30 × 1 vector, reflecting inputs from 30 distinct models. After
D. Ensemble Models combining all features with the predictions of the models, a 150 × 1
Ensemble networks are widely used since they can make inference input vector was created to train the ensemble neural network. The
that other models cannot with combined knowledge of several ensemble neural network, which is shown in Fig. 2, consists of the
models. There are 30 stocks in DOW30. Successfully determining following layers:
the behavior of all stocks in each day is a difficult task. Moreover,
instead of determining which stock prices will increase next day, – Input layer consists of 150 neurons
determining the stock whose price will increase the highest might – Hidden layer consists of 128 neurons with ReLU activation
be more desirable even if it is more challenging. This can be function
accomplished with ensemble models. In our study, a total of 30 – Hidden layer consists of 256 neurons with ReLU activation
models were trained, each predicting the behavior of a stock price function
in the next day independent of the others. Rather than employing – Output layer consists of 30 neurons with softmax activation
a different strategy for each stock, a single unified strategy can be function.

5
M. Yilmaz et al. Applied Soft Computing 163 (2024) 111847

Fig. 2. Ensemble neural network trained to predict the stock with the most percentage increase.

The value of each output layer neuron provides the corresponding In the previous study [2], Graph and Graph-directed models were
stock’s likelihood of being the stock with the largest percentage trained to predict the daily price change. Then the heuristic strategy
increase in the next day. Therefore, the label of the ensemble model of buying the stock with the most anticipated percentage increase
is a 30 × 1 one-hot vector. The main purpose of the ensemble was applied for buy/sell decisions. In that study, all models were
network is to learn the combined behavior of stock prices and trained until 2019 [2]. We retrained Graph and Graph-directed models
choose the best stock to invest for each day. Categorical cross- from 2012 to 2023 in this study. As seen in Table 4, their average
entropy was used to train the ensemble network. Once the ensemble annual returns were 14.15% and 14.43%, respectively, which were
network was trained, the stock with the highest probability was both greater than the average of DOW30, MLP and MLP-Multi. In this
selected for trading in making buy/sell decisions. study, even though the results were promising, we anticipated that
We trained the ensemble DNN model to generate a good and reliable the performance could be further enhanced through the application
buy/sell daily trading strategy since neural networks can create of additional heuristics or strategic modifications. To investigate that,
robust mathematical models and have been used in real world we calculated the annual returns of several greedy strategies includ-
problems. It is also possible to formulate strategies to choose the ing Strategy1, Strategy2, Strategy3 and Strategy4. Strategy1 (and also
best stock fund for investment without using an ensemble model. Strategy3) was not only more greedy than Strategy2 (and Strategy4)
One way of accomplishing that is using greedy algorithms. We but also it was more sensitive to noise due to its single choice approach,
developed several greedy strategies for choosing the best stock yet, in practice, it was able to bring more annual profit than Strategy2
fund for buy/sell decisions in order to use them as baselines and as it can be seen from Table 3. Table 3 illustrates the annual returns
evaluated the performance of the developed ensemble model. The of Graph, Graph-directed, Strategy1, Strategy2, Strategy3, Strategy4
most successful heuristic algorithm involved purchasing the stock and growth rate of DOW30. The average annual returns of Graph and
forecasted to yield the highest profit. Therefore, we compared the Graph-directed models were not greater than the average annual return
annual return of that strategy with the annual return of the en- of Strategy1, 2, 3, 4 except DOW30 Average. Strategy1, with 24.76%
semble model. Ensemble models were trained only for graph based annual profit, gave the best performance among heuristic strategies. De-
models since they performed significantly better than MLP based spite Strategy1’s success, its worst-case scenario, which was −16.16%,
models. For each graph based application, an ensemble model was indicates it is also a risky strategy. Strategy2 was much more reliable
trained and tested. Each ensemble model received the outputs from and less volatile. Strategy3 and Strategy4 were modified versions of
30 models as inputs and generated a prediction for the following day Strategy1 and Strategy2 due to anomalies in the volatility of stock
in the form of a 30 × 1 vector. This vector provided the probability prices in 2020 and 2017. Upon eliminating the anomalies, the average
of a favorable investment choice for each stock. Then, similar to annual returns of Strategy1 and Strategy2 dropped from 24.76% and
the heuristic strategy, we purchased the stock that was projected to 18.48%, to 18.73% and 15.0%, respectively. This adjustment allowed
generate the highest profit. us to evaluate our models without the influence of abnormal increases.
However, Graph and Graph-directed models still could not perform
better than Strategy1, Strategy2, Strategy3 and Strategy4.
4. Performance evaluation and discussions Realizing that Graph and Graph-directed models could not outper-
form the simple greedy algorithms used in Strategy1, Strategy2, Strat-
In order to compare the results of this study with the results of egy3, and Strategy4 from a return performance perspective, we focused
the previous study [2], the same evaluation setup and metrics of the on enhancing their strategies through the use of ensemble models to see
previous study were used. Each model was trained for two years to if we could surpass their average performance. In Ensemble-undirected
predict the daily percentage price change of a stock. In the following and Ensemble-directed models, instead of the heuristic strategy, corre-
year, predictions of the models were used for buy/sell decisions and sponding ensemble networks were used to improve the performance
the annual return of the heuristic strategy or ensemble model was of Graph and Graph-directed models respectively. After the training
calculated. This process was repeated 30 times for the test year and process explained in the previous section, Ensemble-undirected and
their averages are presented for each year. Ensemble-directed models resulted in annual returns of 18.33% and

