Recommender Systems meet Finance: A literature review Dávid Zibriczky1 2 Abstract. The present work overviews the application of recom- on several external factors (like market returns, governmental reg- mender systems in various financial domains. The relevant literature ularizations, currency, etc.); furthermore, expert knowledge is nec- is investigated based on two directions. First, a domain-based cate- essary to judge which one is a good choice. In order to reduce the gorization is discussed focusing on those recommendation problems, risk of such a choice, users tend to formulate stricter expectations to where the existing literature is significant. Second, the application of these products than to conventional e-commerce ones, thus applying various recommendation algorithms and data mining techniques is a recommender system in financial domains is a challenging task. summarized. The purpose of this paper is providing a basis for fur- Users typically protect their personal data, which is especially true ther scientific research and product development in this field. for financial services, causing privacy risk issues in recommender systems [61, 17] and requiring more complex alternative personaliza- tion methods. As privacy issues are significant in financial services, 1 INTRODUCTION personal metadata and individual transactional data are often miss- Recommender Systems [63] are information filtering and decision ing, which causes user cold-start problem for recommender systems. supporting systems that present items in which the user is likely From a business prospective, a common challenge that several fi- to be interested in a specific context. We consider users the active nancial institutions are facing is the lack of an intelligent decision entities that perform interactions (e.g. viewing, purchasing, rating, support system [13]. As sales activities of financial products requires etc.) in the system. We call items the objects with which the user expert knowledge, recommender systems offer great benefits for fi- can interact (e.g. products, movies, songs, etc.). The parameter set- nancial services by either improving the efficiency of sales repre- ting that characterizes the environment (e.g. time, device, location) is sentatives or automatizing decision making process for the clients. defined as context; furthermore, we consider the actual preferences Therefore, a significant demand is observed for these decision sup- (e.g. filters, rules, item types) as constraints of the recommenda- port systems. tions. Both users and items can be described by metadata (e.g. age, In this literature review, we investigate the existing application of gender for users; genre, price for items). Recommender systems ap- recommender system techniques focusing on the financial domains. ply several data mining algorithms such as popularity-based meth- First, we perform domain-based categorization, distinguishing the ods, collaborative- [67] and content-based filtering [58], hybrid tech- most developed fields; then we discuss the applications in less de- niques [9], knowledge-based methods [79, 24] or case-based reason- veloped financial domains. Second, we summarize the most often ing [74] depending on the characteristics of the domain, the quality applied recommender system methods and additional techniques that of available data and the business goals. are indirectly used for recommendations. Recommendation services offer several level of personalization, starting from manually defined ”editorial picks” to complex context- aware hybrid solutions. Businesses often mix various types of 2 DOMAIN-BASED REVIEW carousels in the same page to cover diversified collection of recom- In our terminology, a financial domain is a specific area of finance mendations. Although the majority of the recommender algorithms that can be properly identified, modeled and developed based on its focuses on capturing user preferences, non-personalized techniques specific properties. For example, we consider stocks and portfolios can also be considered as building blocks of a complex service (e.g. as two different domains in this context, because in the first case an first carousel shows personalized recommendations, the second one individual stock should be recommended, but in the second one a contains the most popular items in the last week). composition of financial assets should be selected, which is a dif- Recommender systems appeared in the mid-1990s, however, they ferent recommendation scenario. Based on the work of Burke and are receiving significant attention since the Netflix Prize [3]. Nowa- Ramezani [10], a domain can be characterized by the following as- days, recommender systems are applied in a very broad scale of do- pects: (1) heterogeneity that captures the diversity of items’ prop- mains [48] such as movies (Netflix), books (Amazon) or music (Spo- erties in a domain, (2) churn that characterizes the level of novelty tify). Generally speaking, recommender systems are useful in any and expected lifespan of the items, (3) interaction style that describes domains, where a significant amount of choice exists in the system how the users are able to express their preference, (4) preference sta- and users are interested in just a small portion of items. bility that characterizes the degree of variation of user preferences Compared to the subjects of conventional recommender systems, over time, (5) risk that determines the expected tolerance of the users financial products usually require a long-term significant financial for false recommendations and (6) scrutability that refers to the de- commitment as their utility is not realized immediately depending mand for explanation of recommendations. 1 Department of Finance, Budapest University of Technology and Eco- In the following subsections, we propose a categorization of sci- nomics, Hungary, email: zibriczky@finance.bme.hu entific contribution in financial services considering these properties. 2 ImpressTV, Hungary, email: david.zibriczky@impresstv.com First, we introduce the applications in online banking systems and 3 we discuss two general-purpose multi-domain solutions. Second, we borrowed, the interest rate and the dates of payment. In this domain, walk through on well-defined financial products such as loans, insur- the recommendation problem is finding the right product of the loan ance policies and riders, real estate and stocks. Third, we introduce company for the borrower, which both satisfies his financial needs the standard portfolio selection problem and we discuss various tech- and will be likely to be paid back by the borrower. Felfernig et al. [25] niques of personalized asset allocation. Finally, we collect other less propose a real-time constraint-based recommender application that studied domains. supports sales between the representatives and consumers focusing on loan recommendation problem. Microfinance is a type of banking service that supports low- 2.1 Online banking and multi-domain solutions income individuals and groups, who would otherwise have no oppor- By the rapid growth of information technology, the banking industry tunity to borrow money. In the last couple of years, the peer-to-peer changed significantly in the last decade. With the spreading of on- (P2P) lending became popular, in which individuals or groups have line payment solutions in various devices, a massive online data flow