Regime-Aware Portfolio Management via Retrieval-Augmented LLM-Guided Expert Switching

Ahmad Asadi; Reza Safabakhsh

cs.AIAug 28, 2026

Regime-Aware Portfolio Management via Retrieval-Augmented LLM-Guided Expert Switching

Authors: Ahmad Asadi, Reza Safabakhsh

Organizations: Deep Learning Lab, Computer Engineering Department, Amirkabur University of Technology, Hafez Avenue, Tehran, Iran

Abstract

Financial markets are inherently non-stationary, making the effectiveness of individual portfolio-management strategies highly dependent on changing market conditions. This work proposes a retrieval-augmented expert-switching framework that dynamically selects portfolio management experts based on their historical performance under similar market situations. A dual-stream variational autoencoder represents asset-level and market-wide information, while a retrieval-based knowledge base stores historical situations and expert performance. During inference, an instruction-tuned LLM reasons over the retrieved evidence to identify the most appropriate expert rather than directly generating portfolio actions. We further establish a monotonicity property showing that adding a locally superior expert cannot degrade the switching mechanism's performance. Experiments across cryptocurrency, stock, and foreign-exchange markets show that the proposed selector achieves the highest cumulative return and Sharpe ratio among the evaluated selection strategies in all three markets. In the stock market, for example, cumulative return increases from 26% for the best fixed expert to 34%, while the Sharpe ratio improves from 0.74 to 0.96. Ablation results confirm the importance of both retrieval and LLM reasoning, while experiments with different expert-pool sizes demonstrate the value of complementary expertise. Overall, the findings support retrieval-grounded expert switching as an effective approach to adaptive portfolio management in non-stationary financial environments.

Figures & tables

Figure 1: Overall architecture of the proposed retrieval-augmented adaptive portfolio management framework. The lower pipeline constructs the historical index by encoding market states and evaluating expert performance. The upper pipeline retrieves comparable historical states, combines their performance records with contextual information, and selects the expert used for portfolio allocation.

Condition	Mathematical requirement	Operational constraint
Axiom 1	$\sup_{s}\|\rho_{i}-U_{i}\|\leq\epsilon$	Restrict quantitative assessments to the empirical range of retrieved return and drawdown observations. Missing regimes are handled using historical baseline statistics.
Axiom 2	$\tau\geq 2\epsilon$	Require a candidate expert to exceed the incumbent by a margin larger than the estimated uncertainty before recommending a switch.
Axiom 3	$\delta>\tau+2\epsilon$	Evaluate specialist performance in extreme conditions.

Table 1: Mapping of the switching assumptions to LLM evaluation constraints.

Market	Symbol	Range	Samples
Crypto	VET_USD	2018-08-03-2026-05-31	2859
Crypto	BCH_USD	2017-11-09-2026-05-31	3126
Crypto	ETC_USD	2017-11-09-2026-05-31	3126
Crypto	UNI_USD	2019-10-21-2025-04-17	2006
Crypto	SHIB_USD	2020-08-01-2026-05-31	2089
Crypto	LTC_USD	2014-09-17-2026-05-31	4275

Table 2: Crypto market data summary.

Market	Symbol	Range	Samples
Stock	TMO	2000-01-03-2026-05-29	6641
Stock	MCD	2000-01-03-2026-05-29	6641
Stock	NVDA	2000-01-03-2026-05-29	6641
Stock	DIS	2000-01-03-2026-05-29	6641
Stock	TSLA	2010-06-29-2026-05-29	4004
Stock	CVX	2000-01-03-2026-05-29	6641

Table 3: Stock market data summary.

Market	Symbol	Range	Samples
Forex	USDCAD	2003-09-17-2026-05-29	5904
Forex	EURAUD	2003-12-01-2026-05-29	5853
Forex	EURCHF	2003-01-23-2026-05-29	6059
Forex	GBPCHF	2003-10-01-2026-05-29	5894
Forex	CHFJPY	2003-12-01-2026-05-29	5854
Forex	NZDCHF	2003-12-01-2026-05-29	5825

Table 4: Foreign-exchange market data summary.

Panel of Figure 2 — Figure 2: Daily OHLC price histories for the 30 cryptocurrency symbols used in the experiments.