6
M. Yilmaz et al. Applied Soft Computing 163 (2024) 111847

Table 3 show that graph representation is powerful and that it improves the
Comparison of the average annual return of several heuristic strategies and Pearson
annual return regardless of the graph representation method. Similarly,
based graph models from 2012 to 2023.
the effectiveness of ensemble methods is evident, as shown in Table 7,
Year Graph Graph Strategy1 Strategy2 Strategy3 Strategy4 DOW30
directed average
where all graph types benefit from the use of ensemble models. These
models not only yielded better results but also demonstrated to be
2023 −1.37 −4.99 1.84 −5.90 1.84 −5.90 −1.47
2022 −5.09 −1.63 −6.61 17.7 −6.61 17.7 −2.79
robust across various graph types.
2021 23.58 23.64 38.06 37.9 38.06 37.9 25.49 At any given time step, Graph-combined network was trained with
2020 11.58 9.68 78.24 37.17 39.94 16.14 5.48 18 × 30 × 30 input which was the combination of 6 graph repre-
2019 22.31 26.26 23.60 25.27 23.60 25.27 25.55 sentations. Using the heuristic strategy with Graph-combined network
2018 0.97 2.58 10.77 −3.26 10.77 −3.26 0.29
resulted in 16.69% average annual return. Meanwhile, the ensemble
2017 25.26 22.43 72.7 45.81 38.61 25.06 25.89
2016 15.04 15.67 −16.16 −1.65 −16.16 −1.6 16.52 model trained with Graph-combined network outputs brought 20.09%
2015 2.25 4.81 −2.40 3.14 −2.40 3.14 2.46 average annual return which was the highest achieved average annual
2014 19.49 19.83 −2.67 9.72 −2.67 9.72 15.79 return. This indicates that integrating various graph representations
2013 34.23 35.89 47.87 38.9 47.87 38.9 30.38 enhanced the annual return. Graph-combined network achieved more
2012 21.50 18.98 51.87 16.95 51.87 16.95 14.56
profit than Strategy2, Strategy3, Strategy4 except for Strategy1. Hence,
Average 14.15 14.43 24.76 18.48 18.73 15.0 13.11
we showed that Graph-combined networks can be highly practical to
implement algorithmic trading with less risk and more profit.
Table 4 Tables 8, 9, 10 compares the performance of the graph based models
Comparison of the average annual returns of baseline models and Pearson correlation with the heuristic strategy from 2012 to 2024 using the financial
based graph models with heuristic strategy from 2012 to 2023. evaluation metrics, namely cumulative return (CR), win ratio (WR),
Year MLP MLP-Multi Graph Graph DOW30 sharpe ratio (SR) and maximum drawdown (MD). In Table 8, the CR
directed average
and SR of Graph, Graph Directed models were strictly greater than
2023 −13.65 −1.48 −1.37 −4.99 −1.47 MLP and MLP-Multi. In Table 9, although the WR and MD values of
2022 −18.45 −8.05 −5.09 −1.63 −2.79
all models were within close proximity, the CR and SR values of the
2021 25.42 29.74 23.58 23.64 25.49
2020 −1.33 1.41 11.58 9.68 5.48 graph combined model were strictly greater than the other graph-based
2019 7.88 17.93 22.31 26.26 25.55 models. It can be noticed from Table 9 that Sharpe ratios of all models
2018 −9.05 −3.92 0.97 2.58 0.29 reflected the relative performance observed through the CR values, as
2017 23.21 22.82 25.26 22.43 25.89 expected. Sharpe ratio and cumulative returns of directed graphs were
2016 16.49 16.11 15.04 15.67 16.52
2015 0.80 6.49 2.25 4.81 2.46
satisfactory compared to those of undirected graphs. Consequently, the
2014 20.93 14.53 19.49 19.83 15.79 combined graph model conducted transactions that generated higher
2013 33.83 35.0 34.23 35.89 30.38 profits and avoided riskier ones. Table 10 shows the performance
2012 19.99 17.70 21.50 18.98 14.56 of the ensemble model from 2012 to 2023 using the same financial