opportunity to invest money by lending to another parties using a P2P appeared in bank systems centralizing data from multiple domains. lending marketplace. In this context, the recommendation task is to Banks are forced to change technologies that is capable to handle big find an appropriate pairing between the lenders and individuals who data and exploit business value from the massive information flow. need loans. Choo et al. [15] propose a maximum-entropy-based rec- Yahyapour [84] and Asosheh et al. [2] investigate the introduction of ommendation method to solve this problem using the dataset of Kiva recommender systems into Iranian banking system using Technology P2P lending marketplace. Lee et al. [43] also developed a solution Acceptance Model. Based on the results of their questionnaire, there for Kiva, using collaborative filtering techniques for finding a fair is a significant willingness to introduce such a solution in banking pairing of microfinance. Significant work is published by Guo et al. systems, which primarily depends on perceived ease of use, useful- [37], who formulate an instance-based credit risk assessment model ness and the bank’s attitude. for evaluating risk and return of each individual loan. San Miguel et In order to exploit the value of contextual information of transac- al. [66] introduce a P2P loan recommendation method via social net- tional data, Gallego and Huecas [30] and Vico and Huecas [81] devel- work. They design a data framework architecture, which is capable oped context-aware recommender prototypes. Based on credit card to integrate both public and private data dealing with privacy issues. using history and geolocation data, they implemented a clustering- Bhaskar and Subramanian [7] introduce an adaptive recommender based method that provides personalized recommendation about system that assists microfinance institutions. They discuss the impact money spending opportunities close to the user. They find high user and limitations of such a system in an Indian case study. satisfaction of using such a solution; however, they also consider Based on the properties of this domain, we argue that loans are less the importance of privacy issues. Fano and Kurth [22] introduce a heterogeneous; however, we distinguish between basic loan products concept of interactive management tool that assists in personal re- and microfinance solutions. We think that the churn rate for conven- source (money) allocation. For the optimization of this objective, tional products is low, but for microfinance is typically higher. The they propose an algorithm, which considers expenses, financial goals interaction type is explicit for both opportunities and the individual and time of attainment. Yu [86] introduces a prototype of online transactions are rare. We argue that the preference of a user is unsta- personal finance management tool, which is capable to provide in- ble, because the demand for loans can change by personal financial surance planning, asset allocation and investment recommendation. status. Loans are definitely risky products; therefore, the explanation Overall, a number of works are published for banking sector; how- of recommendations is required. ever, all of them seem to be non-production concept only. Felfernig et al. [27, 26] present two general-purpose knowledge- 2.3 Insurance based recommender systems with intelligent user interface, which can be flexibly applied on various financial products. The au- In the insurance domain, an insurance policy is a contract between thors prefer knowledge-based algorithms over the conventional the insurer and the insured (policyholder). For an initial payment collaborative- and content-based filtering, because they can be ap- (premium), the insurer takes obligation to pay compensation for in- plied more efficiently in multi-criteria-based financial decisions. For sured if loss caused by perils under the terms of policy. As standard those cases, when no results can be shown for a multi-constraint set- policies have little room for customization, insurance riders are in- ting, Felfernig and Stettinger [28] propose a constraint diagnosis and troduced to extend benefits that is purchased separately from the ba- repairing technique. sic policy. Both insurance policy and insurance rider can be the object Related to online banking and multi-domain solutions, the prod- of personalized recommendation problem. ucts are basically heterogeneous. The churn rate depends on the type Mitra et al. [52] discuss a high-level concept of recommending of items accessed by these systems; however, we consider it low in both insurance policies and riders. In their short paper, they sum- banking environment. As these solutions offer interactive user inter- marize the potential business benefits of introducing recommender faces, the interactions are explicit. We argue that the user preference systems in this domain. For insurance policy recommendation, Rah- is unstable, because it strongly depends on the actual goal of the user. man et al. [60] implemented a real-time web-based application. They These systems focus on money management and spending opportu- apply a case-based reasoning algorithm to support insurance sale nities, thus we identify high risk and significant demand for explana- agents to offer the most suitable policies for their clients. Another tion. real-time cloud- and web-based application was developed by Abbas et al. [1], which recommends health insurance policies. The system applies multi-attribute utility-based theory that finds the most simi- 2.2 Loan lar products to the preference of the user based various criteria (e.g. A loan is lending money from one entity (individual or organization) premium, co-pay, co-insurance, benefits). Life insurance recommen- to another one with specified conditions. Under a loan product, we dation problem is also investigated by Gupta and Jain [38]. Their mean a debt with a promissory note specifying the amount of money short paper discusses the application of association rule mining for 4 such problem focusing on cold-start problem; however, it does not on both dividends and the difference of selling-buying price. As stock publish empirical results or architectural description. Rokach et al. market can be volatile depending on economic events and market [65] investigate the main domains of recommender systems compar- news, the estimation of future profit (utility) is very challenging task. ing them to the insurance sector highlighting the main differences. Interpreting the recommendation problem in this context, those prof- In their work, they apply a basic item-to-item collaborative-filtering- itable stocks should be recommended to the investor that meet his based method as a possible solution for the recommending insurance risk-aversion preference and trading behavior. riders. Based on the study published by Rokach et al. [65], the insurance domain is quite small, the interactions are indirect and the attention 2.5.1 Non-personalized stock recommendation span of the users is low; therefore, the size and quality of available The application of decision support systems in stock market has sig- dataset is low. The items are typically complex, the constraints of nificant literature. Most of the contributions focus on improving the users are high; however, they have little expertise. We consider this accuracy of predicting future returns (or