Mean-Variance Efficient Collaborative Filtering for Stock Recommendation

TL;DR

The paper addresses the challenge of providing stock recommendations that respect individual investor preferences while improving portfolio diversification under price uncertainty. It introduces MVECF, a mean-variance regularized extension of weighted matrix factorization that explicitly optimizes for mean-variance efficiency through a regularization term and a restructuring that enables ALS optimization. Empirical results on CRSP data show MVECF yields substantial improvements in portfolio Sharpe ratio with only modest drops in conventional recommendation metrics, and the approach can be extended to graph-based ranking via MV-efficient sampling. This work advances stock recommender systems by integrating risk-return considerations into CF and demonstrates practical applicability for fintech platforms seeking personalized, risk-aware stock suggestions.

Abstract

The rise of FinTech has transformed financial services online, yet stock recommender systems have received limited attention. Personalized stock recommendations can significantly impact customer engagement and satisfaction within the industry. However, traditional investment recommendations focus on high-return stocks or highly diversified portfolios, often neglecting user preferences. The former would result in unsuccessful investment because accurately predicting stock prices is almost impossible, whereas the latter would not be accepted by investors because many investors, including both individuals and institutional portfolio managers, who typically hold focused portfolios based on their investment strategies and interests. Collaborative filtering (CF) also may not be directly applicable to stock recommendations, because it is inappropriate to just recommend stocks that users like. The key is to optimally blend user's preference with the portfolio theory. However, no existing model considers both aspects. We propose a simple yet effective model, called mean-variance efficient collaborative filtering (MVECF). Our model is designed to improve the Pareto optimality in a trade-off between the risk and return by systemically handling uncertainties in stock prices. Experiments on real-world data show our model can increase the mean-variance efficiency of recommended portfolios while sacrificing just a small amount of recommendation accuracy. Finally, we further show MVECF is easily applicable to the graph-based ranking model.

Figures (6)

• Figure 1: Yearly sub-dataset construction
• Figure 2: Performance comparison between MVECF and baseline models across 17 yearly datasets (2006-2022). MVECF$_{\text{wmf}}$ consistently achieves superior portfolio efficiency with $\Delta\text{SR}$ around 0.1 and nearly 100% SR improvement rate, while maintaining competitive recommendation accuracy (MAP@20 $\approx$ 20%, Recall@20 $\approx$ 80%). Traditional recommenders (WMF, LightGCN, BPR) show minimal portfolio improvements, while MPT$_{\text{topSR}}$ exhibits high variance across years.
• Figure 3: Ex-post performance comparison using 5-year realized returns. MVECF$_{\text{wmf}}$ demonstrates robust out-of-sample performance with positive $\Delta\text{SR}$ and dominating SR improvement rate across most years. In contrast, MPT$_{\text{topSR}}$ and 2Step$_{\text{wmf}}$ show unstable performance with frequent Sharpe ratio deterioration.
• Figure 4: Train and validation losses of $\text{MVECF}_\text{WMF}$.
• Figure 5: $\Delta \mu$ and $\Delta \sigma$ of $\text{MVECF}_\text{wmf}$ with various values of $\gamma$.