key: cord-0820049-fziyns10 authors: Thakkar, Ankit; Chaudhari, Kinjal title: Fusion in stock market prediction: A decade survey on the necessity, recent developments, and potential future directions date: 2020-08-26 journal: Inf Fusion DOI: 10.1016/j.inffus.2020.08.019 sha: 54123abc5d32cd47600e3f17e1533b53c0783f58 doc_id: 820049 cord_uid: fziyns10 Investment in a financial market is aimed at getting higher benefits; this complex market is influenced by a large number of events wherein the prediction of future market dynamics is challenging. The investors’ etiquettes towards stock market may demand the need of studying various associated factors and extract the useful information for reliable forecasting. Fusion can be considered as an approach to integrate data or characteristics, in general, and enhance the prediction based on the combinational approach that can aid each other. We conduct a systematic approach to present a survey for the years 2011–2020 by considering articles that have used fusion techniques for various stock market applications and broadly categorize them into information fusion, feature fusion, and model fusion. The major applications of stock market includes stock price and trend prediction, risk analysis and return forecasting, index prediction, as well as portfolio management. We also provide an infographic overview of fusion in stock market prediction and extend our work for other finely addressed financial prediction problems. Based on our surveyed articles, we provide potential future directions and concluding remarks on the significance of applying fusion in stock market. fusion, machine learning, deep learning The financial market is an attractive field of study; it offers a variety of opportunities to investors, market analysts, as well as researchers from various disciplines. The market participation perspectives may differ among individuals, for example, learning the market behaviours, deriving influencial aspects, 5 trading through stocks, predicting future market trend, recommending assets for portfolio management, etc., however, the lack of financial literacy and knowledge of fundamental economic principles can critically affect the investment returns [1] . Therefore, an individual's understanding and approach towards a financial market can determine the type as well as extent of information required to 10 study this domain. The economic market can be considered as a combination of financial investments, transactions, potential earning and/or losing, and several other actions that are performed at a massive level. It has extensively reached to a large number of fields and consequently, it also gets affected by numerous events; with this regard, the financial market can be interpreted as a model of 15 complex systems [2] . In financial markets, various instruments such as stocks, bonds, commodities, derivatives, currencies, etc. are explored, investigated, studied, and traded in an exhaustive manner; such trading is based on buying and selling of instruments. Stock market is a financial market where the new issues of stocks, 20 i.e., initial public offerings (IPOs), are created and sold at the primary market whereas the succeeding buying and selling are carried out at the secondary market [3] . The primary motivation behind investing in a stock market is to gain potential benefits of the investment [4] ; while careful tradings can earn higher returns, the associated risk may sometimes result into lose of valuables. The stock market accumulates the buying and selling of stocks based on 60 the ownership of traders. To analyze such markets, the historical stock market data as well as various factors can be combined to derive market patterns. The efficient market hypothesis (EMH) assumes that the available market information is incorporated within prices [9], however, the validation proof of EMH has been controversial over a period of time. On the other hand, the complex-65 ity levels of economic markets are defined to be the time-series nature, high cross-correlation with various entities, and collective market behaviour during extreme influencing events [2] . The behavioural finance has been studied with a primary focus on how psychology influences the behaviours of financial practitioners and its consequences 70 with respect to the stock markets [10, 11] . Such psychological biases can play a critical role in the market efficiency; therefore, an alternative theory namely, adaptive market hypothesis (AMH), has been proposed in contrast to the point of view where EMH assumes a frictionless market without imperfections [12] . AMH is developed based on the principles of evolutionary biology, i.e., compe- 75 tition, mutation, reproduction, and natural selection [13] ; such an evolutionary perspective of AMH is a reconciliation which argues that the financial markets evolve and its efficiency varies with time [12] . Even being an essentially qualitative and descriptive hypothesis, the arguments of AMH have been widely supported with return predictability of the stock market [14, 15] as well as for- 80 eign exchange rates [16] ; also, its potential implications have been evaluated for the digital currency [12] as well as cryptocurrency market [17] . This emphasizes on the understanding that amongst different factors which can influence the stock market, a careful selection of complementing and/or contradicting aspects may be fused to enhance the collected information and thus, improve the 85 predictions. The fusion techniques consist of different data sources, processed information, derived features, and/or prediction methods that are combined under appropriate scenarios. For a nebulous stock market, fusion can aid to extract knowledge base from a diverse set of data sources; it can extend the stength 90 of a model to overcome weakness of the other model; it can exploit the search space to derive effective solutions. Also, the dependency on specific market concepts can be explored with fusion. Therefore, to address and demonstrate the necessity of fusion in stock market prediction, this article presents a focused survey on the recent advances of fusion-based stock market prediction. The 95 primary motivation behind this survey is to identify how limitations of one or more entities are overcome with the others and to denote the importance of fusion for real-world stock market applications. To the best of our knowledge, our review work is the foremost survey that is devoted to fusion in stock market prediction and the same can serve as the foundation in this field. 