Panel of Figure 3 — Figure 3: Daily OHLC price histories for the 30 stock symbols used in the experiments.

Panel of Figure 4 — Figure 4: Daily OHLC price histories for the 30 foreign-exchange symbols used in the experiments.

Panel of Figure 5 — Figure 5: Market-level cryptocurrency indices.

Panel of Figure 6 — Figure 6: Market-level United States equity reference indices for the stock-market experiments.

Market	Expert	Cum. Return	Sharpe	Sortino	MDD	Weeks Dominant
Crypto	A2C	63.00 %	1.61	2.44	-25 %	19 (34.5%)
Crypto	DDPG	71.26 %	1.74	2.73	-26 %	06 (10.9%)
Crypto	PPO	60.49 %	1.57	2.38	-27 %	03 (05.5%)
Crypto	CrossQ	53.66 %	1.44	2.24	-28 %	09 (16.4%)
Crypto	TQC	76.16 %	1.71	2.63	-29 %	18 (32.7%)
Stock	A2C	32.16 %	0.90	1.35	-23 %	17 (19.3%)

Table 5: Standalone performance of the five expert models in each market.

Panel of Figure 7 — Figure 7: Illustrative expert-selection frequencies for the proposed switcher in the three markets.

Market	Method	Cum. Return	Ann. Return	Ann. Vol.	Sharpe	Sortino	MDD
Crypto	Best Fixed	68 %	62 %	0.32	1.64	2.54	-27 %
Crypto	Recent Gating	69 %	62 %	0.32	1.66	2.57	-27 %
Crypto	Proposed RAG	71 %	64 %	0.31	1.73	2.72	-26 %
Stock	Best Fixed	26 %	14 %	0.21	0.74	1.11	-26 %
Stock	Recent Gating	33 %	17 %	0.20	0.91	1.37	-24 %
Stock	Proposed RAG	34 %	18 %	0.19	0.96	1.45	-26 %

Table 6: Effect of expert-selection strategy in each market.

Market	# Experts	Fixed Sharpe	RAG Sharpe	$\Delta$ Sharpe (%)	Fixed Sortino
Crypto	1	0.81	0.81	0.0	1.11
Crypto	2	0.84	0.85	1.2	1.16
Crypto	3	0.89	0.98	10.1	1.22
Crypto	4	0.96	1.15	19.8	1.34
Stock	1	0.43	0.43	0.0	0.58
Stock	2	0.51	0.52	2.0	0.68

Table 7: Effect of expert-pool size on the proposed switcher. Fixed Sharpe values are consistent with the reported RAG Sharpe values and percentage improvements.

Market	Configuration	Cum. Return	Sharpe	MDD
Crypto	NR (No Retrieval)	+64.8%	1.622	-26.7%
Crypto	R $-$ LLM (No LLM)	+69.1%	1.661	-27.1%
Crypto	Full (R+L+U)	71 %	1.73	-26 %
Stock	NR (No Retrieval)	+33.1%	0.911	-24.4 %
Stock	R $-$ LLM (No LLM)	+25.5%	0.735	-26.5%
Stock	Full (R+L+U)	34 %	0.96	-26 %

Table 8: Component ablation across the three markets.

Model	Success	Failure
Google/Gemma4-E4B	28%	72%
Google/Gemma4-31B-it	96%	4%
AliBaba/Qwen3.6-27B-it	100%	0%

Table 9: Success and failure rates for each indexing LLM.

Market	Inference LLM	Cum. Return	Ann. Return	Ann. Vol.	Sharpe	Sortino	MDD
Crypto	Google/Gemma4-31B-it	68.3 %	61.7 %	0.32	1.65	2.55	-27 %
Crypto	Google/Gemma4-E4B	63.5 %	57.4 %	0.32	1.57	2.42	- 28 %
Crypto	AliBaba/Qwen3.6-27B-it	71 %	64 %	0.31	1.73	2.72	-26 %
Stock	Google/Gemma4-31B-it	25.1 %	13.7 %	0.21	0.73	1.09	-26.5 %
Stock	Google/Gemma4-E4B	24.5 %	13.5 %	0.21	0.71	1.06	-27 %
Stock	AliBaba/Qwen3.6-27B-it	34 %	18 %	0.19	0.96	1.45	-26 %

Table 10: Sensitivity to the LLM used for inference.

