Impact of Expert-Following Strategies in Financial Asset Recommendation

Technical Summary: Expert-Following Strategies in Financial Asset Recommendation

Problem Statement
Financial Asset Recommendation (FAR) faces a fundamental trade-off between two competing objectives: maximizing actual investment returns (Return on Investment, or ROI) and ensuring alignment with investor preferences (measured by normalized Discounted Cumulative Gain, or nDCG). Existing approaches typically optimize for only one of these metrics. Return-based strategies often ignore individual preferences, resulting in low relevance, while preference-based strategies (such as collaborative filtering or popularity models) tend to yield returns comparable to the market average without significant alpha. Previous research [2] has confirmed that no existing method simultaneously outperforms baselines in both ROI and nDCG, creating a need for a framework that can jointly optimize these criteria using behavioral signals.

Methodology
The paper proposes Expert-Following Strategies, a framework designed to bridge the gap between profitability and relevance by leveraging the transaction histories of high-performing investors. The methodology operates as follows:

1. Expert Identification: The system identifies a subset of "expert" investors ($E_X$) from the training period based on their historical ROI. Specifically, the top $X\%$ of investors (where $X \in \{5, 10, 15, 20\}$) are selected. An investor's ROI is calculated as the portfolio return over the training window $[t_0, t]$, considering net share holdings (purchases minus sales).
2. Asset Scoring: Instead of using complex model learning, the framework scores assets based on the collective behavior of these identified experts. For each asset $i$, a recommendation score $S_i$ is computed as the sum of ROI-weighted purchase frequencies:
   $$S_i = \sum_{u \in E_X} \max(\text{ROI}_u(t_0, t), 0) \cdot C_{u,i}$$
   Here, $C_{u,i}$ represents the number of times expert $u$ purchased asset $i$. This formulation ensures that assets purchased frequently by high-performing investors receive higher scores, effectively combining collective preference signals with profitability weighting.
3. Recommendation: The top-$k$ assets (with $k=10$) with the highest scores are recommended to the user.

Experimental Setup
The study utilizes the FAR-Trans dataset, covering transaction histories from January 2018 to November 2022. The evaluation employs a robust multi-instance walk-forward validation framework:

• Validation: 10 sliding-window instances were generated, each with a training period of approximately 14 months and an evaluation period of 10 months.
• Metrics: Performance was evaluated using ROI@10 (mean portfolio return), SR@10 (annualized Sharpe Ratio), and nDCG@10 (relevance of purchased items within the holding period).
• Baselines: The proposed method was compared against return-based strategies (Sharpe Ratio Optimization, Random Forest), preference-based strategies (Popularity), and a Market Average baseline.

Key Results
The experiments demonstrate that Expert-Following Strategies successfully resolve the ROI–nDCG trade-off:

• Simultaneous Improvement: Unlike baselines that excel in one metric at the expense of the other (e.g., Random Forest achieved high ROI but near-zero nDCG; Popularity achieved high nDCG but no significant ROI gain), the Expert-Following variants achieved statistically significant improvements over the Market Average baseline in both ROI and nDCG simultaneously across all four expert thresholds.
• Top Performance: The top-5% expert variant achieved the highest ROI (0.246) while maintaining a significant nDCG (0.176), occupying the high-ROI/high-nDCG region in the performance landscape.
• Statistical Significance: The improvements were validated using paired t-tests and Wilcoxon signed-rank tests with Bonferroni correction ($\alpha' \approx 0.0071$).

Significance and Claims
The paper claims that Expert-Following Strategies offer a simple yet effective solution to the dual-objective optimization problem in financial recommendation. The core significance lies in the finding that high-performing investors inherently exhibit transaction patterns that reflect both profit-maximizing strategies and realistic purchasing behaviors that align with general investor preferences. By identifying and following these experts, the framework achieves a robust balance without requiring complex model training. The authors note that while the approach assumes zero transaction costs (which may overestimate ROI for high-turnover strategies), it provides a strong empirical foundation for future work involving realistic fees and graph-based models.