trends) [89, 47, 14], pro- domain homogeneous with low item churn rate. We think that the viding buy/sell signals [16, 83, 12] or introducing automatic trading user preference is more stable for insurance than loans; however, we solutions [19, 40]; however, majority of these papers ignore the per- also note that a user is likely not to be interested in a same product sonalization factor. Nonetheless, global ranking of available stocks after contracting one. Insurance products are less risky than loans, can be considered as non-personalized recommendations. A number but the demand for explanation is still high. of papers pointed out on the observation that groups have greater knowledge than individuals and they can provide better market pre- dictions, calling it the ”wisdom of crowds” [39, 80, 36]. Eickhoff 2.4 Real estate and Muntermann [20] present significant correlation between the Real estate is a property consisting of the land, its natural resources prediction power of stock analysts and a set of social media users. and the buildings on it. The purchase of real estate is a rare and ex- Stephan and Von Nitzsch [75] report that individuals cannot beat pensive transaction, which may be undertaken for investment or for the market substantially; however, inexperienced investors can take personal residence. Therefore, buyers pay special attention to find the benefits from online communities. Several works consider the ap- proper choice considering several various preferences, which leads to plication of natural language processing methods on financial news a multi-criteria decision problem. In this review, we primarily con- [70, 69, 32, 49] and social networks texts [64, 4]. A comprehensive sider real estate as a type of investment. review about techniques of opinion mining and sentiment analysis is The application of recommender systems in real estate domain published by Ravi and Ravi [62]. has relatively weak literature, relevant papers were presented in the last five years only. One of the most significant contribution is pub- lished by Yuan et al. [87]. They propose a combination of ontolog- 2.5.2 Personalized stock recommendation ical structure and case-based reasoning for real estate recommenda- In order to provide personalized recommendations, individual infor- tion problem; furthermore, they implement a web-based application mation is required about the investor; however, explicit user prefer- with map visualization interface. Daly et al. [18] introduce a trans- ences are not available in most of the cases. One way to overcome portation time calculator to extend conventional metadata of real es- this difficulty is providing a user interface, where investor can spec- tate. In their work, they also propose a method to find the trade-off ify his preferences. An early solution was implemented by Liu and between multi-criteria. Wang et al. [82] apply a simple similarity- Lee [45], which offers a set of features for analyzing and picking based collaborative-filtering method for personalized ranking of real stocks based on preferences specified by the investor. Yoo et al. [85] estate; however, their data were collected by questionnaires. Quan- propose a graphical user interface, which calculates personalized rec- titative and qualitative criteria for decision making is investigated ommendations based on Moving Average Convergence Divergence by Ginevičius et al. [33], who present a study about the application (MACD) indicator and user interactions. Seo et al. [72] introduce of recommender systems for real estate management. Another study a management tool that applies multiple agents to collect informa- is published by Kafi et al. [42], which discusses a ”fuzzification” tion about the stocks and provides stock recommendations based on method on the metadata of real estate and the implementation of their what the investor is holding. Chalidabhongse and Kaensar [11] de- solution, called Fuzzy Expert System. sign a framework, which uses stochastic technical indicator on stock As real estate can be described by the same well-defined features returns. The solution considers both explicit preferences and user in- (e.g. price, size, rooms), this domain is homogeneous. We argue that teractions for personalized recommendations. the churn rate of items is significant, because items usually become Some of the works assume that user attributes and individual user unavailable after a purchase. The interactions can be both implicit transactions are available in the data set. Yujun et al. [88] propose (e.g. browsing) and explicit (e.g. purchasing), we argue that browsing a stock recommender algorithm based on big order net inflow. They data is frequent but purchase events are quite rare. We consider the argue that using just big orders underscores low-valued stocks and preference of a user stable; however, it can change over the time in reduce computational requirement for advanced algorithms. They in- long term. Purchasing real estate is expensive and risky transaction, troduce a fuzzy-based method, which recommends stocks that were thus proper explanation is required. selected by similar users. Taghavi et al [78] propose a concept of classical recommender system for ranking stocks. In their work, they combine hybrid techniques with various information collector 2.5 Stocks agents. Although their concept is quite close to conventional recom- A stock is a type of security, which represents ownership in a com- mender systems in e-commerce, they do not publish empirical re- pany and claims on its assets, earnings and dividends. Stocks are sults. The application of standard collaborative-filtering methods is traded in stock market, where the prices are controlled by traders’ also investigated by Sayyed et al. [68]; however, they present a pre- bids (buy price) and offers (sell price). They are held to gain profit liminary concept only. 5 2.5.3 Characteristics of stock market allocation methods in terms of average yield and intra-list-diversity of portfolios. Garcia-Crespo et al. [31] and Gonzalez-Carrasco et al. Due to its variability over time, stock market is more difficult to char- [34] introduce a fuzzy model that transforms the ontology of investor acterize than previous domains. We argue that stocks are heteroge- (education, age, income, risk-aversion, etc.) and the ontology of port- neous, because they represent companies from various sectors. The folio (market risk, interest rate, liquidity, returns, etc.) to a unified churn rate is low, because companies leaves stocks exchange very bi-dimensional matrix, where dimensions are psychological and so- rarely. Considering bidding and trading transactions, the interaction cial behavior features. Portfolios are recommended based on the dis- style is rather implicit with very high volume. We argue that the user tance of investor and portfolio models. The authors also discuss the preference is unstable, because it is strongly driven by news and the architecture the solution and compare the value of applied accuracy ever changing global economy. Recommending stocks is very risky; measures with other domains. Beraldi et al. [5] present a decision therefore, a particular good explanation is required; however, it is a support system for assisting strategic asset allocation using stochas- quite challenging task. tic optimization method. In their solution, an investor can define his strategy by setting its