LLM-Based Portfolio Recommender

• LLM-based personalized portfolio recommender is an integrated framework that fuses semantic feature encoding, graph neural networks, and reinforcement learning to optimize asset allocation.
• The approach utilizes memory augmentation and multi-modal data fusion to enhance context awareness, risk profiling, and interpretability in recommendation systems.
• Empirical results show improved risk-adjusted returns and interpretability, outperforming traditional models in metrics like Sharpe ratio and NDCG.

A LLM-based personalized portfolio recommender denotes an integrated recommender framework that leverages the deep semantic reasoning abilities and adaptive representational power of neural LLMs—often in conjunction with graph neural networks, reinforcement learning, or memory architectures—to optimize asset allocation and item recommendation at the individual investor level. Recent advances have enabled direct conditioning of multi-modal and conversational signals on underlying risk preferences and market dynamics, offering accuracy and interpretability advantages over traditional collaborative filtering and conventional optimization methods (Zhao et al., 6 Jun 2025, Li et al., 15 Dec 2025, Zhu et al., 2024, Chen, 3 May 2025, Ebrat et al., 2 Aug 2025).

1. Core Architectural Principles

LLM-based portfolio recommenders are founded on several architectural paradigms:

• Semantic Feature Encoding: Portfolio items and investor states are embedded via pre-trained LLMs (e.g., BERT, GPT-2/4, FinBERT), which are fine-tuned or prompt-tuned on financial corpora, user reviews, and textual market signals (Zhao et al., 6 Jun 2025, Li et al., 15 Dec 2025, Zhu et al., 2024).
• Heterogeneous Graph Construction: User-instrument interaction networks are formalized as multi-type graphs, with nodes representing users, assets, and (optionally) social/trust entities, and edges capturing interactions (e.g., holdings), co-holdings, and social links (Zhao et al., 6 Jun 2025).
• Adaptive Personalization Streams: Several frameworks maintain independent LoRA modules (low-rank adaptation layers) per user, gated via meta-learned user embeddings, enabling lifelong personalization even at scale (Zhu et al., 2024).
• Memory Augmentation: External memory banks encode user history events as structured, retrievable records; the LLM dynamically retrieves the most relevant historical allocations, supporting efficient context injection for recommendation (Chen, 3 May 2025).

2. Information Fusion and Message Passing

The fusion of multi-modal semantic information and relational graph signals is realized through graph neural networks (GNNs), attention mechanisms, and parallel optimization streams:

• Joint LLM-GNN Embeddings: Text-based features $h_i^{(0)} = \mathrm{LLM}(t_i)$ are fused with graph-structured signals via relational GNN message passing, often leveraging graph-attention coefficients $a_{v,u}^{(r)}$ specific to neighbor relations (Zhao et al., 6 Jun 2025).

$$h_v^{(l+1)} = \sigma \left( \sum_{r \in R} \sum_{u \in \mathcal N_v^{(r)}} \frac{1}{c_{v,r}} a_{v,u}^{(r)} (W_r^{(l)} h_u^{(l)} + b_r^{(l)}) \right)$$

• Parallel Optimization Streams: Many models employ pseudo-label branches (learning interpretable risk/sector labels from embeddings) and late-fusion mechanisms (learned combinations of text-only and graph-only representations) (Zhao et al., 6 Jun 2025).
• Meta-LoRA Personalization: User-specific LoRA modules adapt LLM weights per individual, with gating vectors produced by a CRM (ID-based recommendation module) magnifying small finetune sets to full-data knowledge (Zhu et al., 2024).

3. Personalization via Risk Preference Modeling and RL

Robust personalization hinges on direct estimation of investor risk profiles and their incorporation into allocation policy optimization:

• Risk Profiling: LLM hidden states from user dialogue $h_t$ are projected to bounded risk vectors $r \in [0,1]^d$; a scalar CRRA (constant relative risk aversion) parameter $\gamma_i$ is extracted and informs both utility modeling and the RL reward signal (Li et al., 15 Dec 2025).

$$U_i(x) = \frac{x^{1-\gamma_i} - 1}{1-\gamma_i}, \quad \gamma_i > 0$$

$r = \sigma(W_r h_t + b_r)$

• Policy Optimization via RL: Personalized portfolio allocation is framed as an MDP with state comprising market features, the LLM-derived risk vector, and portfolio weights. RL agents (e.g., PPO) optimize allocations, trading off return, risk penalty, and alignment to inferred investor preferences (Li et al., 15 Dec 2025).

$$r_t = w_t^\top R_{t+1} - \lambda_i \mathrm{Var}(w_t^\top R_{t+1}) + \eta \mathrm{sim}(r, w_t)$$

• Conversational Feedback Loop: The LLM agent both processes user inputs and generates explanatory outputs, enabling iterative update of risk preferences and allocation policy (Li et al., 15 Dec 2025).

4. Memory, Retrieval, and Context Integration

The use of external, trainable memory stores is a defining feature in several LLM-based personalization frameworks:

• Dynamic Memory Profile: User histories are recorded as sets of memory vectors $m_i$ encoding allocations, realized returns, volatility, risk level, and market features via MLP encoders (Chen, 3 May 2025).
• Similarity-Based Retrieval: For each new recommendation request, top-$k$ relevant memory entries are extracted by cosine or risk-weighted similarity, enhancing context relevancy and reducing prompt length (Chen, 3 May 2025).
• Prompt Construction: Retrieved memory is formatted as concise, interpretable list objects and injected into the LLM prompt alongside risk constraints, horizon, and diversification objectives (Chen, 3 May 2025).