100 In this article, we briefly categorize fusion into information fusion, feature fusion, and model fusion. With each fusion category, we explore the existing research work having primary applications of stock price and/or trend prediction, portfolio management, risk/return forecasting, as well as other stock market 105 concepts; though the reviewed articles are mainly based on fusion techniques, some of them might not have explicitly referred their work as a fusion approach. For collecting research articles for this survey, we applied a systematic strategy as follows. Using Google Scholar search engine, we conducted an initial search with terms "stock market prediction" and "fusion"; for a major concen- 110 tration on the recent studies, we restricted our search within years 2011 − 2020. We extended the search using additional terms such as "information fusion", "feature fusion", and "model fusion" whereas the potential applications related to stock markets were gathered using the terms "stock price prediction", "stock trend prediction", "portfolio management", "risk prediction", "return predic-115 tion", and replacing "prediction" by "forecasting" in each of these search operations, as well as using other financial market terms; to ensure that these terms were closely related to the actual work of the articles, we restricted our search such that articles having such terms in their titles were put together. We also set the exclusion criteria for articles resulting with the considered search operations; 120 in case of having the above-mentioned terms mainly occuring in the reference section, the articles were examined to find if the terms were being referred for the related work and not for the actual proposed approach; in such scenario, the articles were excluded from being considered in our survey. To minimize the redundancy, we eliminate similar articles from being reviewed, if found. On the 125 other hand, it was found that though specific fusion methods were applied in a set of articles, "fusion" term had not been specifically used to describe the same; therefore, we closely studied a large number of articles to identify whether they represented fusion at any level; thus, for an exhaustive survey, we advanced our collection by including such fusion-based articles as well as the articles related 130 to the ones downloaded using the aforementioned search strategies. It is observed that majority of the existing review articles have not precisely focused on fusion in stock market, however, they might have referred to fusionbased articles as a part of the survey. Therefore, we compare our survey with 135 the existing surveys even if they are partially related to the central focus of our survey. The computational approaches include various techniques that are adopted to predict stock market; artificial neural network (ANN) is one of the machine learning approaches that have been largely applied to stock market prediction. Some of the ANN-based models including a fusion model, as well as other techniques, were studied in Ref. [18] . On the other hand, a classifier fusion-based survey was carried out in study [19] where authors discussed the financial applications where the multiple classifier systems were integrated. A detailed study of fusion in opinion mining was conducted in review article [20]; along with [20] [21] [22] [23] [24] [25] Our survey nomics based on fusion approaches whereas the survey on multi-source knowl-150 edge fusion [23] , stock market prediction was discussed in the multi-modal-based aspect. For the quantitative financial applications, various machine learning techniques were reviewed in Ref. [24] . Subsequently, a fusion-based approach in financial text mining applications was covered under the survey on deep learning for financial applications [25] . Hence, we prepare a comparative analysis of our 155 survey with the existing surveys under various criteria associated with fusion and stock market as shown in Table 1 . It can be noted that amongst the existing surveys, a limited work has been regulated around fusion in stock market. In the presented survey, we take primitive steps of studying fusion perspectives based on stock market applications. Figure 1 ; our survey is primarily focused on the three fusion approaches and associated stock market prediction techniques. This presentation is a general overview of fusion in stock market prediction and the same can be extended 170 as well as modified as per the requirements. In our survey, the presented infographic is aimed to direct the readers of this survey through the wider scope of potential work based on fusion in the field of stock market; it can be considered as a walk-through summary to integrate fusion at various levels of operations completeness of understanding how specific fusion may be incorporated to an approach, the readers would be directed to refer to the study given by Ref. [7] . The organization of the remaining article can be summarized as follows: for the thorough coverage of all the criteria (C1−C9) mentioned in Table 1 , Section 180 2 reviews information fusion in stock market which includes fusion techniques based on raw data as well as processed data; Section 3 describes feature fusion in stock market and elaborates its potential applications; Section 4 presents model fusion in stock market including learning models as well as evolutionary models; Section 5 discusses various aspects of fusion in stock market and the 185 potential future directions; Section 6 provides the concluding remarks on the presented survey. The time-series stock market data present the market tradings; in its raw form, the historical data generally includes open, close, high, low, and volume The major concern around stock market analysis includes fundamental as 230 well as technical analyses; while the concerned features of such analyses differ, their fusion can be beneficial for a stronger information source while predicting the stock market. One of the data fusion techniques proposed to integrate economic and financial parameters for the fundamental analysis with the stock price change over time-based technical analysis [27] . As the stock price was 235 proven to have a lognormal distribution [28] , it was demonstrated that the logarithm of a stock price at the given time followed normal distribution with