parameters (initial cash, period, type of assets 2.6 Asset allocation and portfolio management and currency). Based on these criteria, portfolios are generated maxi- mizing the tradeoff between expected final wealth, Conditional Value A portfolio is a composition of finite number financial assets with at Risk and risk aversion parameter. The authors provide a detailed various weights. It is well observed phenomena, that diversification high-level architecture and performance measurement of their solu- reduces the risk of an investment, because the specific risk of each tion. component become insignificant; therefore, portfolios offers better risk-return tradeoff than individual stocks. The technique of portfolio composition is often called asset allocation. In this context, the rec- 2.6.3 Characteristics of portfolio management ommendation tasks are selecting assets and estimating their optimal As portfolios can contain various assets, the portfolio management is weights in portfolio meet individual preferences and risk-aversion. heterogeneous. Although the churn rate may vary by the type of do- mains, we consider it low, because the assets are purchased for long- 2.6.1 Modern Portfolio Theory term investment. On the other hand, portfolios are basically unique and they always change if reallocation is performed. Assuming an One of the most well-known portfolio selection model (Modern Port- interactive user interface, the interaction type is explicit, because in- folio Theory, MPT) was published by Markowitz [50]. His model vestors can specify both their preferences or the desired weight of can be interpreted as a two-step recommendation problem. First, assets in portfolios. The stability of user preference may vary over well-diversified portfolios offer the best risk-return tradeoff for ev- time, but it is less unstable than stock exchange, because portfolios ery risk level, these set of portfolios are the object of recommen- are typically composed for long-term investment. The risk of such dation. Second, an investor is modeled by his risk-aversion utility investment is still high and explanation is desired in this domain. function, which scores every investment opportunity based on risk and expected return. Investors select those portfolios that maximize his utility function. The practical drawback of this theoretical model 2.7 Other financial domains is finding efficient portfolios requires complex calculation and esti- In this subsection, we discuss the financial domains that have weak mating the individual utility function itself is challenging task. literature in recommender systems. We mention only the most signif- Based on MPT, several works are published for asset allocation icant differences in characteristics from the aforementioned domains. [73, 8]; however, the first concepts of automated solutions appears An emerging domain of investment opportunities is venture fi- in the early 2000s. Elton and Gruber [21] argues that investors of- nance. Venture capital is a type of private equity that is offered for ten make irrational decisions; therefore, automatized recommenda- startup companies as seed funding. This kind of investment is typ- tions are advantageous for preventing irrational portfolio selections. ically risky, but expects high returns on promising companies. As Sycara et al. [77] present an overview of the application of intel- companies typically need only a few rounds of funding, the item ligent agents in portfolio management. They highlight the speci- churn is high in this case. The goal in this domain is to find an advan- ficity of this domain such as heterogeneity of information, dynamic tageous matching between the venture capital firms and their invest- change of environment, time-dependency and cost-constraints. Sev- ment partners. Related to this problem, Stone et al. [76] published eral researchers extend MPT by fuzzy techniques for modeling risk- a relevant work focusing on the application of collaborative filter- aversion [90], estimating risk of portfolios [6] and composing opti- ing. They report that the domain is characterized by extremely sparse mal portfolios [23, 57]. For generating efficient portfolios, Nanda et long-tailed data, thus the efficient use of conventional recommender al. [55] integrate a stock clustering method, Raei and Jahromi [59] system methods is challenging. Continuing their work, Zhao et al. apply two types of multi-criteria decision methods. Although the [91] investigate diversification techniques in this field. The authors aforementioned works propose various type of sophisticated portfo- propose 5 algorithms for ranking startups and a quadratic portfolio lio weighting methods, they are just non-personalized models. weight optimization method considering risk-aversion levels. Stock fund is a fund that principally invests in stocks. The compo- 2.6.2 Personalized portfolio selection sition of stock fund is defined by fund manager focusing on a cer- tain sector or a level of risk. Due to its diversification level, stock Musto et al. [54, 71, 53] propose a case-based reasoning methodolo- funds are less risky than stocks; however, they often cannot be traded gies for asset allocation, which consider user metadata for personal- in stock market thus the amount of transactions is low. Matsatsinis ization. In their work, recommended portfolios are calculated based and Manarolis [51] introduce a hybrid application for stock fund rec- on what similar users selected applying various combining strategies. ommendation problem. To reduce the sparsity issues, they propose The authors provide empirical results of neighbor selection- and asset the combination of collaborative filtering and multi-criteria decision 6 analysis. Lacking individual real data on transactions, they evaluate management. Daly et al. [18] presents housing recommender system, the proposed model on simulated investment behavior. which considers not just the metadata of a home, but the transporta- Jannach and Bundgaard-Joergensen [41] apply knowledge-based tion opportunities to the user specified locations. A metadata-based techniques to design a web-based advisory tool to improve the com- solution for peer-to-peer lending is proposed by San Miguel et al. pleteness of a business plans. In this context, the personalization of [66]; however, it is different from the conventional content-based fil- related questions is considered as a type of recommendation prob- tering. The authors introduce a framework that capable to represent lem. The application also provides a summary of financials, level of user data in vector-based- and semantic user models. We conclude completeness and aggregated advices. The risk of recommendation that pure metadata-based methods are not typical in financial do- is low and the explanation is not critical in this case. mains. 