In this work we demonstrate the usefulness of the application of Recommender Systems in the financial domain. Specifically we investigate a dataset, made available by a major European bank, containing the purchases of a large set of investment assets by 200k investors. We also present some preliminary results of the application of network analysis via statistical validation to identify clusters of investment assets.
Banking and Financial Services, being them provided by incumbent Banks or by FinTech companies, are looking seriously at machine learning and information retrieval fields in order to leverage the data at their disposal to provide tailored services and customized experiences to their customers.
One of the fields of computer science which can support this attempt is the one represented by Recommender Systems (RecSys), which has been heavily investigated in the last years by the research community as well as the most promising companies in the ecommerce and entertainment fields.
In this work we show the usefulness of some RecSys algorithms in suggesting investment assets to a large panel of investors. This is done by using a large dataset provided by a major European bank and comparing the performance of three diferent RecSys against two baseline models in the task of suggesting investment assets. The recommender system implementation and analysis have been done on a dataset with financial investment information, made available to us by a European bank during a research collaboration program, which contains 224,885 clients, 1,288,315 transactions and information related to 7 diferent asset types, 23 rating levels, 6 order channels, 12 industrial sectors, 8 maturity buckets, 5 coupon types, 2 product complexity levels.
The records span a period of twelve months and all data entries are properly hashed, anonymized and organized as a table, where each record represents a purchase defined by: execution date, user data (client, branch and account identifiers) and traded item data (type of asset, transaction currency, asset country, time to maturity, complexity, industrial sector, industrial group, industrial sub-group, rating, coupon type, trading channel, buy-sell type).
Having no information on the traded volume per transaction nor the client total wealth at the time of each trade, we model the recommendation problem on the basis of the binary information purchased/not purchased item.
Implicit feedback recommender for financial investments To better capture clients’ preferences we compare three diferent RecSys algorithms. All of them required to test diferent combinations of features at our disposal in order to define user and item entities. After some trials and analysis, we defined the user as a combination of client ID and bank branch, and the item as a combination of asset type, country, time to maturity, coupon type, industrial sector and rating.The results for other aggregations are qualitatively analogous.
The first algorithm we tested is the Bayesian Personalized
Ranking algorithm [
The RecSys algorithms mentioned above are evaluated through
various tests against two benchmark algorithms based on most
popular items by number of users (POP.u) or by number of transactions
(POP.trans):
(1) Average Accuracy of the user preference structure (see [
Other metrics (e.g. novelty and coverage) have been calculated but are left out for lack of space. Diferent train/test sampling methodologies were used: (1) leave-one-out: removing randomly from train one purchased asset for each user (who has at least 5 purchases); (2) leave-last-out: removing from train the last (in time) asset purchased by each user; (3) 20% level sampling: removing randomly from train 20% of interactions. Due to limited amount of space, we here report the results for the leave-last-out case only.
Given that a good RecSys should give suggestions relevant and specific to the user and expand user’s taste into neighboring areas, we run the above tests after removing n = {0, 20, 50} most popular items from the test set. In this way if a RecSys performs well with 0 popular items removed and poorly with 50, it is reasonable to deduce that maybe it is just good in suggesting popular items but not items related to the specific interests of the user.
Table 1 displays the results of our study for the leave-last-out train/test sampling case.It shows that all the RecSys we propose perform extremely well on the dataset at our disposal, in terms of both average accuracy and ranking structure (expected percentile ranking - Rank - and AUC). BPR-MF is the best performer when no popular items is excluded from the test set and its advantage doesn’t reduce when we increase the number of popular purchased item removed from the test set. ALS performs similarly well, while Asset Embedding performs better than POPs when the number of popular items excluded from the test set is at least 50, but it never beats BPR-MF and ALS.
Toward a network-based RecSys for banking and
ifnancial services
Besides the training of the recommender system shown above and
the detailed test previously mentioned, we performed an analysis
of the dataset seen as it were a bipartite network users → items.
We implemented a statistical validation procedure [
the number of purchased objects in the diferent communities.
As an example of the possible network analysis, Figure 1 shows the statistically filtered network derived by applying the validation algorithm to the bipartite network with the same specification of users and items as in Table 1, for 1% confidence threshold. There are 4 big connected communities, 2 smaller ones (but still connected) and 6 small isolated communities. Color of nodes (i.e. of assets) refers to diferent value of sector attribute. As an example, the light-blue sector, Governmental assets, results to be statistically over-expressed in the rightmost community, and under-expressed in the leftmost and in the bottom one. This evidence indicates that statistically filtered investors’ decisions could be used to cluster assets: a promising starting point to build a statistically guided algorithm for recommendations. This is part of a work in progress for future publication.
The authors would like to thank MPS Bank for supporting the collaboration, Francesco Mainieri (MPS) for essential support in extracting the data and Franco Maria Nardini (CNR, Pisa) for useful comments. FL and DR acknowledge support by the European Community’s H2020 Program under the scheme INFRAIA-1- 2014-2015: Research Infrastructures, Grant Agreement No. 654024 SoBigData: Social Mining & Big Data Ecosystem.