Average 8.83 12.36 14.15 14.43 13.18 evaluation metrics. The ensemble of directed graphs performed better
than the ensemble of undirected graphs except the Pearson directed
graph. The ensemble model, which integrated all the graphs, achieved
Table 5
Comparison of the average annual returns of Strategy1, Strategy3, Pearson correlation the highest Cumulative Return (CR) and Sharpe Ratio (SR) compared to
based models with heuristic strategy and ensemble networks from 2012 to 2023. other ensemble and graph models, with these scores being significantly
Year Graph Graph Ensemble Ensemble Strategy1 Strategy 3 higher than the rest.
directed undirected directed Fig. 3 displays the cumulative returns for MLP, MLP-Multi, Graph,
2023 −1.37 −4.99 −8.574 −8.60 1.84 1.84 and Graph-Directed models using the heuristic strategy from 2012 to
2022 −5.09 −1.63 −6.92 −9.51 −6.61 −6.61 2023. Among these, the Graph and Graph-Directed models achieved the
2021 23.58 23.64 25.54 19.31 38.06 38.06
highest profit compared to MLP and MLP-Multi. As clearly illustrated
2020 11.58 9.68 8.540 15.81 78.24 39.94
2019 22.31 26.26 32.26 31.34 23.60 23.60
in Fig. 4, the Graph-combined model secured the highest profit across
2018 0.97 2.58 2.03 −2.33 10.77 10.77 all time periods. Fig. 5 presents the cumulative returns of the ensemble
2017 25.26 22.43 25.25 29.58 72.7 38.61 models from 2012 to 2023, where the ensemble approach of the Graph-
2016 15.04 15.67 28.76 26.96 −16.16 −16.16 combined model delivered the greatest cumulative and average profit
2015 2.25 4.81 9.79 13.82 −2.4 −2.4
among the models.
2014 19.49 19.83 26.20 24.91 −2.67 −2.67
2013 34.23 35.89 46.64 37.33 47.87 47.87
2012 21.50 18.98 30.42 33.73 51.87 51.87 5. Conclusion and future work
Average 14.15 14.43 18.33 17.69 24.76 18.73
There are several studies that use graph representation for financial
forecasting. The main motivation of this study is to focus on the annual
return of the investment strategy and try to develop a reliable and
17.69% respectively in the test period. Hence, annual return of the better performing graph based trading strategy. For this purpose, we
heuristic strategy was overwhelmingly boosted by the ensemble net- investigated several graph representations of the stock market. For each
works as presented in Table 5. Both ensemble models also outperformed stock, a DNN was trained with graph and time series data. From these
Strategy4 and they had comparable performances with Strategy3 and models a final ensemble model was trained. The algorithmic trading
even Strategy2, which included anomalies. strategy based on the decisions made by the graph ensemble model
In addition to Pearson correlation based models (Graph and Graph- was applied according to the output of this model and 20.09% average
directed), Spearman, Spearman-directed, Euclidean, Euclidean-directed annual profit was achieved.
and Graph-combined models are also implemented. Their average an- Our results indicate that using various graph models outperforms
nual returns with heuristic strategies and ensemble models are pre- the performance of financial forecasting by extracting additional fea-
sented in Tables 6 and 7 respectively. As it can be seen from Table 6, tures from graph information and feeding them into DNN networks.
the results of directed graphs are better than their corresponding undi- Moreover, adding ensemble models to graph models and training them
rected graphs. However, Pearson, Spearman and Euclidean graphs boosts their performance by comparing the heuristic human based
provided similar results regardless of their types. Hence, these results strategies. These models not only brought more profit, but they were