3 METHOD-BASED REVIEW 3.3 Knowledge-based recommendation In this section, we categorize relevant scientific contributions based Knowledge-based recommender systems (KBRS) [79] focus on for- on the applied methodologies. First, we walk through the standard malizing the knowledge about a domain based on its specificity, recommendation methods such as collaborative-filtering, content- various constraints and ontology of items. The information about a based filtering, knowledge- and case-based recommender systems. user is usually collected by a knowledge acquisition interface, per- Second, we discuss various hybrid techniques and additional data sonalized recommendation is calculated based on the representa- mining and machine learning methods that indirectly applied for rec- tion of knowledge about the user and available items. The advan- ommendation problems in financial services. Further domain-related tage of knowledge-based methods is that the recommendations rely studies, architectures and user interface designs are not discussed in only on the domain-knowledge and constraints of the user prefer- this section. ences; furthermore, they are easy to be explained. On the other hand, the knowledge base itself should be built up and maintained, which can be a significant overhead in operating such an interactive deci- 3.1 Collaborative filtering sion support systems and the conflict should be resolved by heuris- One of the most often used technique in recommender systems is tics when there is no matching item based on the actual constraints collaborative filtering (CF) [67]. As this method require interactions [28]. As knowledge-based methods are able to handle complex user only, it can be applied in various domains. Collaborative filtering is preferences that is typical for financial domains, they can be poten- able to extract latent behavioral pattern in transactional data that can- tially effective solutions assuming that the knowledge acquisition in- not be modeled by metadata; therefore, collaborative filtering meth- terface is implemented and knowledge about the domain is acquired. ods usually have higher accuracy than metadata-based methods. On Felfernig et al. propose several solutions for recommending various the other hand, their efficiency strongly depends on the sparsity of financial products using constraint-based reasoning, which is a type data and the novelty of items (cold-start problem); furthermore, it is of knowledge-based methods [27, 25, 26]. KBRS is also applied for quite challenging to explain the output of CF algorithms, which is a personalizing questions of business plan analysis [41]. strong disadvantage for risky financial domains. Among collaborative filtering-based solutions, the majority of works apply item-based nearest-neighbor methods for recommend- 3.4 Case-based recommendation ing insurance riders [65], real estate, [82] and venture capital [76]. Case-based recommender systems (CBRS) [46, 74] apply case-based We also find preliminary concept of the application of similarity- reasoning (CBR) that solves the recommendation problem based on based recommendations for stock market [68]. Lee et al. [43] ap- old similar cases. A case is defined in various ways (like product ply matrix factorization for Bayesian personalized ranking in micro- description, user preference, search criteria and outcome of case). finance services. They propose a fairness-aware optimization with CBRS relies on the first two step of case-based reasoning, which stochastic gradient descent (SGD). A significant contribution is pub- is (1) retrieve that finds relevant old cases to the current case and lished by Zhao et al. [91], who propose five different collaborative- (2) reuse that applies the knowledge from relevant old cases. An ac- filtering methods for venture capital domain. CF is also applied in tual case of the user is defined by user profile data or via interactive several other hybrid methods; however, we discuss those in a later user interface. In order to find similar cases, similarity of attributes, section. collaborative patterns or knowledge of the domain are usually ap- plied. On one hand, CBRS can be used for complex problems and it provides explainable recommendations. Based on Musto et al. [53], 3.2 Content-based filtering CBR has better properties than collaborative filtering for financial do- Content-based filtering (CBF) [58] recommends items based on the mains. On the other hand, these methods require a significant amount metadata of items in user history and other available items; therefore, of data about the cases. this method requires metadata and individual interactions only. CBF In financial domains, we find a number of case-based recom- algorithms can cope with the cold start problem and their recom- mender systems. Rahman et al. [60] propose a CBR-based applica- mendations are easy to explain by meta words; however, the models tion for recommending insurance policies. Musto et al. [54, 71, 53] strongly rely on the quality of metadata and they are usually less ac- introduce case-based reasoning for portfolio recommendation. In curate than collaborative filtering methods. their works, the authors also propose a diversification technique for We find that the metadata-based recommendation problem is usu- weighting candidate solutions in revise step. Yuan et al. [87] intro- ally associated with multiple-criteria decision analysis (MCDA) [29]. duce a real estate recommender that combines case-based reasoning Due to the complexity of real estate selection problem, MDCA mod- and ontology of items. Guo et al. [37] applies instance-based method els are often applied in that field. Ginevičius et al. [33] propose a for peer-to-peer recommendation problem and employ kernel regres- model that handles quantitative and qualitative criteria for real estate sion to find similarity weights of instances in the past. 7 3.5 Hybrid methods a decent number of applications recommending both insurance poli- cies and riders. There are a few papers dealing with real estate recom- We consider the combination of the different decision support meth- mendation; a decent part of them is empirical study only. There is a ods as hybrid method [9]. Generally, hybrid recommenders benefit huge literature dealing with stock market. A significant part of pub- from the advantages of applied techniques, while their weaknesses lications focuses on predicting stock prices and providing buy/sell are reduced. Hybrid methods can be more precise than conventional signals; however, these methods are non-personalized. Several works models; however, the efficient implementation of such solutions can introduce interactive user interface for managing stocks, but only a be very difficult for complex problems. few number of papers propose machine learning methods for per- We find hybrid solutions that incorporate credit card transactions sonalized stock recommendation. We also find a significant literature in various domains to provide context-aware recommendations based for asset allocation. On the basis of modern portfolio theory, several on the location of the user [30, 81]. We argue that hybrid filter- methods are introduced to find efficient portfolios for various risk- ing is an efficient solution for cross-domain recommendation. An- aversion levels; however, the personalization is realized in selecting other hybrid application focuses on finding the most profitable stocks risk level only. Some of the works apply machine learning methods to at a right time based on the investor preference [78]. They apply compose personalized portfolios based on individual attributes. Fur- collaborative- and content-based filtering in algorithm level and so- thermore, we present promising applications of recommender sys- cial, economic and semantical agents in system level. CF and CBF tems for venture finance, stock funds and business plan-related ques- is also combined by Mitra et al. [52] for recommending insurance tionnaire. product and by Choo et