Market	Method	Cum. Return	Ann. Return	Ann. Vol.	Sharpe	MDD
Crypto	TAC [ Asadi and Safabakhsh, 2025 ]	24 %	21.6 %	0.94	1.29	-31 %
Crypto	AlphaMixRL [ Sun, 2024 ]	22 %	19.8 %	1.00	1.38	-29 %
Crypto	LLM-MAS [ Luo et al., 2025 ]	33 %	30 %	0.96	1.50	-28 %
Crypto	Proposed	71 %	64 %	0.31	1.73	-26 %
Stock	TAC [ Asadi and Safabakhsh, 2025 ]	10.4 %	5.6 %	0.25	0.74	-28 %
Stock	AlphaMixRL [ Sun, 2024 ]	11.8 %	6.3 %	0.31	0.82	-27 %

Table 11: Comparison with representative state-of-the-art methods in each market.

Adaptive Model Routing

Adaptive model routing selects models or specialized experts per input to balance prediction quality, computational cost, latency, and resource constraints.

See all 91 papers

TradingMoE: Routing the Right Experts in Evolving Markets
Aug 12, 2026Chang Zhou, Xingtong Yu, Minbin Huang +4Mixture-of-Experts Language Models Sparse Mixture-of-Experts
University of Science and Technology of China · The Chinese University of Hong Kong · The University of Tokyo +1
Regime-Adaptive Continual Learning for Portfolio Management
May 29, 2026Chaofan Pan, Lingfei Ren, Linbo Xiong +3Continual Learning Portfolio Optimization
Southwestern University of Finance and Economics · Shanxi University
PortBench: A Correlation-Aware, Full-Pipeline Benchmark for LLM-Driven Portfolio Management
May 27, 2026Yuxuan Zhao, Sijia Chen, Ningxin SuLLM Evaluation Quantitative Finance
Yantai Research Institute of Harbin Engineering University · The Hong Kong University of Science and Technology (Guangzhou)

Aug 12, 2026J/K move · Enter open · S save

TradingMoE: Routing the Right Experts in Evolving Markets

Chang Zhou, Xingtong Yu, Minbin Huang, Zhennan Wu+3

University of Science and Technology of China · The Chinese University of Hong Kong · The University of Tokyo+1

Mixture-of-Experts Language Models Sparse Mixture-of-Experts Expert Routing LLM Routing Algorithmic Trading

Open paper AskPDF arXiv

Large language models (LLMs) have shown strong potential for financial analysis and trading, but direct trading remains challenging because the predictive capabilities required can vary across assets, decision fields, and market conditions. Existing LLM-based trading systems either coordinate human-defined external experts or adopt conventional internal Mixture-of-Experts (MoE) routers that do not directly evaluate how individual experts contribute to trading decisions. Moreover, these routers receive no direct signal indicating when an inactive expert has become more suitable as market conditions change. We find that native router scores poorly reflect how much individual experts improve trading decisions, frequently leaving better alternatives unselected. We further reveal that token-specific expert usefulness exhibits a compact low-dimensional structure. Based on these findings, we propose TradingMoE, a trading-oriented sparse MoE that augments a frozen dense LLM with lightweight residual experts. We introduce a Query-Key router that represents the expertise required by each token under the current market context as a low-dimensional query and matches it with learnable expert keys. We further propose a sparse expert selection update mechanism that samples a few inactive experts during training and estimates whether they should replace the weakest expert in the current Top-k route. This mechanism enables the router to update expert selection as market conditions change while preserving sparse computation. Experiments against 22 baselines on stock and cryptocurrency markets show that TradingMoE improves cumulative return over the best-performing baselines by 30.89% and 30.7%, respectively. Rolling paper-trading experiments further demonstrate that its advantage persists under forward-only deployment.

Abstract

Adaptive Model Routing

Explore similar work

TradingMoE: Routing the Right Experts in Evolving Markets

Regime-Adaptive Continual Learning for Portfolio Management

PortBench: A Correlation-Aware, Full-Pipeline Benchmark for LLM-Driven Portfolio Management