7
M. Yilmaz et al. Applied Soft Computing 163 (2024) 111847

Table 6
Comparison of the average annual returns of graph based models with heuristic strategy from 2012 to 2023.
Year Graph Graph Spearman Spearman Euclidean Euclidean Graph
directed directed directed combined
2023 −1.37 −4.99 −3.21 −1.39 −0.18 −0.21 2.440
2022 −5.09 −1.63 1.89 −4.74 −2.91 −0.804 −0.36
2021 23.58 23.64 17.82 29.97 21.13 20.96 22.28
2020 11.58 9.68 17.81 11.17 8.64 10.35 15.49
2019 22.31 26.26 22.98 32.51 27.61 30.29 25.84
2018 0.97 2.58 −0.27 7.07 2.40 4.53 3.30
2017 25.26 22.43 28.47 24.18 26.53 26.08 26.60
2016 15.04 15.67 17.11 18.50 17.80 17.05 18.21
2015 2.25 4.81 0.87 4.52 2.02 3.58 10.52
2014 19.49 19.83 15.02 10.41 18.59 14.09 13.15
2013 34.23 35.89 35.29 33.71 28.43 34.05 42.48
2012 21.50 18.98 20.93 19.12 14.50 18.37 20.27
Avg 14.15 14.43 14.56 15.42 13.71 14.86 16.69

Table 7
Comparison of the average annual returns of graph based models with ensemble model from 2012 to 2023.
Year Graph Graph Spearman Spearman Euclidean Euclidean Graph
directed directed directed combined
2023 −8.57 −8.60 −12.42 −12.80 −8.19 −11.54 −5.64
2022 −6.92 −9.51 −8.70 −8.80 −7.93 −5.10 −7.96
2021 25.54 19.31 21.0 25.89 25.05 20.84 23.08
2020 8.54 15.81 14.13 23.40 15.25 20.29 18.16
2019 32.26 31.34 28.65 26.29 26.54 27.25 33.82
2018 2.03 −2.33 −3.99 −3.79 −4.36 −4.38 3.62
2017 25.25 29.58 28.49 25.92 28.48 26.03 26.08
2016 28.76 26.96 27.54 35.26 29.73 30.66 35.34
2015 9.79 13.82 14.65 14.39 8.42 11.24 13.54
2014 26.20 24.91 29.05 28.98 25.51 28.94 27.04
2013 46.64 37.33 44.10 47.37 38.67 43.94 49.50
2013 30.42 33.73 28.68 31.76 29.86 21.84 24.52
Avg 18.33 17.70 17.60 19.49 17.25 17.50 20.09

Fig. 3. Cumulative returns of graph based models from 2012 to 2023.

Table 8 In the future, these novel approaches will be used on various financial
Assessing the performance metrics MLP, MLP-Multi [2] and Pearson correlation graph datasets and stock prices in order to determine the relationship between
based models from 2012 to 2023. The table includes four acronyms: CR (Cumulative different stock prices and investment strategies. Therefore, its general
Return), WR (Win Ratio), SR (Sharpe Ratio), and MD (Maximum Drawdown). structure is appropriate for various datasets. Another study that can
MLP MLP-Multi Graph Graph be considered as future work is that this ensemble neural network
directed
structure can be used with different types of models that have been
CR(%) 138.40 281.42 366.06 395.34 studied in the literature previously or will be studied next because
WR 0.515 0.519 0.519 0.521
our ensemble neural network model structure is highly suitable for
SR 0.353 0.518 0.586 0.594
MD −0.581 −0.493 −0.482 −0.471
any model type with sufficient effort, such as fine-tuning this structure
or changing its network size for any other problems. In the future,
combining various models with this graph-based model can potentially
yield higher profits. Therefore, our graph models and ensemble model
also more reliable and robust. Hence, using this ensemble structure structure can be integrated into existing financial forecasting; they
with graph-based models is a novel approach to financial forecasting. can outperform existing financial forecasting strategies or be studied

8
M. Yilmaz et al. Applied Soft Computing 163 (2024) 111847

Table 9
Comparison of the evaluation scores of graph based models with ensemble networks from 2012 to 2023.
Year Graph Graph Spearman Spearman Euclidean Euclidean Graph
directed directed directed combined
CR(%) 366.06 395.34 389.83 451.03 371.07 374.77 508.83
WR 0.519 0.521 0.521 0.521 0.521 0.52 0.523
SR 0.587 0.595 0.595 0.638 0.586 0.599 0.669
MD −0.483 −0.472 −0.458 −0.453 −0.479 −0.477 −0.497