al. [15] for microfinancing. In order to re- Several domains can be characterized by homogeneous products; duce sparsity issues for stock fund recommendation, Matsatsinis and however, we argue that stock exchange, portfolio management and Manarolis [51] propose a combination if collaborative filtering and multi-domain solutions are rather heterogeneous. The item churn rate multi-criteria decision analysis. is basically low among the financial domains, except for real estate, There are a few applications of association rule mining (ARM) where the offers are available until only one transaction by nature. [44] in financial domains. A web-based hybrid association rule min- Assuming that user interface is provided, the interaction style is ex- ing method is proposed for personalized recommendation of insur- plicit, otherwise implicit data or user profile metadata can be used ance products, which also deals with cold-start problem [38]. ARM only. We find that the preference stability is various in these do- is used in stock market for predicting trading-based relationships be- mains depending on individual financial status and the changes of tween stocks [56]. global market. As the object of recommendations are usually related to money spending transactions, we consider all financial domains; 3.6 Complementary methods therefore, the demand for proper explanation about the recommen- dations is significant. In this section, we also discuss additional complementary techniques Based on our method-based analysis, we conclude that collabora- that are integrated to conventional recommender methods. We find tive filtering is applied in various domains where the product itself that fuzzy methods are primarily introduced for stock market and is well-defined; however, it is limited to handle complex recommen- asset allocation. Yujun et al. [88] introduce a fuzzy-based cluster- dation problems. We find a small number of applications using pure ing for stock recommendations. A fuzzy-based transformation is in- content-based filtering. Due to the specificity of financial domains, troduced by Garcia-Crespo et al. [31] and Gonzalez-Carrasco et al. multiple-criteria decision analysis and case-based reasoning has sig- [34] for portfolio recommendation problem. Fuzzy-based expert sys- nificant advantage over collaborative- and content-based filtering. tems are proposed for real-estate- [42] and portfolio recommenda- Assuming that a well designed user interface is available, knowledge- tions [23, 35]. Several variations of fuzzy-based extensions of mod- based methods has great benefits for assisting personalization prob- ern portfolio theory are introduced [90, 6, 57]. lems. We find several hybrid methods combining collaborative- and We find applications of artificial neural networks (ANN) for de- content-based filtering, we argue that application of association rules signing trading decision support systems [16] and extracting infor- is less significant. Investigating other methods, we find that fuzzy mation from news [32]. In stock price forecasting, semantic methods techniques are basically applied for portfolio selection problem; fur- are also considered for processing web texts [70] and emotions ex- thermore, artificial neural networks and support vector machines are pressed in Twitter messages [64]. Based on our research, classifica- typically used in stock market decision systems. tion methods are usually applied for stock markets. Support vector Summarizing our work, we state that an extensive work is being in machines (SVM) are used for incorporating information from finan- progress for investigating applications of recommendation systems in cial news [69, 49], forecasting stock returns [89, 47] and providing financial services; however, there remain several unexploited oppor- stock buy/sell signals [83]. tunities in this field for both scientific research and product develop- ment. 4 CONCLUSION ACKNOWLEDGEMENTS In this review, we have discussed the scientific contributions that were addressed to the recommendation problems in financial ser- I am grateful to ImpressTV for funding this research and support- vices in the last 15 years. We have performed a two-way investigation ing me by flexible work hours. I thank Mihály Ormos from Budapest based on financial domains and applied recommendation techniques. University of Technology and Economics for giving valuable advices Considering the domains, our finding is the following. Banking in- for literature review. Least but not least, I am thankful to the orga- stitutes have a significant willingness to introduce decision support nizers of 2nd International Workshop on Personalization and Rec- systems; however, we find just concepts for that problem. There is ommender Systems in Financial Services for extending the deadline a great support for personalizing peer-to-peer lending than conven- of submission, providing valuable feedbacks and encouraging me to tional loan services. Although insurance domain is small, we find finish this paper. 8 REFERENCES Zanker, ‘Constraint-Based Recommender Systems’, chapter Con- straint, 161–190, Springer US, Boston, MA, (2015). [25] Alexander Felfernig, K Isak, K Szabo, and P Zachar, ‘The VITA finan- cial services sales support environment’, Proceedings of the National [1] Assad Abbas, Kashif Bilal, Limin Zhang, and Samee U Khan, ‘A cloud Conference on Artificial Intelligence, 2(Felfernig), 1692–1699, (2007). based health insurance plan recommendation system: A user centered [26] Alexander Felfernig, Michael Jeran, Martin Stettinger, Thomas Ab- approach’, Future Generation Computer Systems, 4344, 99–109, (feb senger, Thomas Gruber, Sarah Haas, Emanuel Kirchengast, Michael 2015). Schwarz, Lukas Skofitsch, and Thomas Ulz, ‘Human Computation [2] Abbas Asosheh, Sanaz Bagherpour, Nima Yahyapour, and A Asosheha, Based Acquisition Of Financial Service Advisory Practices’, Organi- ‘Extended Acceptance Models for Recommender System Adaption, zational Support, 27, (2015). Case of Retail and Banking Service in Iran’, WSEAS Transactions on [27] Alexander Felfernig and A Kiener, ‘Knowledge-based interactive sell- Business and Economics, 5(5), 189–200, (2008). ing of financial services with FSAdvisor’, Proceedings of the National [3] James Bennett and Stan Lanning, ‘The netflix prize’, in Proceedings of . . . , 100, 1475–1482, (2005). KDD cup and workshop, volume 2007, p. 35, (2007). [28] Alexander Felfernig and Martin Stettinger, ‘Conflict Management in [4] Janek Benthaus and Roman Beck, ‘It’s more about the Content than Interactive Financial Service Selection’, Organizational Support, 3, the Users! The Influence of Social Broadcasting on Stock Markets’, in (2015). European Conference on Information Systems, (2015). [29] José Figueira, Salvatore Greco, and Matthias Ehrgott, Multiple criteria [5] P Beraldi, A Violi, and F De Simone, ‘A decision support system for decision analysis: state of the art surveys, volume 78, Springer Science strategic asset allocation’, Decision Support Systems, 51(3), 549–561, & Business Media, 2005. (jun 2011). [30] Daniel Gallego and Gabriel Huecas, ‘An empirical case of a context- [6] J D Bermudez, J V Segura, and E Vercher. A fuzzy ranking strategy for aware mobile recommender system in a banking environment’, Pro- portfolio selection applied to the Spanish stock market, 2007. ceedings - 2012 3rd FTRA International Conference on Mobile, Ubiq- [7] Tarun Bhaskar and Gopi Subramanian, ‘Loan recommender system for uitous, and Intelligent Computing, MUSIC 2012, 13–20, (2012). microfinance loans: Increasing efficiency to assist growth’, J Financ [31] Ángel Garcı́a-Crespo, José Luis López-Cuadrado, Israel González- Serv Mark, 15(4), 334–345, (mar 2011). Carrasco, Ricardo Colomo-Palacios, and Belén Ruiz-Mezcua, ‘SIN- [8] Michael J Brennan, Eduardo S Schwartz, and Ronald Lagnado, ‘Strate- VLIO: Using semantics and fuzzy logic to provide individual invest- gic asset allocation’, Journal of Economic Dynamics and Control, ment portfolio recommendations’, Knowledge-Based Systems, 27, 103– 21(89), 1377–1403, (jun 1997). 