5. Training Protocols, Loss Functions, and Hyperparameter Choices

Training LLM-based portfolio recommenders involves multi-stage optimization and hybrid objectives:

• Joint Loss Formulation: Loss functions typically combine supervised cross-entropy or BPR ranking loss with auxiliary terms (pseudo-label BCE, distillation KL, risk regularization), plus $\ell_2$ norm regularization for stability (Zhao et al., 6 Jun 2025, Zhu et al., 2024, Ebrat et al., 2 Aug 2025).

$$\mathcal{L} = \alpha_{\text{sup}} \mathcal{L}_{\text{sup}}(\hat{y}, y) + \alpha_{\text{pseudo}} \sum_{i} \mathrm{BCE}(D_i, y_i) + \lambda \| \Theta \|_2^2$$

• RL Losses in PPO (Li et al., 15 Dec 2025):

$$L^{\text{PPO}}(\theta) = \mathbb E_t \left[ \min\left( r_t(\theta) \widehat{A}_t, \operatorname{clip}(r_t(\theta), 1-\epsilon, 1+\epsilon) \widehat{A}_t \right) \right]$$

$$L(\theta) = -L^{\text{PPO}}(\theta) + c_1 \mathbb E_t \left[ (V_\theta(s_t) - V_t^{\text{target}})^2 \right] - c_2 \mathbb E_t \left[ \pi_\theta(\cdot | s_t) \ln \pi_\theta(\cdot | s_t) \right]$$

• Hyperparameters: Typical configurations include LLM/GNN embedding dimensions (128–384), GAT heads (4), dropout rates (0.2), AdamW learning rates ($1e^{-3}$), LoRA ranks (8–32), retrieval windows (short/long histories), and batch sizes (64–1024) (Zhao et al., 6 Jun 2025, Zhu et al., 2024, Li et al., 15 Dec 2025, Ebrat et al., 2 Aug 2025).

6. Empirical Performance, Interpretability, and Comparative Metrics

Empirical benchmarks demonstrate superior recommendation quality, return, risk-adjusted performance, and interpretability across multiple datasets and baselines:

• Portfolio Metrics: Top-K hit rate, cumulative/average daily return, Sharpe ratio, diversity (sector entropy), calibration, NDCG@10, MRR, and risk-regularizer error are all employed (Zhao et al., 6 Jun 2025, Li et al., 15 Dec 2025, Zhu et al., 2024, Ebrat et al., 2 Aug 2025).
• Results Table—Portfolio Recommendation (Li et al., 15 Dec 2025):

Model	AR (%)	SR	MDD (%)	IR	CR	UAS	CSS
MVO	8.42	0.94	22.6	0.47	0.38	0.52	0.60
DRL-PPO	11.87	1.21	18.3	0.64	0.52	0.66	0.71
BERT-FA	10.54	1.12	19.7	0.59	0.46	0.74	0.82
L-PPR	14.63	1.45	15.1	0.78	0.63	0.89	0.93

All metrics for L-PPR (LLM-based recommender) improve over baselines at $p < 0.01$.

• Ablations and Insights: Removing graph or text streams, pseudo-label losses, or LLM initialization components notably degrades performance—up to 5% NDCG loss in cold-start scenarios (Zhao et al., 6 Jun 2025, Ebrat et al., 2 Aug 2025).
• Interpretability: Attention weights and pseudo-label branches elucidate drivers of personalization (e.g., risk, sector preference), while natural-language explanations are generated by the LLM for transparency (Zhao et al., 6 Jun 2025, Li et al., 15 Dec 2025, Ebrat et al., 2 Aug 2025).

7. Limitations, Open Problems, and Future Directions

LLM-based personalized portfolio recommenders exhibit several open technical and practical issues:

• Market Realism: Simulated market environments may omit real transaction costs, slippage, or regime shifts, limiting live applicability (Li et al., 15 Dec 2025).
• User Data Heterogeneity: Synthetic user dialogues and constrained history profiles may underrepresent population variability (Li et al., 15 Dec 2025).
• LLM Biases: Prompt engineering and domain drift in LLM risk inference remain unsolved; interpretability may be limited by opaque neural outputs (Li et al., 15 Dec 2025, Zhu et al., 2024).
• Scalability: Training on full user histories or across extremely large financial datasets challenges both efficiency and accuracy, motivating hybrid retrieval and memory-based designs (Zhu et al., 2024, Chen, 3 May 2025).
• Research Directions: Future work includes integrating real-time news, live-trading execution data, continual learning for model drift robustness, multi-agent RL for equilibrium analysis, and improving risk-awareness in semantic profiling (Li et al., 15 Dec 2025, Zhao et al., 6 Jun 2025).

A plausible implication is that continued fusion of LLMs with graph structures, memory modules, and RL policies—augmented with strong risk modeling and retrieval strategies—will be central to the next generation of adaptive, scalable, and interpretable portfolio recommendation platforms.