mean and variance values; such lognormal models were developed for financial assets such as stock prices [29] . Subsequently, authors in Ref. [27] inferred from the histogram of stock prices for the companies of Iran that the stock 240 price change followed normal distribution. An extended Kalman filter (EKF) was adopted to combine the analyses information; the predicted next-day stock price trend using data fusion as well as change in error using EKF structure indicated performance improvements as copared to regression and ANN models [27, 30] . On the other hand, the intuition that stocks with higher correlation could be affected by the same event was considered and an information fusion approach was applied to predict the stock trend in study [32] . Authors proposed to fuse quantitative, event-specific, and sentiment information, i.e., historical stock 265 prices, financial Web news-based events, and financial discussion board-based user sentiments, respectively. The collected data were categorized based on similar events and further combined from such heterogeneous sources to derive their internal relations. The correlations enhanced the shared knowledge-based learning and hence, improved the prediction performance. Based on the col-270 lected data from various sources, another fusion approach was proposed to predict the fluctuation trend of the given stock [33] ; using a tensor-based sub-mode coordinate (SMC) algorithm, the subspace dimensions were reduced according to the stock similarity. Authors proposed to fuse data collected from financial Web news, sentiments of the investors extracted from social media platform, and 275 quantitative data; the enhanced features were further given to long short-term memory (LSTM) model for the prediction. Hence, the diverse data sources can help for potentially fusing information to build a knowledge base that can be utilized to derive inherent market patterns. Another way of approaching the market sentiments is through social media were required for such prediction [34] . Considering that a specific lag period 290 represents strong correlation between the social sentiments and market prices, an SVM-based non-linear granger causality approach was proposed in study [35] . It fused the sentiments with historical stock market data for extracting a higher level of statistical significance at the particular lag period. The optimal lag periods were evaluated and daily stock prices were predicted for four such trade-off is considered to be a fundamental aspect in finance [37] . Though some investors may choose to keep a safe trading and therefore, accept lower return on behalf of low risk, investors having risk-averse nature may desire higher return and hence, take high risk with the trading. Thus, it is important to forecast the potential return, inherent risk, as well as relation between them 320 during stock valuation. Also, a considerable number of studies have been carried out to demonstrate the distributions of market returns such as fat-tailed and skewed return distribution [38, 39] . Such financial statistics can be utilized by the learning models in order to infer the characteristics of stock market assets; the role of information fusion can be examined to identify data distribution 325 and corresponding predictions can be evaluated. One of the primary concerns of an investment is to make an early-warning that can be helpful in preventing the financial crisis of an investor as well as the investing firm; an information fusion approach was proposed for such financial early-warning prediction in [40] . Authors adopted the Dempster-Shafer theory 330 (DST) that inferred based on an aggregation of independent information from different data sources to obtain a higher degree of belief. A total of 25 variables were selected to derive the financial crisis information and the prediction results were calculated using SVM as well as logistic models; these data were fused using DST to prepare the final rating of each company's probability of having 335 an early-warning. The market trends can be visualized using historical stock data; similarly, the market news can reflect various events, their impact on stock prices, discussions on corresponding analyses, as well as market behaviours. Such data from diverse perspectives can be fused to develop broader viewpoints that can be utilized for source. Subsequently, sensor data that provide timely measurements of various events, along with social media data, were considered to derive informative features of the given stock [42] ; for characterization of such time-series, the internal dependences were approximated using copula theory. Also, the social media data were assisted using Google trends and fusion was applied to infer 350 events for stock return prediction. While the investments in stock market are inherently risky, the logistics as well as lenders may provide a loan to the customers and their firms; the concern raises with the possibility of unability of the borrower to repay the lended money. Hence, it arises the potential credit risk; in order to prevent the large possibilities 355 of facing such issues, identification of various factors associated with the credit risk must be carried out before making an investment. An information fusion approach was proposed to forecast credit risk of the customers [43] ; the internal history of the customers such as financial and operational data, credit history as well as customer characteristics information were collected along with the 360 external dynamics such as industry index as well as relevant news. Such data were prepared into features and aligned for risk prediction. The improved results indicated the potential fusion of internal and external information for credit risk prediction and hence, for possible decision-making. A daily stock return was predicted using three models, adaptive neuro-fuzzy 365 inference system (ANFIS), ANN, and SVM in Ref. [44] ; here, an information fusion-based sensitivity analysis was carried out to evaluate relative importance of individual variables given as input to the prediction model so as to reduce uncertainty of models and increase the extracted knowledge. Considering a decision support required for investment in financial markets, future indices on short-term as well as long-term horizons were predicted using 380 regulatory disclosures in Ref. [45] . The proposed text mining approach collected ad-hoc announcements from