Financial Asset Recommendation

Overview

The asset recommender leverages a hybrid recommendation pipeline that integrates:

• Collaborative Filtering (CF): Uses customers' past buy transactions.
• Content-Based Filtering (CB): Uses asset features and profitability data.
• Demographic Based Scoring: Incorporates customer risk profiles and demographics.

A Streamlit frontend is provided to allow user interaction and parameter tuning.

---

Dataset Source

This system is built upon the FAR-Trans dataset, a comprehensive financial asset recommendation dataset, provided by a European financial institution.

Citation:

Sanz-Cruzado, J., Droukas, N., & McCreadie, R. (2024).
FAR-Trans: An Investment Dataset for Financial Asset Recommendation.
IJCAI-2024 Workshop on Recommender Systems in Finance (Fin-RecSys), Jeju, South Korea.
arXiv:2407.08692

License: CC-BY 4.0
Link: https://creativecommons.org/licenses/by/4.0/

The dataset includes:

• Customer demographics and investment profiles
• Detailed financial product metadata
• Historical transaction logs
• Time-series pricing and profitability data
• MiFID-aligned structure for risk profiling

---

Data Sources

1. Customer Information

   • File: customer_information.csv
   • Details: Contains customer identifiers, type, risk level, investment capacity, and timestamps.

2. Asset Information

   • File: asset_information.csv
   • Details: Contains ISIN, asset name, asset categories/subcategories, market identifier, sector, industry, and update timestamps.

3. Transactions

   • File: transactions.csv
   • Details: Contains customer transactions (Buy/Sell) with monetary values, units, channels, and market information.
   • Note: Preprocessed to use only "Buy" transactions as positive interaction signals.

4. Limit Prices

   • File: limit_prices.csv
   • Details: Contains profitability data (ROI), first/last dates, and extreme values for every asset.

---

System Components

1. Data Loading & Preprocessing

• CSV Loading:
  Load all dataset files assuming CSV formatting and UTF-8 encoding.

• Preprocessing Transactions:

  • Filter to include only "Buy" transactions.
  • Sort by timestamp for proper train-test splitting.
  • Build a customer × asset rating matrix using transaction counts.

• Train-Test Split:

  • Use leave-one-out split for evaluation.
  • For each user, hold out their last transaction as test data.

---

2. Recommendation Pipeline

A. Collaborative Filtering

• Matrix Factorization:
  • Use Truncated SVD with 5 components.
  • Compute latent factors for users and assets.
  • Output: Predicted ratings for customer-asset pairs.

B. Content-Based Filtering

• Asset Profile Building:

  • One-hot encode categorical features (category, subcategory, sector, industry, market).
  • Include profitability as a numerical feature.
  • Handle missing values with appropriate defaults.
  • Output: Feature matrix for all assets.

• User Profile & Scoring:

  • Build user profile as mean of their purchased assets' features.
  • Compute cosine similarity between user profile and all assets.
  • Handle cold-start with neutral scores.
  • Output: Content-based similarity scores.

C. Demographic Based Scoring

• Risk Profile Matching:
  • Map customer risk levels and investment capacity to numeric scores.
  • Compute weighted similarity between user demographics and asset categories.
  • Output: Demographic scores for assets.

D. Hybrid Scoring

• Component Weights:

  • Allow dynamic adjustment of weights for each component.
  • Weights can be set independently (no sum constraint).
  • Default weights: CF (0.4), CB (0.3), Demographic (0.3).

• Score Combination:

  • Normalize each component's scores to [0,1] range.
  • Apply weighted combination.
  • Output: Final composite scores.

E. Recommendation Generation

• Filtering and Ranking:
  • Remove previously purchased assets.
  • Rank remaining assets by composite score.
  • Select Top-N recommendations.

---

3. Evaluation & Frontend

A. Evaluation Metrics

• RMSE:

  • Compute on held-out test transactions.
  • Handle edge cases and insufficient data.

• Precision@N & Recall@N:

  • Evaluate recommendation quality at specified N.
  • Robust handling of edge cases and errors.

B. Streamlit Frontend

• User Interface:

  • Customer selection dropdown.
  • Component weight sliders.
  • Top-N parameter setting.
  • Evaluation metrics toggle.

• Risk Assessment:

  • Interactive questionnaire for risk profiling.
  • Questions on risk appetite, investment expectations.
  • Automatic profile updates.

• Recommendation Display:

  • Detailed asset information.
  • Formatted scores and metrics.
  • Profitability and price information.

---

Data Flow Diagram

graph LR
    A[Customer Information]
    B[Asset Information]
    C[Transactions]
    D[Limit Prices]
    
    A -->|Preprocessing| F(Customer Profile)
    B -->|Feature Encoding| G(Asset Features)
    C -->|Filter & Aggregate| H(Rating Matrix)
    D -->|Merge with B| G
    
    H -->|SVD| J(CF Scores)
    G -->|Cosine Similarity| K(CB Scores)
    F -->|Risk Matching| L(Demographic Scores)
    
    J --> M[Score Normalization]
    K --> M
    L --> M
    
    M -->|Weighted Combination| N(Final Scores)
    N -->|Rank & Filter| O(Top-N Recommendations)
    
    O --> P[Streamlit UI]
    P -->|Questionnaire| Q[Risk Assessment]
    Q -->|Update| F

Loading