Recommender Systems for Banking and Financial Services Andrea Gigli Fabrizio Lillo Daniele Regoli MPS Capital Services Università di Bologna Scuola Normale Superiore Viale Mazzini, 23 Viale Quirico Filopanti 5 Piazza dei Cavalieri 7 Siena, Italy 53100 Bologna, Italy 40126 Pisa, Italy 56126 andrea.gigli@mpscs.it fabrizio.lillo@unibo.it daniele.regoli@sns.it ABSTRACT data (client, branch and account identifiers) and traded item data In this work we demonstrate the usefulness of the application of (type of asset, transaction currency, asset country, time to maturity, Recommender Systems in the financial domain. Specifically we complexity, industrial sector, industrial group, industrial sub-group, investigate a dataset, made available by a major European bank, rating, coupon type, trading channel, buy-sell type). containing the purchases of a large set of investment assets by Having no information on the traded volume per transaction 200k investors. We also present some preliminary results of the nor the client total wealth at the time of each trade, we model the application of network analysis via statistical validation to identify recommendation problem on the basis of the binary information clusters of investment assets. purchased/not purchased item. KEYWORDS Implicit feedback recommender for financial Finance, Collaborative Filtering, Networks, Statistical Validation investments ACM Reference format: To better capture clients’ preferences we compare three different Andrea Gigli, Fabrizio Lillo, and Daniele Regoli. 2017. Recommender Systems RecSys algorithms. All of them required to test different combina- for Banking and Financial Services. In Proceedings of RecSys 2017 Posters, tions of features at our disposal in order to define user and item Como, Italy. Copyrights held by the authors. , August 27-31, 2 pages. entities. After some trials and analysis, we defined the user as a com- bination of client ID and bank branch, and the item as a combination Introduction of asset type, country, time to maturity, coupon type, industrial sec- tor and rating.The results for other aggregations are qualitatively Banking and Financial Services, being them provided by incumbent analogous. Banks or by FinTech companies, are looking seriously at machine The first algorithm we tested is the Bayesian Personalized Rank- learning and information retrieval fields in order to leverage the ing algorithm [3] where we use a matrix factorization method data at their disposal to provide tailored services and customized maximizing the posterior probability of user preference structure, experiences to their customers. and tune model’s parameters via 5-fold cross-validation. The sec- One of the fields of computer science which can support this ond one is the Alternating Least Squares algorithm [1] using 30 attempt is the one represented by Recommender Systems (RecSys), latent factors and a regularization factor equal to 0.01. The third which has been heavily investigated in the last years by the research one is an adaptation of the Word2Vec algorithm [2] that we call community as well as the most promising companies in the e- Asset-Embedding in the following. In this case we treated the clients’ commerce and entertainment fields. portfolios as they were documents, each asset as a word, and vector- In this work we show the usefulness of some RecSys algorithms represented each asset by the portfolio it belongs to via continuous in suggesting investment assets to a large panel of investors. This bag of words in a 300 dimension space. is done by using a large dataset provided by a major European bank The RecSys algorithms mentioned above are evaluated through and comparing the performance of three different RecSys against various tests against two benchmark algorithms based on most pop- two baseline models in the task of suggesting investment assets. ular items by number of users (POP.u) or by number of transactions (POP.trans): The Dataset (1) Average Accuracy of the user preference structure (see [3]); The recommender system implementation and analysis have been (2) Expected percentile ranking, as defined in [1] (the lower, the done on a dataset with financial investment information, made better); available to us by a European bank during a research collaboration (3) the Area Under the ROC curve. program, which contains 224,885 clients, 1,288,315 transactions and information related to 7 different asset types, 23 rating levels, 6 Other metrics (e.g. novelty and coverage) have been calculated but order channels, 12 industrial sectors, 8 maturity buckets, 5 coupon are left out for lack of space. Different train/test sampling method- types, 2 product complexity levels. ologies were used: The records span a period of twelve months and all data entries (1) leave-one-out: removing randomly from train one purchased are properly hashed, anonymized and organized as a table, where asset for each user (who has at least 5 purchases); each record represents a purchase defined by: execution date, user (2) leave-last-out: removing from train the last (in time) asset purchased by each user; RecSys 2017 Poster Proceedings, August 27-31, Como, Italy. Copyrights held by the authors. (3) 20% level sampling: removing randomly