the publishing service provider and integrated ma- While the information fusion is significantly applied to the price, trend, and 390 index forecasting, as well as risk and return prediction techniques, it has sparsely covered other financial applications. In stock market, a large number of events can have influence on the market etiquettes; apart from the direct correlations between similar or dependent companies, the indirect associations among various firms can also affect the stock 395 market. Thus, identification and study of such market events can be a critical task. A knowledge-based complex event processing system was proposed in Ref. [46] to monitor the stock market based on the available real-time stock events and background knowledge; here, the companies having joint-stocks were considered and their publicly available data were fused to prepare a knowledge base. Initially, this approch was employed for addressing user queries by extracting the company knowledge base and fusing it with event data streams [47] . J o u r n a l P r e -p r o o f Apart from various predictions about the future stock market price trend, the market status of a company can also be a significant interest of the traders. Ref. [48] was proposed to predict the one-month-ahead status about how promising 405 an industry would be; considering that the securities companies provided ratings to various industries, authors adopted the DST and presented the degree of belief regarding the investment value of an industry. Based on the historical ratings, predictions were made for each trading day to determine the most promising industries for the next month; also, the higher average rise rate was observed 410 for the predicted industries. It can be observed that various factors affect the market dynamics of a stock index; hence, development of a correlation among such influenctial factors can be helpful for future predictions. Ref. [49] was proposed with a focus on the Bayesian networks to predict future stock price movement direction; while the 415 graphical representation of these networks could be useful to visualize market etiquette, the multivariate nodes could also indicate associate risks; such beneficial approach was employed to predict and analyze the S&P market index [49] . The network topology was prepared using several factors such as interest rate, consumer price index, unemployment rate, money supply, housing start, and into higher equity profits. In this section, we review the stock-related forecasting applications that are addressed using information fusion. As observed in the reviewed articles, the historical stock data can be appended, combined, integrated, as well as fused 430 with the collected or derived information and further applied for the prediction task. and being trapped in local optima using ANN [53] . Therefore, it is essential to identify the appropriateness of a specific approach for the target stock market application. Also, the motivation to fuse specific data as well as the rationale behind integrating diversity may require higher attention. On the other hand, identification of the amount of information diversity required to obtain desired 460 prediction performance is critical; one of the potential limitations may be identification of an appropriate arrangement of the available data [54] such that the application of information fusion should enhance the prediction performance. It can also provide valuable features which can increase the prediction model's learning capabilities; hence, the steps following information fusion must also be carefully selected. Such research directives can be addressed to improve forecasting models. and/or operation is adopted with a specific input type and applied to the stock-495 related forecasting application(s) as given by Figure 1 . This section reviews the stock-related forecasting applications based on feature fusion. The analysis of market trends is important for decision-making; it determines whether money should be invested to a specific stock as well as the length individually, dimensions of these features were reduced using independent component analysis (ICA). Further, the extracted feature were fused using canonical correlation analysis (CCA) to derive union feature that was used for predicting 510 next day's close price. On the other hand, a minimal variation order weighted average (OWA) operator was employed to fuse price data periods in [56] ; here, authors proposed to multiply features with their corresponding weights and summarize as a single aggregated factor; clustering and fuzzy rules were combined to train ANFIS and future stock prices were predicted. in study [57] . Authors proposed to use the time-series stock market data along 520 with stock chart images; the prepared stock charts were evaluated for price prediction and combined with volume information to create fusion chart images. investor [59] . The maximization of the expected returns is a primary concern of portfolio management. For the diversified stocks of a portfolio, long-term management Hence, combinational fusion based on various attributes can also be extended for portfolio management. In a prediction model, features are crucial factors that define specific aspect of the input data and help in deriving useful information for the model to learn; turn, is likely to reduce the prediction performance; for example, nonlinearity of CCA [62] . Therefore, it is necessary to evaluate the technical aspects of the feature constructing approaches. It can also be noticed that stock market data are presented in different forms; the same may be extended at fusion level for 580 retrieving rich set of features that can improve the prediction accuracy. Information fusion and feature fusion are applied to stock market prediction; these approaches respectively focus on the data and derived attributes that can enhance the intrinsic perspectives. Another way to improve prediction Direct or fusion-based techniques can be applied on the data processing 615 as well as feature-related steps of a prediction approach and they are further given as inputs to the prediction model; one of the expected outcomes of a prediction model is reliable forecasting. Model fusion can be considered as an application to the stock price/trend prediction; this fusion may combine two or more architectures or it may apply modules of different prediction models in 620 order to fuse the capabilities of stock