Paper: arXiv 2308.01208

Authors: Ashraf Ghiye, Baptiste Barreau, Laurent Carlier, Michalis Vazirgiannis

Abstract

Classical recommender systems often assume that historical data are stationary and fail to account for the dynamic nature of user preferences, limiting their ability to provide reliable recommendations in time-sensitive settings. This assumption is particularly problematic in finance, where financial products exhibit continuous changes in valuations, leading to frequent shifts in client interests. These evolving interests, summarized in the past client-product interactions, see their utility fade over time with a degree that might differ from one client to another. To address this challenge, we propose a time-dependent collaborative filtering algorithm that can adaptively discount distant client-product interactions using personalized decay functions. Our approach is designed to handle the non-stationarity of financial data and produce reliable recommendations by modeling the dynamic collaborative signals between clients and products. We evaluate our method using a proprietary dataset from BNP Paribas and demonstrate significant improvements over state-of-the-art benchmarks from relevant literature. Our findings emphasize the importance of incorporating time explicitly in the model to enhance the accuracy of financial product recommendation.

Complexity vs Empirical Score

  • Math Complexity: 4.5/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper employs moderate mathematical complexity through graph neural network formulations and embedding decomposition but focuses heavily on empirical validation using a proprietary real-world dataset from BNP Paribas with clear benchmark comparisons and ablation studies.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Dynamic Financial Product Recommendations"] --> B{"Data Input:<br>BNP Paribas Proprietary Dataset"}
  B --> C["Preprocessing &<br>Time-Stamping Interactions"]
  C --> D["Adaptive Collaborative Filtering:<br>Personalized Time Decay Functions"]
  D --> E["Model Training:<br>Handling Non-Stationary Data"]
  E --> F["Evaluation:<br>Comparison with State-of-the-Art Benchmarks"]
  F --> G["Outcome:<br>Significant Improvement in Recommendation Accuracy"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a time‑dependent collaborative filtering algorithm with personalized time decay functions (ACF) to adaptively discount past client‑product interactions.

Source sentence
We propose personalized time decay functions to aggregate the past, assuming that the utility of past interactions decreases monotonically with time.

Assumes 2

  • Interest persists for a certain period after interactions and then fades over time. model
    Source sentence
    In this work, we assume that interest persists for a certain period after the interactions have occurred, after which the utility of the interactions progressively fades over time.
  • The utility of past interactions decreases monotonically with time. model
    Source sentence
    We propose personalized time decay functions to aggregate the past, assuming that the utility of past interactions decreases monotonically with time.

Data

  • proprietary database of G10 Bonds RFQs · BNP Paribas other proprietary
    Source sentence
    We conduct an extensive study on a proprietary database of G10 Bonds RFQs to test the effectiveness of different aggregation functions in encoding the dynamic collaborative signals.

Costs

Not applicable The paper reports no trading or execution result.

Finds 2

  • ACF outperforms all baselines, achieving up to 16.86% gain in s‑mAP and 12.65% gain in s‑MRR compared to HCF. performance
    Source sentence
    The best model, ACF-LE, shows substantial improvements in both metrics - with gains of up to 16.86% in s-mAP and 12.65% in s-MRR compared to HCF.
  • Time‑decayed approaches outperform attention mechanisms and recurrent‑based models, despite both being better than a simple average. performance
    Source sentence
    Our experiments show that time-decayed approaches outperform attention mechanisms with time embeddings, although both perform better than a simple average (HCF).

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 7 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 2 papers in the archive (1 by independent authors)

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.