118, (mar 2012). [9] Robin Burke, ‘Hybrid recommender systems: Survey and experiments’, [32] Tomer Geva and Jacob Zahavi, ‘Empirical evaluation of an automated User modeling and user-adapted interaction, 12(4), 331–370, (2002). intraday stock recommendation system incorporating both market data [10] Robin Burke and Maryam Ramezani, ‘Matching Recommendation and textual news’, Decision Support Systems, 57, 212–223, (jan 2014). Technologies and Domains’, Recommender Systems Handbook, 54, [33] Tomas Ginevičius, Jurgita Alchimovien, Paulius Kazokaitis, and Artras 367–386, (2011). Kaklauskas, ‘Recommender System for Real Estate Management’, Ver- [11] T H Chalidabhongse and C Kaensar. A Personalized Stock Recommen- slas: teorija ir praktika, (3), 258–267, (2011). dation System using Adaptive User Modeling, 2006. [34] Israel Gonzalez-Carrasco, Ricardo Colomo-Palacios, Jose Luis Lopez- [12] Ya-Wen Chang Chien and Yen-Liang Chen, ‘Mining associative Cuadrado, Ángel Garca-Crespo, and Belén Ruiz-Mezcua, ‘PB- classification rules with stock trading data A GA-based method’, ADVISOR: A private banking multi-investment portfolio advisor’, In- Knowledge-Based Systems, 23(6), 605–614, (aug 2010). formation Sciences, 206, 63–82, (nov 2012). [13] Hilary Cheng, Yi-Chuan Lu, and Calvin Sheu, ‘An ontology-based [35] M Hassani Goodarzi, ‘A web-based Implementation of a Portfolio Ad- business intelligence application in a financial knowledge management visor System based on Fuzzy Expert Systems’, in Conf. Rec. IEEE 6th system’, Expert Systems with Applications, 36(2, Part 2), 3614–3622, Int. Conf. Information & Communication Technology and System ICTS, (mar 2009). Surabaya, Indonesia, pp. 15–22, (2010). [14] Vincent Cho, ‘MISMIS A comprehensive decision support system for [36] Jörg Gottschlich and Oliver Hinz, ‘A decision support system for stock stock market investment’, Knowledge-Based Systems, 23(6), 626–633, investment recommendations using collective wisdom’, Decision Sup- (aug 2010). port Systems, 59, 52–62, (mar 2014). [15] Jaegul Choo, Daniel Lee, Bistra Dilkina, Hongyuan Zha, and Haesun [37] Yanhong Guo, Wenjun Zhou, Chunyu Luo, Chuanren Liu, and Hui Park, ‘To Gather Together for a Better World: Understanding and Lever- Xiong, ‘Instance-based credit risk assessment for investment decisions aging Communities in Micro-lending Recommendation’, in Proceed- in P2P lending’, European Journal of Operational Research, 249(2), ings of the 23rd International Conference on World Wide Web, WWW 417–426, (mar 2016). ’14, pp. 249–260, New York, NY, USA, (2014). ACM. [38] Abdhesh Gupta and Anwiti Jain, ‘Life Insurance Recommender Sys- [16] Seng-Cho Timothy Chou, Chau-Chen Yang, Chi-Huang Chan, and tem Based on Association Rule Mining and Dual Clustering Method Feipei Lai. A rule-based neural stock trading decision support system, for Solving Cold-Start Problem’, International Journal of Advanced 1996. Research in Computer Science and Software Engineering, 3(7), 1356– [17] Lorrie Faith Cranor, ‘I didn’t buy it for myself’, in Designing person- 1360, (2013). alized user experiences in eCommerce, 57–73, Springer, (2004). [39] Shawndra Hill and Noah Ready-Campbell, ‘Expert Stock Picker: The [18] Elizabeth M Daly, Adi Botea, Akihiro Kishimoto, and Radu Marinescu, Wisdom of (Experts in) Crowds’, International Journal of Electronic ‘Multi-criteria Journey Aware Housing Recommender System’, in Pro- Commerce, 15(3), 73–102, (2011). ceedings of the 8th ACM Conference on Recommender Systems, RecSys [40] Yong Hu, Kang Liu, Xiangzhou Zhang, Lijun Su, E W T Ngai, and ’14, pp. 325–328, New York, NY, USA, (2014). ACM. Mei Liu, ‘Application of evolutionary computation for rule discovery in [19] L Dymova, P Sevastianov, and K Kaczmarek, ‘A stock trading ex- stock algorithmic trading: A literature review’, Applied Soft Computing, pert system based on the rule-base evidential reasoning using Level 36, 534–551, (nov 2015). 2 Quotes’, Expert Systems with Applications, 39(8), 7150–7157, (jun [41] Dietmar Jannach and Uffe Bundgaard-Joergensen, ‘SAT: A Web-Based 2012). Interactive Advisor for Investor-Ready Business Plans.’, in ICE-B, eds., [20] Matthias Eickhoff and Jan Muntermann, ‘Stock Analysts Vs. The Joaquim Filipe, David A Marca, Boris Shishkov, and Marten van Sin- Crowd: A Study on Mutual Prediction’, (2015). deren, pp. 99–106. INSTICC Press, (2007). [21] Edwin J Elton and Martin J Gruber, ‘The Rationality of Asset Allo- [42] Amir Hossein Kafi, Kazemipoor Hamid, and Mohammad Ali Afshar cation Recommendations’, The Journal of Financial and Quantitative Kazem, ‘Design and Implementation of Fuzzy Expert System for Real Analysis, 35(1), 27–41, (2000). Estate Recommendation’, International Journal of Information, Secu- [22] Andrew Fano and Scott W Kurth, ‘Personal Choice Point: Helping rity and Systems Management, 2(1), 142–147, (jan 2013). users visualize what it means to buy a BMW’, Control, 46–52, (2003). [43] Eric L Lee, Jing-kai Lou, Wei-ming Chen, Yen-chi Chen, Shou-de Lin, [23] Mehdi Fasanghari and Gholam Ali Montazer, ‘Design and implemen- Yen-sheng Chiang, and Kuan-ta Chen, ‘Fairness-Aware Loan Recom- tation of fuzzy expert system for Tehran Stock Exchange portfolio rec- mendation for Microfinance Services’, 1–5, (2014). ommendation’, Expert Systems with Applications, 37(9), 6138–6147, [44] Weiyang Lin, Sergio A Alvarez, and Carolina Ruiz, ‘Efficient adaptive- (sep 2010). support association rule mining for recommender systems’, Data min- [24] Alexander Felfernig, Gerhard Friedrich, Dietmar Jannach, and Markus 9 ing and knowledge discovery, 6(1), 83–105, (2002). sentiment’, Proceedings - IEEE International Conference on Data Min- [45] N K Liu and K K Lee, ‘An intelligent business advisor system for stock ing, ICDM, 21–26, (2007). investment’, Expert Systems, 14(3), 129–139, (1997). [71] Giovanni Semeraro and Cataldo Musto, ‘Personalized Wealth Manage- [46] Fabiana Lorenzi and Francesco Ricci, ‘Case-based Recommender Sys- ment through Case-Based Recommender Systems’, 2–4, (2014). tems: A Unifying View’, in Proceedings of the 2003 International Con- [72] Young-Woo Seo, Joseph A Giampapa, and Katia Sycara, ‘Financial ference on Intelligent Techniques for Web Personalization, ITWP’03, news analysis for intelligent portfolio management’, (2004). pp. 89–113, Berlin, Heidelberg, (2005). Springer-Verlag. [73] William F Sharpe, ‘Asset allocation: Management style and perfor- [47] Chi-Jie Lu, Tian-Shyug Lee, and Chih-Chou Chiu, ‘Financial time se- mance measurement’, The Journal of Portfolio Management, 18(2), 7– ries forecasting using independent component analysis and support vec- 19, (1992). tor regression’, Decision Support Systems, 47(2), 115–125, (may 2009). [74] Barry Smyth, ‘Case-based recommendation’, in The adaptive web, [48] Jie