market prediction. A fusion model was proposed to predict the trend of stock close price in [64] ; authors considered ANFIS model to advance the capabilities of ANN and Sugeno fuzzy system where the membership function parameters were adjusted using clonal selection-based immune algorithm. Authors provided historical 625 stock market data along with technical indicators to the ANFIS model where the data served as antigen and the prepared fuzzy rules were treated as the antibody; comparison indicated performance improvement using the proposed fusion model. On the other hand, considering wavelet analysis as an effective approach to process data information at various scales, it was fused with two 630 models namely, ANN and ANFIS for predicting next-day close price in Ref. [65]. Authors adopted one dimensional discrete wavelet with level 3 and applied 25 J o u r n a l P r e -p r o o f denoising on the normalized historical stock data; ANN and ANFIS models, with and without wavelet analysis, were evaluated individually and the prediction results were compared. These approaches indicated the significance of fusion 635 over neural networks as well as fuzzy systems. A large number of historical stock market data are analyzed and various technical indicators can be derived; such data can be appended to prepare a set of feature that can be further utilized for stock market prediction. However, operations with such large amount of data having non-linear complex relation-640 ships may be difficult. An OWA approach was fused with ANFIS model in study [66] for stock price prediction; here, OWA was adopted to reduce the higher dimensional stock dataset into an aggregate value and fuzzy rules were created using ANFIS. The predicted stock price indicated higher profits as compared to previous work. On the other hand, a Bayesian-regularized feed-forward ANN 645 model was proposed in Ref. [67] to predict the one-day-ahead stock market trend; the prediction results indicated comparable mean absolute percentage error (MAPE) with one of the existing fusion models [68] . A two-stage fusion approach was proposed in Ref. [69] to predict future close price; authors proposed to integrate SVR in the first stage and fused SVR, 650 ANN, and random forest (RF) in the second stage that combined SVR-ANN, SVR-RF, and SVR-SVR models for prediction. Results indicated performance improvement with fused models as compared to single prediction model. Subsequently, for predicting the stock price, a blended model fusion approach was proposed in study [70] where the regression and backpropagation using ANN 655 algorithms were integrated to enhance the prediction. Another approach for the stock trend prediction integrated recurrent neural network (RNN) and representative pattern discovery (RPD) in study [71] ; the advantages of deriving short-term and long-term trends of time-series data using RNN and RPD, respectively, were fused to build an improved model. The rep- investments [74] and hence, diversify portfolios at domestic and international levels based on the geopolitical risks and financial stresses [75] to handle tur-680 bulences of economic conditions; other potential approach may diversify the investment styles using fundamental analysis and/or technical analysis; also, the concept of financial cognitive dissonance, i.e., changing the investment beliefs, may be adapted by the investors [76] . Stock market offers a veriety of assets; selecting beneficial assets that can 685 gain higher returns and hence, creating an optimal portfolio is a challenging task. Apart from the econometric models [77] , computational models can also be applied, as well as fused, to address portfolio management. One of such fusion approaches was proposed in Ref. [78] with SVM and mean-variance (MV) for portfolio selection. Authors evaluated cardinality of the selected portfolios 690 as well as associated risks and returns; the proposed approach (SVM + MV) indicated better performance as well as higher net profitability. The desire of earning higher profits of the investment can enforce to move towards risky assets. Hence, it is necessary to estimate the expected returns 695 and corresponding level of risk prior to an investment. Also, the duration of investment such as short-term, medium-term, or long-term, should be carefully chosen. While the investors worry about the returns of their investment in stocks, the lenders may consider the potential risk of not receiving the owed amount as well 700 as its interest; the credit risk concept can be related to this kind of a scenario of having possible loss in case the borrower fails to repay the borrowed amount. This becomes a critical factor that demands systematic study of the investment and potential credit risk. A fusion-based credit risk assessment approach was proposed in Ref. [79] ; authors proposed to apply an enhanced decision support 705 model (EDSM) using relevance vector machine and decision tree to interpret the forecasting results. Also, the effectivenes of various criteria was evaluated and the future measurement designs were examined with the help of corporate transparency and information disclosure feasibility. Such assessments can be helpful in forecasting credit risks for the investors and hence, taking a well-710 examined decision. It can be noticed that diversity is one of the major concerns while forecasting the stock market; such diversities are explored using various fusion techniques. For stock risk and return predictions, multiple diverse base classifiers were proposed to be fused for a common input data and a Meta classifier was developed to 715 learn from outputs of such individual classifiers [80] . Authors included Bagging, Boosting, and AdaBoost methods with various base classifiers and analyzed the prediction performance. The financial market offer a wide range of interesting concepts; an investor's predict EPS using the model fusion [83] . The sentiments associated with stock-related text are important attributes; an analytical approach was proposed with SVM and bootstrapping techniques to evaluate the stock sentiments [84] . Authors derived the stock review blogs using a web crawler and classified the emotions within the review text using SVM 735 and reconstructed the set using bootstrapping classifier. The results indicated performance improvement using such model fusion. One of the major concerns while investing in a company is its revenue; it indicates the turnover