Table 10
Comparison of the evaluation scores of graph based models with ensemble networks from 2012 to 2023.
Year Graph Graph Spearman Spearman Euclidean Euclidean Graph
directed directed directed combined
CR(%) 617.14 590.85 563.96 647.27 574.04 531.91 740.71
WR 0.521 0.520 0.518 0.522 0.520 0.519 0.522
SR 0.656 0.655 0.628 0.689 0.632 0.649 0.725
MD −0.522 −0.511 −0.546 −0.525 −0.530 −0.507 −0.507

Fig. 4. Cumulative returns of graph based models from 2012 to 2023.

Fig. 5. Cumulative returns of ensemble models from 2012 to 2023.

in future works. In the future, we believe further enhancements are attention, effort, and training time compared to deep neural networks
possible through alternative heuristics or strategy adjustments. Instead on financial problems due to their highly dynamic properties. These
of using ensemble models with our graph-based models, algorithms ideas will be examined and compared with our neural network en-
such as genetic algorithms and reinforcement learning can be used semble models. This study indicates that using and training various
to find investment strategies. These types of algorithms find different graph models are innovative approaches that have the potential to
strategies where deep neural networks cannot, but they require more achieve good performance. Furthermore, ensemble neural networks

9
M. Yilmaz et al. Applied Soft Computing 163 (2024) 111847

trained with time graphs present high performance, improvement, and [16] W. Li, R. Bao, K. Harimoto, D. Chen, J. Xu, Q. Su, Modeling the stock relation
robustness. with graph network for overnight stock movement prediction, in: No. CONF,
2020, pp. 4541–4547.
[17] R. Kim, C.H. So, M. Jeong, S. Lee, J. Kim, J. Kang, Hats: A hierarchical
CRediT authorship contribution statement graph attention network for stock movement prediction, 2019, arXiv preprint
arXiv:1908.07999.
Muhammed Yilmaz: Writing – original draft, Supervision, Method- [18] W. Chen, M. Jiang, W.-G. Zhang, Z. Chen, A novel graph convolutional feature
ology, Conceptualization. Mustafa Mert Keskin: Writing – original based convolutional neural network for stock trend prediction, Inform. Sci. 556
(2021) 67–94.
draft, Validation, Methodology, Data curation. Ahmet Murat Ozbayo-
[19] Y. Chen, Z. Wei, X. Huang, Incorporating corporation relationship via graph
glu: Writing – review & editing, Writing – original draft, Supervision, convolutional neural networks for stock price prediction, in: Proceedings of the
Methodology. 27th ACM International Conference on Information and Knowledge Management,
2018, pp. 1655–1658.
Declaration of competing interest [20] D. Loe, S.-l. Chang, J. Chau, Stock market movement prediction using graph
convolutional networks, UCSD Data Sci. Capstone Proj. 2021 (2020).
[21] D. Matsunaga, T. Suzumura, T. Takahashi, Exploring graph neural networks for
The authors declare that they have no known competing finan- stock market predictions with rolling window analysis, 2019, arXiv preprint
cial interests or personal relationships that could have appeared to arXiv:1909.10660.
influence the work reported in this paper. [22] P. Patil, C.-S.M. Wu, K. Potika, M. Orang, Stock market prediction using ensemble
of graph theory, machine learning and deep learning models, in: Proceedings
Data availability of the 3rd International Conference on Software Engineering and Information
Management, 2020, pp. 85–92.
[23] X. Ying, C. Xu, J. Gao, J. Wang, Z. Li, Time-aware graph relational attention net-
Data will be made available on request. work for stock recommendation, in: Proceedings of the 29th ACM International
Conference on Information & Knowledge Management, 2020, pp. 2281–2284.
Declaration of Generative AI and AI-assisted technologies in the [24] J. Ye, J. Zhao, K. Ye, C. Xu, Multi-view graph convolutional networks for
writing process relationship-driven stock prediction, 2020, arXiv preprint arXiv:2005.04955.
[25] J. Ye, J. Zhao, K. Ye, C. Xu, Multi-graph convolutional network for relationship-
driven stock movement prediction, in: 2020 25th International Conference on
During the preparation of this work the author(s) used chatGPT in
Pattern Recognition, ICPR, IEEE, 2021, pp. 6702–6709.
order to rephrase some of the sentences and explanations for better [26] F. Soleymani, E. Paquet, Deep graph convolutional reinforcement learning for
clarity. After using this tool/service, the author(s) reviewed and edited financial portfolio management–DeepPocket, Expert Syst. Appl. 182 (2021)
the content as needed and take(s) full responsibility for the content of 115127.
the publication. [27] Y.-L. Hsu, Y.-C. Tsai, C.-T. Li, Fingat: Financial graph attention networks for
recommending top-𝑘 k profitable stocks, IEEE Trans. Knowl. Data Eng. 35 (1)
(2021) 469–481.
References
[28] A. Jafari, S. Haratizadeh, GCNET: graph-based prediction of stock price move-
ment using graph convolutional network, Eng. Appl. Artif. Intell. 116 (2022)
[1] O.B. Sezer, M.U. Gudelek, A.M. Ozbayoglu, Financial time series forecasting with 105452.
deep learning: A systematic literature review: 2005–2019, Appl. Soft Comput. 90 [29] S. Xiang, D. Cheng, C. Shang, Y. Zhang, Y. Liang, Temporal and heterogeneous