Lu, Dianshuang Wu, Mingsong Mao, Wei Wang, and Guangquan 342–376, Springer, (2007). Zhang, ‘Recommender system application developments: A survey’, [75] Philipp Stephan and Rüdiger von Nitzsch, ‘Do individual investors’ Decision Support Systems, 74, 12–32, (jun 2015). stock recommendations in online communities contain investment [49] Chao Ma and Xun Liang. Online mining in unstructured financial in- value?’, Financial Markets and Portfolio Management, 27(2), 149–186, formation: An empirical study in bulletin news, 2015. (2013). [50] Harry Markowitz, ‘Portfolio selection’, The journal of finance, 7(1), [76] Thomas Stone, Weinan Zhang, and Xiaoxue Zhao, ‘An empirical study 77–91, (1952). of top-n recommendation for venture finance’, Proceedings of the 22nd [51] Nikolaos F Matsatsinis and Eleftherios A Manarolis, ‘New Hybrid Rec- ACM international conference on Conference on information & knowl- ommender Approaches: An Application to Equity Funds Selection’, in edge management - CIKM ’13, 1865–1868, (2013). Algorithmic Decision Theory, 156–167, Springer, (2009). [77] K P Sycara, D Zeng, and K Decker, ‘Intelligent Agents in Portfo- [52] Sanghamitra Mitra, Nilendra Chaudhari, and Bipin Patwardhan, ‘Lever- lio Management’, in Agent Technology SE - 14, eds., NicholasR. Jen- aging Hybrid Recommendation System In Insurance Domain’, 3(10), nings and MichaelJ. Wooldridge, 267–281, Springer Berlin Heidelberg, (2014). (1998). [53] Cataldo Musto, Giovanni Semeraro, Pasquale Lops, Marco de Gemmis, [78] Mona Taghavi, Kaveh Bakhtiyari, and Edgar Scavino, ‘Agent-based and Georgios Lekkas, ‘Personalized finance advisory through case- computational investing recommender system’, Proceedings of the 7th based recommender systems and diversification strategies’, Decision ACM conference on Recommender systems - RecSys ’13, 455–458, Support Systems, 77, 100–111, (sep 2015). (2013). [54] Cataldo Musto, Giovanni Semeraro, Pasquale Lops, and Marco De [79] Shari Trewin, ‘Knowledge-based recommender systems’, Encyclope- Gemmis, ‘Financial Product Recommendation through Case-based dia of library and information science, 69(Supplement 32), 180, (2000). Reasoning and Diversification Techniques’, (2014). [80] M Velic, T Grzinic, and I Padavic. Wisdom of crowds algorithm for [55] S R Nanda, B Mahanty, and M K Tiwari, ‘Clustering Indian stock mar- stock market predictions, 2013. ket data for portfolio management’, Expert Systems with Applications, [81] Daniel Gallego Vico and Gabriel Huecas, ‘Generating Context-aware 37(12), 8793–8798, (dec 2010). Recommendations using Banking Data in a Mobile Recommender Sys- [56] P Paranjape-Voditel and U Deshpande. An Association Rule Mining tem’, ICDS 2012 : The Sixth International Conference on Digital Soci- Based Stock Market Recommender System, 2011. ety Generating, 73–78, (2012). [57] Preeti Paranjape-Voditel and Umesh Deshpande, ‘A stock market port- [82] Lei Wang, Xiaowei Hu, Jingjing Wei, and Xingyu Cui, ‘A Collabo- folio recommender system based on association rule mining’, Applied rative Filtering Based Personalized TOP-K Recommender System for Soft Computing, 13(2), 1055–1063, (feb 2013). Housing’, in Proceedings of the 2012 International Conference of Mod- [58] Michael J Pazzani and Daniel Billsus, ‘Content-based recommendation ern Computer Science and Applications SE - 74, ed., Zhenyu Du, vol- systems’, in The adaptive web, 325–341, Springer, (2007). ume 191 of Advances in Intelligent Systems and Computing, 461–466, [59] Reza Raei and M Jahromi, ‘Portfolio optimization using a hybrid of Springer Berlin Heidelberg, (2013). fuzzy ANP, VIKOR and TOPSIS’, Management Science Letters, 2(7), [83] Qinghua Wen, Zehong Yang, Yixu Song, and Peifa Jia, ‘Automatic 2473–2484, (2012). stock decision support system based on box theory and SVM al- [60] S S A Rahman, A A Norman, and K J Soon, ‘MyINS: A CBR e- gorithm’, Expert Systems with Applications, 37(2), 1015–1022, (mar Commerce Application for Insurance Policies’, Electronic Commerce 2010). Research, 5(1), 373–380, (2006). [84] Nima Yahyapour, ‘Determining Factors Affecting Intention to Adopt [61] Naren Ramakrishnan, Benjamin J Keller, Batul J Mirza, Ananth Y Banking Recommender System’, Technology, (2008). Grama, and George Karypis, ‘Privacy risks in recommender systems’, [85] Jungsoon Yoo, Melinda Gervasio, and Pat Langley, ‘An adaptive stock IEEE Internet Computing, 5(6), 54, (2001). tracker for personalized trading advice’, IUI ’03: Proceedings of the [62] Kumar Ravi and Vadlamani Ravi, ‘A survey on opinion mining and 8th international conference on Intelligent user interfaces, 197–203, sentiment analysis: Tasks, approaches and applications’, Knowledge- (2003). Based Systems, 89, 14–46, (nov 2015). [86] Chien-Chih Yu, ‘Designing A Consumer-Oriented Intelligent Decision [63] Francesco Ricci, Lior Rokach, and Bracha Shapira, Introduction to rec- Support System For Personalized Financial Planning Services’, (2004). ommender systems handbook, Springer, 2011. [87] Xiaofang Yuan, Ji-Hyun Lee, Sun-Joong Kim, and Yoon-Hyun Kim, [64] Marten Risius, Fabian Akolk, and Roman Beck, ‘Differential Emotions ‘Toward a user-oriented recommendation system for real estate web- and the Stock Market-The Case of Company-specific Trading’, (2015). sites’, Information Systems, 38(2), 231–243, (apr 2013). [65] Lior Rokach, Guy Shani, Bracha Shapira, Eyal Chapnik, and Gali Si- [88] Yang Yujun, Li Jianping, and Yang Yimei, ‘An Efficient Stock Rec- boni, ‘Recommending insurance riders’, Proceedings of the 28th An- ommendation Model Based on Big Order Net Inflow’, Mathematical nual ACM Symposium on Applied Computing - SAC ’13, 253, (2013). Problems in Engineering, 2016, (2016). [66] Beatriz San Miguel, Jose M del Alamo, and Juan C Yelmo, ‘A Per- [89] Zhi-yong Zhang, Chuan Shi, Su-lan Zhang, and Zhong-zhi Shi, ‘Stock sonal Data Framework for Exchanging Knowledge about Users in New Time Series Forecasting Using Support Vector Machines Employing Financial Services’, Organizational Support, 19, (2015). Analyst Recommendations’, in Advances in Neural Networks - ISNN [67] Badrul Sarwar, George Karypis, Joseph Konstan, and John Riedl, ‘Item- 2006 SE - 66, eds., Jun Wang, Zhang Yi, JacekM. Zurada, Bao-Liang based collaborative filtering recommendation algorithms’, in Proceed- Lu, and Hujun Yin, volume 3973 of Lecture Notes in Computer Science, ings of the 10th international conference on World Wide Web, pp. 285– 452–457, Springer Berlin Heidelberg, (2006). 295. ACM, (2001). [90] Zili Zhang and Chengqi Zhang, ‘An Agent-Based Hybrid Intelligent [68] Fr Sayyed, Rv Argiddi, and Ss Apte, ‘Generating Recommendations System for Financial Investment Planning’, in PRICAI 2002: Trends in for Stock Market Using Collaborative Filtering’, Ijces.Org, 3(1), 46– Artificial Intelligence SE - 39, eds., Mitsuru Ishizuka and Abdul Sattar, 49, (2013). volume 2417 of Lecture Notes in Computer Science, 355–364, Springer [69] Robert P Schumaker and Hsinchun Chen, ‘Textual Analysis of Stock Berlin Heidelberg, (2002). Market Prediction Using Breaking Financial News: The AZFin Text [91] Xiaoxue Zhao, Weinan Zhang, and Jun Wang, ‘Risk-Hedged Venture System’, ACM Transactions on Information System, 27(2), 12:1—- Capital Investment Recommendation’, in Proceedings of the 9th ACM 12:19, (2009). Conference on Recommender Systems, RecSys ’15, pp. 75–82, New [70] Vivek Sehgal and Charles Song, ‘SOPS: Stock prediction using web York, NY, USA, (2015). ACM. 10