of a firm and hence, investors may look forward to the companies having a promising sale and therefore, higher returns of the investments. For predicting a company's future revenue, a model fusion was proposed in Ref. for example, curse of dimensionality problem with VAR [88] and ANFIS [89] ; overfitting because of complex RF trees [90] . Therefore, it becomes essential to reduce the potential limitations of the associated models in order to minimize the expansion of individual's drawback(s). As the model selection is crucial to 770 support the usefulness of pre-processing as well as feature selection/extraction steps, it is also required to choose appropriate operations in order to assist the forecasting model. A considerable amount of work provides the intuition towards fusing specific models. The same may be extended with information as well as feature fusion techniques in order to inherit exploration and exploitation 775 capabilities of the corresponding models. The model fusion may also be dependent on the input data types; hence, a generalized conceptualization may be developed to study how specific kinds of models should be fused for the given In the previous sections, different stock-related forecasting applications were covered under specific fusion categories, however, there are other extensions and enhancements on fusion approaches. It has been observed that fusion of predic-805 tion results could provide improved predictions [91, 92] . Hence, predictions and prediction models were treated as information and sources, respectively whereas the combinational approach was considered as fusion in Ref. [93] ; authors derived the sentiment scores based on such information fusion and analyzed the short-term performance of IPOs. The financial markets are association of a large number of inherent correlations; with a growth of the related trading strategies, the traders' actions also get inline with the market correlations which then get reflected in market dynamics as well [94] . Such correlations are important to study because the fragment of an overall component may result into market crash over a period of 815 time. Based on the prospect theory [95] and price fluctuations being considered as a dynamic process, Ref. [96] proposed to develop a multiparametric analysis framework for decision making in financial investments (MIAMI model) on short time-frames; the approaches including knowledge discovery, technical analysis, information fusion, and soft computing were applied along with fuzzification 820 such that the contributions were fused in energy and entropy decision variables to derive useful trend analysis. While fusion-fission approach was applied to study and predict a market crash [94] , a potential integration of such statistical approaches can be desired for higher-level of fusion. Subsequently, a combinational fusion analysis (CFA) was integrated for the task of portfolio management 825 in Ref. [97] ; as a combination of selecting appropriate attributes and specific approaches, the portfolio management was proposed with multiple algorithms, instead of single one, based on the diversity or performance strength and the results were given scores for deriving a combinational approach. On the other hand, as an application of individuals' sentiments on the market research, a fu-830 sion approach was proposed [98] ; while stock market data could be considered as a potential application, authors studied the impact of sentiment analysis-based fusion. Apart from the traditional fusion, an approach to inject the knowledge to derive the actual information was proposed in study [99] ; it was developed as 835 a defusing technique to infer the relevant base data from the aggregated historical data. Another application of fusion was given as a one-vs-one scheme with optimizing decision directed acyclic graph (ODDAG) to predict the listing status of companies [100] . One of the recent applications included the uncertain possibility-probability information fusion for stock selection based on the rela-840 tive closeness of the alternative stocks [101] . The hyperlinks on Twitter that 33 J o u r n a l P r e -p r o o f Journal Pre-proof referred to specific companies in stock market, i.e., cashtags, were integrated with natural language processing (NLP) and data fusion approaches in Ref. [102] ; authors prepared the company-specific corpora based on such collected tweets and classified whether the corresponding tweet was related to a particular The basic understanding of how fusion can be applied to various domains is an important task. It has been noticed that specification of the specific category of fusion and/or type of fusion technique is given in a limited number of articles; also, the words such as "fusion", "ensemble", and "hybrid" are interchangeably 860 used in some articles and hence, maintaining the principle conceptual meaning of the appropriate terminology becomes a crucial responsibility for future articles. While we have primarily considered the articles having closest explanation about fusion at some point in the prediction phase, other articles that have not been exclusively said to have used fusion, however, a part of it represents fusion 865 approach can be studied and explored. Various information fusion-based methods have considered quantitative, derived, and appended data. The data sources in a stock market play a critical role; while selection of such sources may differ based on the target application, a combination of static and active data sources [103] can also be carried out 870 on a dynamic level for information fusion. Also, the rationale behind choosing 34 J o u r n a l P r e -p r o o f a specific group of data can be advisable for more insights to the approach. One of the important concerns is perspective analysis; while perspective information fusion is applied to various domains [104], the same can be adopted for personalized recommendations on stock market applications. The data fusion 875 is expanded to information fusion, however, an exhaustive extraction can be carried out to prepare a knowledge base and integrate the same for knowledge fusion [105] for stock market. It can be observerd that comparatively a restricted amount of work has been obtained based on feature fusion; though these features come along the process-880 ing step with various attributes, an appropriate fusion of such data can enhance the prediction performance. One of the possibilities