Optimal Client Recommendation for Market Makers in Illiquid Financial Products

Dieter Hendricks, Stephen J. Roberts

TL;DR

Market makers in illiquid financial markets often struggle to offload unwanted inventory. This paper applies Natural Language Processing (NLP) techniques—specifically Latent Dirichlet Allocation (LDA)—to transaction data, treating financial products as "documents" and clients as "words" to generate high-quality, prioritized recommendation lists for sales traders.

What Problem Does This Paper Address?

In financial markets, "market makers" (like large banks) provide liquidity by acting as the counterparty to a trade when a natural buyer or seller isn't immediately available. This often leaves the market maker holding a "proprietary position"—essentially an unwanted inventory of a specific asset, such as a corporate bond.

For illiquid products, the market maker cannot simply wait for a buyer to show up. Instead, sales traders must proactively "target" clients who might be interested. Currently, this relies heavily on a trader's intuition and personal relationships. As the number of clients and products grows, this manual process becomes an "arduous task." The authors seek to automate this by inferring client profiles and providing a curated, ranked list of recommendations to expedite the liquidation of these positions.

Methods

The core innovation is the use of a topic-modeling analogy to frame the recommendation problem:

  • Word $\rightarrow$ Client: The basic unit of data.
  • Document $\rightarrow$ Product (Bond): A collection of clients who have expressed interest in that bond.
  • Topic $\rightarrow$ Client Group: A latent (hidden) group of clients who share similar investment interests.
  • Corpus $\rightarrow$ Market: The set of all products the market maker trades.

The authors primarily utilize Latent Dirichlet Allocation (LDA), a three-level hierarchical Bayesian model. By observing which clients trade which bonds, the model learns the "topic mixture" (what kind of interest groups a bond appeals to) and "word mixture" (which clients belong to those interest groups).

Key technical details include: * TF-IDF Preprocessing: To handle the high sparsity of financial data (where a few clients dominate trading), the authors use Term Frequency-Inverse Document Frequency to down-weight ubiquitous clients and highlight meaningful, specific interests. * Variational Bayes: They use the scikit-learn implementation of online variational Bayes to make the posterior inference tractable. * Baselines: They compare LDA against Empirical Term-Frequency (ETF) (a simple historical count) and Non-negative Matrix Factorization (NMF).

Datasets

The study uses a proprietary dataset provided by BNP Paribas (BNPP):

  • Scope: Daily recorded transactions from January 5, 2015, to February 10, 2017.
  • Features: Anonymized Client and Product IDs, trade direction (Buy/Sell), trade status (e.g., "Done," "Passed," "IOI"), and bond metadata (currency, seniority, maturity).
  • Processing: The authors collapsed various trade statuses into a single "interest" indicator to reduce data sparsity. They focused specifically on "active" bonds (those issued before the training period and maturing after the test date).

Key Insights and Ideas

  • The "Bag-of-Clients" Approach: By treating a bond as a collection of interested clients, the model can find "latent" relationships that aren't obvious. For example, it can identify that a group of clients consistently trades high-yield energy bonds, even if they haven't traded a specific new energy bond yet.
  • Redefining "False Positives": In a typical recommender, a "False Positive" is a failure. However, the authors introduce Relevant False Positives (RFP). This occurs when a recommended client doesn't trade the suggested bond but does trade a different bond on that same day.
  • Sales Lead Generation: For a sales trader, an RFP is still a "win." It identifies a client who is active in the market, providing an opening for a conversation that could lead to a different trade.

Key Results

  • Superiority over Random: All models significantly outperformed random client sampling, proving that historical transaction data contains strong predictive signals.
  • LDA vs. ETF: While the simple ETF model performed surprisingly well for "Top 10" recommendations, LDA offered superior accuracy as the list grew (up to "Top 100"). This suggests that the latent structure captured by LDA is better at refining the long-tail of potential interests.
  • High-Quality Leads: For their best-performing LDA model (using a 500-day rolling window), 79% of the "incorrect" recommendations were actually Relevant False Positives. This means nearly 4 out of 5 "wrong" guesses still pointed the trader toward a client who was actively trading that day.

Limitations and Future Work

  • Temporal Dynamics: The standard LDA model assumes "exchangeability" (order doesn't matter). However, financial interests are highly temporal—a client's interest in a bond changes as it nears maturity or as market conditions shift. The authors suggest moving toward more sophisticated dynamic topic models.
  • Metadata Integration: The current model relies almost entirely on transaction history. Incorporating bond metadata (like coupon rates, duration, or industry sector) directly into the LDA algorithm (e.g., via a Structural Topic Model) could significantly improve recommendations for newly issued bonds with no trade history.
  • Data Imbalance: The "long-tail" of infrequent traders remains a challenge. Future work could explore more advanced data balancing techniques to ensure these clients aren't ignored by the model.

Summary updated

Abstract

The process of liquidity provision in financial markets can result in prolonged exposure to illiquid instruments for market makers. In this case, where a proprietary position is not desired, pro-actively targeting the right client who is likely to be interested can be an effective means to offset this position, rather than relying on commensurate interest arising through natural demand. In this paper, we consider the inference of a client profile for the purpose of corporate bond recommendation, based on typical recorded information available to the market maker. Given a historical record of corporate bond transactions and bond meta-data, we use a topic-modelling analogy to develop a probabilistic technique for compiling a curated list of client recommendations for a particular bond that needs to be traded, ranked by probability of interest. We show that a model based on Latent Dirichlet Allocation offers promising performance to deliver relevant recommendations for sales traders.
BibTeX
@inproceedings{hendricks2017optimal,
  author = {Dieter Hendricks and Stephen J. Roberts},
  title = {{Optimal Client Recommendation for Market Makers in Illiquid Financial Products}},
  booktitle = {ECML/PKDD},
  year = {2017},
  url = {https://openreview.net/forum?id=oq9nMwSGrY},
}
Figure