from train 20% of interactions. RecSys 2017 Poster Proceedings, August 27-31, Como, Italy. Copyrights held by the authors. Andrea Gigli, Fabrizio Lillo, and Daniele Regoli Table 1: Evaluation metrics for leave-last-out test method- ology, with variable number of most purchased items ex- cluded from test set. most purchased recommender Average Rank AUC excluded items system Accuracy BPR-MF 0.961 3.878 0.970 Asset Embedding 0.951 4.975 0.950 0 ALS 0.954 4.590 0.954 POP.u 0.949 5.119 0.958 POP.trans 0.950 5.044 0.958 BPR-MF 0.941 5.919 0.951 Asset Embedding 0.906 9.389 0.903 20 ALS 0.909 9.080 0.906 POP.u 0.916 8.404 0.926 POP.trans 0.917 8.286 0.927 BPR-MF 0.885 11.537 0.914 Asset Embedding 0.859 14.105 0.874 50 ALS 0.917 8.259 0.913 Figure 1: Communities (pink regions) of assets detected POP.u 0.825 17.472 0.819 on the statistically filtered asset graph projection. Color de- POP.trans 0.820 18.032 0.817 notes sector attribute. Due to limited amount of space, we here report the results for the the number of purchased objects in the different communi- leave-last-out case only. ties. Given that a good RecSys should give suggestions relevant and As an example of the possible network analysis, Figure 1 shows specific to the user and expand user’s taste into neighboring areas, the statistically filtered network derived by applying the validation we run the above tests after removing n = {0, 20, 50} most popular algorithm to the bipartite network with the same specification of items from the test set. In this way if a RecSys performs well with users and items as in Table 1, for 1% confidence threshold. There are 0 popular items removed and poorly with 50, it is reasonable to 4 big connected communities, 2 smaller ones (but still connected) deduce that maybe it is just good in suggesting popular items but and 6 small isolated communities. Color of nodes (i.e. of assets) not items related to the specific interests of the user. refers to different value of sector attribute. As an example, the Table 1 displays the results of our study for the leave-last-out light-blue sector, Governmental assets, results to be statistically train/test sampling case.It shows that all the RecSys we propose over-expressed in the rightmost community, and under-expressed perform extremely well on the dataset at our disposal, in terms of in the leftmost and in the bottom one. This evidence indicates that both average accuracy and ranking structure (expected percentile statistically filtered investors’ decisions could be used to cluster ranking - Rank - and AUC). BPR-MF is the best performer when assets: a promising starting point to build a statistically guided no popular items is excluded from the test set and its advantage algorithm for recommendations. This is part of a work in progress doesn’t reduce when we increase the number of popular purchased for future publication. item removed from the test set. ALS performs similarly well, while Asset Embedding performs better than POPs when the number of ACKNOWLEDGMENTS popular items excluded from the test set is at least 50, but it never beats BPR-MF and ALS. The authors would like to thank MPS Bank for supporting the collaboration, Francesco Mainieri (MPS) for essential support in Toward a network-based RecSys for banking and extracting the data and Franco Maria Nardini (CNR, Pisa) for useful comments. FL and DR acknowledge support by the European Com- financial services munity’s H2020 Program under the scheme INFRAIA-1- 2014-2015: Besides the training of the recommender system shown above and Research Infrastructures, Grant Agreement No. 654024 SoBigData: the detailed test previously mentioned, we performed an analysis Social Mining & Big Data Ecosystem. of the dataset seen as it were a bipartite network users → items. We implemented a statistical validation procedure [4] to get a sta- REFERENCES tistically significant projection on the item module of the network. [1] Yifan Hu, Yehuda Koren, and Chris Volinsky. 2008. Collaborative filtering for im- We used this statistically filtered network: plicit feedback datasets. In Data Mining, 2008. ICDM’08. Eighth IEEE International Conference on. Ieee, 263–272. • to identify items’ communities: each community represents [2] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient the set of items that are purchased together by users in a Estimation of Word Representations in Vector Space. CoRR abs/1301.3781 (2013). http://arxiv.org/abs/1301.3781 statistically over-expressed way with respect to a random [3] Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme. rewiring of the bipartite network keeping fixed the assets’ 2009. BPR: Bayesian personalized ranking from implicit feedback. In Proceedings degrees; of the twenty-fifth conference on uncertainty in artificial intelligence. AUAI Press, 452–461. • to identify the features that are statistically over-expressed [4] Michele Tumminello, Salvatore Micciche, Fabrizio Lillo, Jyrki Piilo, and Rosario N (or under-expressed) inside communities; Mantegna. 2011. Statistically validated networks in bipartite complex systems. PloS one 6, 3 (2011), e17994. • to compute, for each user, a raking of communities of items based on the p-values of the hyper-geometric distribution of