of adopting feature fusion can be the usage of image fusion techniques that can derive useful patterns from the stock charts [106] . Subsequently, appropriate combination of information fusion and feature fusion may be helpful in achieving higher-order knowledge 885 base as well as description for the given approach. The model fusion is presented in a broad term; other ways of applying related fusion include classifier fusion, architecture fusion, as well as decision fusion. Such fusion techniques are further classified as per the proposed outline of the article, for example, hierarchy of techniques that combine classifiers [107] . While the classification categories may 890 differ, unification of such wide spread classes and hence, showing a contextual presentation may be a potential future work. Subsequently, the integration of fusion at multiple levels, i.e., information, feature, model, etc. may widen the abilities of an approach to study complex market patterns from diverse perspectives. Also, the balanced approaches integrating economical aspects with 895 computational intelligence can increase the opportunities of reliable predictions; the dependency of such predictions for stock market data must be evaluated for an economic collapse such as the COVID-19 outbreak. Our survey primarily focuses on various stock market applications addressed using different fusion techniques. It can be observed that the amount of work [108, 109] can be consolidated for fusion; on the other hand, the computational intelligence [21, 85, 110] can also be supplemented with such models to increase the learning capabilities of fusion techniques. The considera-905 tion of various financial aspects can redirect the existing work towards a remarkable enhancement; one of such approaches is the financial behaviour which may be further adapted for cognitive analysis and potential stock market predictions. Hence, it is worth mentioning that data fusion, in general, has leveraged stock market predictions and has a prospective future scope. Fusion can take place at diverse components within a prediction process; 920 we elaborate the necessity of applying fusion in stock market and we consider the three kinds, namely, information fusion, feature fusion, and model fusion to cover its major aspects. The target area where fusion is integrated can decide on its impact on the overall prediction performance as well. We also study fusionbased stock market applications such as stock price/trend prediction, index fore- Debt paying ability, share index, profitability, operating capacity, capital structure [40] Customer history, characteristics, industry index Sensor-based events Investors' transaction records, public market information [36] Quantitative data, financial news, investor sentiments from social media platform Historical stock data, financial news-based events, financial discussion board-based user sentiments [32] Historical stock data Stock market price, Twitter sentiments [34] Technical indicators, Wikipedia hits Real-time stock events Stock market data, ad-hoc announcement-based sentiments consumer price index, unemployment rate, money supply, housing start Early fusion method, combinations of candlestick and bar charts, line and bar charts, filled line and bar charts Order weighted average operator-based aggregation value of attributes Canonical correlation analysis-based union of technical indicators Model Fusion Deep random subspace ensembles Recurrent neural network and representative pattern discovery [71] Relevance vector machine Regression, backpropagation using artificial neural network Order weighted average operator, adaptive neuro-fuzzy inference system Wavelet analysis, artificial neural network, adaptive neuro-fuzzy inference system Support vector regression, artificial neural network Bayesian-regularized feed-forward artificial neural network Adaptive neuro-fuzzy inference system, clonal selection-based immune algorithm [64] Support vector machine Autoregressive model, adaptive neuro-fuzzy inference system Bagging, Boosting Financial literacy and stock market participation Levels of complexity in financial markets CREST: Cross-Reference to Exchange-based Stock Trend Prediction using Long Short-Term Memory Investing in stocks: The influence of financial risk attitude and values-related money and stock market attitudes Computational intelligence and financial markets: A survey and future directions Technical, fundamental, and combined information for separating winners from losers, Pacific-Basin Finance A review of data fusion techniques Editorial on the special issue Hybrid and ensemble techniques in soft computing: recent advances and emerging trends The efficient market hypothesis: Review of specialized literature and empirical research Inefficient markets: An introduction to behavioural finance Handbook of Research on Behavioral Finance and Investment Strategies: Decision Making in the Financial Industry: Decision Making in the Financial Industry Adaptive market hypothesis and evolving predictability of bitcoin Reconciling efficient markets with behavioral finance: the adaptive markets hypothesis Stock return predictability and the adaptive markets hypothesis: Evidence from century-long US data, 975 Are stock markets really efficient? Evidence of the adaptive market hypothesis Exchange-rate return predictability and The adaptive market hypothesis in the high frequency cryptocurrency market A survey on stock market prediction techniques A survey of multiple classifier systems as hybrid systems Opinion mining and information fusion: a survey Soft computing techniques for stock market prediction: A literature survey A survey of data fusion in smart city applications 2019 IEEE Fourth International Con-1000 ference on Data Science in Cyberspace (DSC) Machine Learning for Quantitative Finance Applications: A Survey Deep learning for financial 1005 applications: A survey Sensor models and multisensor integration A new approach to forecasting stock price with EKF data fusion Why stock prices have a lognormal distribution Lognormal model for stock prices Forecasting Trend and Stock Price with Adaptive Extended Kalman Filter Data Fusion Stock market one-day ahead movement prediction using disparate data sources Improving stock market prediction via heterogeneous information fusion A tensor-based sub-mode coordinate algorithm for stock prediction Information fusion of stock prices and sentiment in