(2020) 106181. graph neural network for financial time series prediction, in: Proceedings of the
[2] M.M. Keskin, M. Yilmaz, A.M. Ozbayoglu, A deep neural network model for stock 31st ACM International Conference on Information & Knowledge Management,
investment recommendation by considering the stock market as a time graph, in: 2022, pp. 3584–3593.
2021 2nd International Informatics and Software Engineering Conference, IISEC, [30] W. Yin, Z. Chen, X. Luo, B. Kirkulak-Uludag, Forecasting cryptocurrencies’ price
IEEE, 2021, pp. 1–6. with the financial stress index: a graph neural network prediction strategy, Appl.
[3] R. Flanagan, L. Lacasa, Irreversibility of financial time series: a graph-theoretical Econ. Lett. 31 (7) (2024) 630–639.
approach, Phys. Lett. A 380 (20) (2016) 1689–1697.
[31] Ö. Ekmekcioğlu, M.Ç. Pınar, Graph neural networks for deep portfolio
[4] J. Long, Z. Chen, W. He, T. Wu, J. Ren, An integrated framework of deep
optimization, Neural Comput. Appl. 35 (28) (2023) 20663–20674.
learning and knowledge graph for prediction of stock price trend: An application
[32] A. Lazcano, P.J. Herrera, M. Monge, A combined model based on recurrent neural
in Chinese stock exchange market, Appl. Soft Comput. 91 (2020) 106205.
networks and graph convolutional networks for financial time series forecasting,
[5] Y. Long, Visibility graph network analysis of gold price time series, Phys. A 392
Mathematics 11 (1) (2023) 224.
(16) (2013) 3374–3384.
[33] H. Xu, Y. Zhang, Y. Xu, Promoting financial market development-financial stock
[6] E.J. Ruiz, V. Hristidis, C. Castillo, A. Gionis, A. Jaimes, Correlating financial
classification using graph convolutional neural networks, IEEE Access (2023).
time series with micro-blogging activity, in: Proceedings of the Fifth ACM
[34] J.M.-T. Wu, Z. Li, N. Herencsar, B. Vo, J.C.-W. Lin, A graph-based CNN-LSTM
International Conference on Web Search and Data Mining, 2012, pp. 513–522.
stock price prediction algorithm with leading indicators, Multimedia Syst. 29 (3)
[7] L. Wang, Z. Wang, S. Zhao, S. Tan, Stock market trend prediction us-
(2023) 1751–1770.
ing dynamical Bayesian factor graph, Expert Syst. Appl. 42 (15–16) (2015)
[35] D. Cheng, F. Yang, S. Xiang, J. Liu, Financial time series forecasting with
6267–6275.
[8] D. Xiao, J. Wang, Graph based and multifractal analysis of financial time series multi-modality graph neural network, Pattern Recognit. 121 (2022) 108218.
model by continuum percolation, Int. J. Nonlinear Sci. Numer. Simul. 15 (5) [36] C. Luo, T. Ma, M. Cucuringu, Spatial-temporal stock movement prediction and
(2014) 265–277. portfolio selection based on the semantic company relationship graph, 2024,
[9] Y. Xiu, G. Wang, W.K.V. Chan, Crash diagnosis and price rebound prediction Available at SSRN.
in NYSE composite index based on visibility graph and time-evolving stock [37] Z. Sun, A. Harit, A.I. Cristea, J. Wang, P. Lio, Money: Ensemble learning for
correlation network, Entropy 23 (12) (2021) 1612. stock price movement prediction via a convolutional network with adversarial
[10] J.R. Abrams, J. Celaya-Alcalá, D. Baldwin, R. Gonda, Z. Chen, Analysis of equity hypergraph model, AI Open 4 (2023) 165–174.
markets: A graph theory approach, Soc. Ind. Appl. Math. (2016) doi 10. [38] H. Oukhouya, H. Kadiri, K. El Himdi, R. Guerbaz, Forecasting international stock
[11] S. Jeon, B. Hong, V. Chang, Pattern graph tracking-based stock price prediction market trends: Xgboost, LSTM, LSTM-XGBoost, and backtesting XGBoost models,
using big data, Future Gener. Comput. Syst. 80 (2018) 171–187. Statist. Optim. Inf. Comput. 12 (1) (2024) 200–209.
[12] C.K.-S. Leung, R.K. MacKinnon, Y. Wang, A machine learning approach for stock [39] S. Carta, A. Corriga, A. Ferreira, A.S. Podda, D.R. Recupero, A multi-layer
price prediction, in: Proceedings of the 18th International Database Engineering and multi-ensemble stock trader using deep learning and deep reinforcement
& Applications Symposium, 2014, pp. 274–277. learning, Appl. Intell. 51 (2021) 889–905.
[13] H. Liu, B. Song, Stock price trend prediction model based on deep residual [40] I.E. Livieris, E. Pintelas, S. Stavroyiannis, P. Pintelas, Ensemble deep learning
network and stock price graph, in: 2018 11th International Symposium on models for forecasting cryptocurrency time-series, Algorithms 13 (5) (2020) 121.
Computational Intelligence and Design, ISCID, Vol. 2, IEEE, 2018, pp. 328–331. [41] H. Song, H. Choi, Forecasting stock market indices using the recurrent neural
[14] J.M.-T. Wu, Z. Li, N. Herencsar, B. Vo, J.C.-W. Lin, A graph-based CNN-LSTM network based hybrid models: Cnn-lstm, gru-cnn, and ensemble models, Appl.
stock price prediction algorithm with leading indicators, Multimedia Syst. (2021) Sci. 13 (7) (2023) 4644.
1–20. [42] F. Gu, Z. Jiang, J. Su, Application of features and neural network to enhance the
[15] Y. Wang, C. Zhang, S. Wang, S.Y. Philip, L. Bai, L. Cui, Deep co-investment performance of deep reinforcement learning in portfolio management, in: 2021
network learning for financial assets, in: 2018 IEEE International Conference on IEEE 6th International Conference on Big Data Analytics, ICBDA, IEEE, 2021,
Big Knowledge, ICBK, IEEE, 2018, pp. 41–48. pp. 92–97.