social media using Granger causality 2019 IEEE Intl Conf 40 , Autonomic and Secure Computing, Intl Conf on Pervasive Intelligence and Computing, Intl Conf on Cloud and Big Data Computing, Intl Conf on Cyber Science and Technology Congress An integrated framework of deep learning and knowledge graph for prediction of stock price trend: An application in Chinese stock exchange market The risk return tradeoff in the long run Stable Paretian models in finance Fat-tailed and skewed asset return distributions: implications for risk management, portfolio selection, and 1050 option pricing Based on information fusion technique with data mining in the application of finance early-warning Enhancing quantitative intra-day stock 1055 return prediction by integrating both market news and stock prices information Social media data assisted inference with application to stock prediction Port customer credit risk prediction based on Internal and external information fusion A data analytic approach to 1065 forecasting daily stock returns in an emerging market Long-term stock index forecasting based on text mining of regulatory disclosures Proceedings of the 6th ACM International Conference on Distributed Event-Based Systems Knowledge-based processing of complex stock market events Belief fusion of predictions of industries in Chinas stock market Market analysis and trading strategies with Bayesian 1080 networks The application of AUV navigation based on adaptive extended Kalman filter Decision forest: Twenty years of research Advantage and drawback of support vector machine functionality Predicting the direction of stock market index movement using an optimized artificial neural network model Information fusion techniques for change 1095 detection from multi-temporal remote sensing images A feature fusion based forecasting model for financial time series Minimal variability OWA operator com-1100 bining ANFIS and fuzzy c-means for forecasting BSE index Forecasting stock prices with a feature fusion LSTM-CNN model using different representations of the same data Technical analysis and sentiment embeddings for market trend prediction Portfolio choice with illiquid assets Long-term portfolio management using attribute selection and combinatorial fusion The Omnibus model: a new model of data 1115 fusion? Discriminative extended canonical correlation analysis for pattern set matching Prediction of Listed Companies' Revenue Based 1120 on Model-Fused A fusion model integrating ANFIS and artificial immune algorithm for forecasting Indian stock market Stock price prediction using a fusion model of wavelet, fuzzy logic and ANN OWA-based ANFIS model for TAIEX forecasting A Bayesian regularized artificial neural network for stock market forecasting A fusion model of HMM, ANN and GA for stock market forecasting Predicting stock market index using fusion of machine learning techniques Financial market predic-1140 tion using hybridized neural approach 2017 18th Interna-1145 tional Conference on Parallel and Distributed Computing, Applications and Technologies (PDCAT) Combining the wisdom of crowds and technical analysis for financial market prediction using deep random subspace ensembles Diversification across economic sectors and implication on portfolio investments in Malaysia International portfolio diversification benefits: Cross-country evidence from a local perspective Geopolitical Risk and Trading Patterns of Foreign and Domestic Investors: Evidence from Korea Impact of behavioral finance in investment decisions 1160 and strategies-a fresh approach Portfolio choice problems Decision-1165 making for financial trading: A fusion approach of machine learning and portfolio selection Credit risk assessment and decision making by a fusion approach Fusion of multiple diverse predictors 1170 in stock market the Mediation Influence of Value Relevance of Accounting Information, Investment Decision and Dividend Policy on the Relationship Between Profitability and the CompanyS Value A fusion ANFIS model for forecasting EPS of leading industries in Taiwan Fusion anfis model based on ar for forecasting eps of leading industries Research and Implementation of SVM and Bootstrapping Fusion Algorithm in Emotion Analysis of Stock Review Texts A Comprehensive Survey on Portfolio Optimization, Stock Price and Trend Prediction using Particle Swarm Optimization Surveying stock market forecasting 1190 techniques-Part II: Soft computing methods An analysis of applications and possibilities of neural networks (fuzzy, logic and genetic algorithm) in finance and accounting Short-term inflation projections: A Bayesian vector autoregressive approach Adaptive neuro-fuzzy inference 1200 system: Overview, strengths, limitations, and solutions Improving random forests by neighborhood projection for effective text classification, Information Systems A comparative analysis of machine learning techniques for student retention management An investigation of data and text mining methods for real world deception detection Analyzing initial public offerings' short-term performance using decision trees and SVMs The Asian correction can be quantitatively 1215 forecasted using a statistical model of fusion-fission processes Prospect theory: An analysis of decisions under risk Multi indicator 1220 approach via mathematical inference for price dynamics in information fusion context Combining multiple algorithms for portfolio management using combinatorial fusion Computing (ICCI* CC) Fusion with sentiment scores for market research H-fuse: 1230 Efficient fusion of aggregated historical data SIAM International Conference on Data Mining One versus one multi-class classification fusion using optimizing decision directed acyclic graph for predicting listing 1235 status of companies An uncertain possibility-probability information fusion method under interval type-2 fuzzy environment and its application in stock selection A methodology for the 1240 resolution of cashtag collisions on Twitter-A natural language processing & data fusion approach Multi-data source fusion Prescriptive information fusion From data fusion to knowledge fusion A new method of feature fusion and its application in image recognition A survey of deci-1255 sion fusion and feature fusion strategies for pattern classification Stock market forecasting techniques: literature survey Stock market forecasting techniques: A survey Stock Market Forecasting Using Computational Intelligence: A Survey