10
M. Yilmaz et al. Applied Soft Computing 163 (2024) 111847

[43] J.-S. Kim, S.-H. Kim, K.-H. Lee, Diversified adaptive stock selection using [47] X. Yu, W. Wu, X. Liao, Y. Han, Dynamic stock-decision ensemble strategy based
continual graph learning and ensemble approach, IEEE Access (2023). on deep reinforcement learning, Appl. Intell. 53 (2) (2023) 2452–2470.
[44] R. Du Plooy, P.J. Venter, A comparison of artificial neural networks and boot- [48] Y. Li, S. Jiang, Y. Wei, S. Wang, Take bitcoin into your portfolio: a novel
strap aggregating ensembles in a modern financial derivative pricing framework, ensemble portfolio optimization framework for broad commodity assets, Financ.
J. Risk Financ. Manag. 14 (6) (2021) 254. Innov. 7 (1) (2021) 63.
[45] Z. Hao, H. Zhang, Y. Zhang, Stock portfolio management by using fuzzy ensemble [49] Z. Zhou, Z. Song, T. Ren, L. Yu, Two-stage portfolio optimization integrating
deep reinforcement learning algorithm, J. Risk Financ. Manag. 16 (3) (2023) 201. optimal sharp ratio measure and ensemble learning, IEEE Access 11 (2022)
[46] K. He, Q. Yang, L. Ji, J. Pan, Y. Zou, Financial time series forecasting with the 1654–1670.
deep learning ensemble model, Mathematics 11 (4) (2023) 1054.

11