How H3M‑SSMoEs Combines Hypergraph Multimodal Learning and LLM Reasoning to Predict Stock Direction

Stock prediction is challenging due to low signal‑to‑noise ratios, non‑stationarity, complex inter‑stock dependencies, and the need to fuse numerical and textual modalities. Existing models struggle to capture group‑level dynamics, align modalities, and remain computationally efficient.

Problem definition : Build a scalable framework that simultaneously models structural stock relationships, semantic fusion of quantitative and news data, and adaptive regime awareness to improve prediction accuracy, risk control, and returns.

Method

H3M‑SSMoEs consists of four modules:

Context multimodal hypergraph – two hypergraphs capture different temporal scales. The Local Context Hypergraph (LCH) treats each stock‑time instance as a node and builds four sub‑hyperedges (quant‑quant, news‑news, quant‑news, news‑quant) using an adaptive projection network. Edge weights are derived from Jensen‑Shannon divergence to retain informative edges. The Global Context Hypergraph (GCH) aggregates long‑term inter‑stock relations via multi‑head self‑attention and cross‑attention, producing global context features.

LLM‑enhanced reasoning – A frozen Llama‑3.2‑1B (hidden size 2048) is equipped with lightweight adapters. GCH outputs are concatenated, projected into the LLM input space, and processed to obtain high‑dimensional semantic representations that incorporate financial knowledge.

Style‑structured expert mixture (SSMoEs) – A shared market expert captures regime‑level signals (e.g., bull/bear markets) and industry‑specific experts capture sector characteristics (e.g., supply‑chain shocks). Each expert is parameterized by a learnable style vector and activated by a sparse top‑K gating mechanism, producing market‑level and industry‑level embeddings that are concatenated with stock features.

Loss function – Composite loss combines cross‑entropy classification loss with an auxiliary expert‑balance loss that regularizes expert utilization, preventing redundancy.

Datasets and features

Three major indices: DJIA (30 stocks), NASDAQ‑100 (91 stocks), S&P‑100 (99 stocks) from 2020‑01‑01 to 2025‑08‑31, split 7:1:2 for train/val/test.

Numerical features: Yahoo Finance price data plus Alpha158/Alpha360 technical indicators, z‑score normalized.

Textual features: Daily news generated by FinRobot, encoded by Llama‑3.2‑1B.

Timestamp embeddings: Date strings encoded by the same LLM and added to quantitative features.

Baselines and evaluation

15 baselines covering stock‑prediction models (SFM, Adv‑ALSTM, DTML), time‑series models (DLinear, iTransformer, TimeMixer), graph models (GCN, GraphSAGE, GAT), and time‑series LLMs (GPT4TS, aLLM4TS, Time‑LLM).

Metrics: Accuracy (ACC), Precision (PRE), Annualized Return (AR), Sharpe Ratio (SR, risk‑free 2%), Calmar Ratio (CR), Maximum Drawdown (MDD).

Results

H3M‑SSMoEs outperforms all baselines on every dataset. Example: DJIA ACC = 57.47 % (2nd = 57.34 %), PRE = 62.01 % (2nd = 62.44 %); AR = 50.00 % (2nd = 31.70 %), SR = 1.585 (highest), CR = 3.377 (2nd), MDD = 14.81 % (lowest).

NASDAQ‑100 ACC = 58.60 % (highest), PRE = 69.97 % (highest); AR = 70.80 % (2nd = 71.75 %), SR = 2.100 (highest), CR = 4.380 (highest), MDD = 16.17 % (lowest).

S&P‑100 ACC = 56.91 % (highest), PRE = 66.04 % (highest); AR = 29.62 % (2nd = 30.62 %), SR = 1.351 (highest), CR = 2.075 (highest), MDD = 14.27 % (lowest).

Ablation study

Removing LCH drops DJIA AR from 50.00 % to 16.47 % and NASDAQ‑100 CR from 4.380 to 0.331.

Removing the LLM reduces NASDAQ‑100 AR from 70.80 % to 9.78 % and SR from 2.100 to 0.451.

Removing SSMoEs lowers DJIA AR to 16.52 % and NASDAQ‑100 AR to 12.20 %.

Visualization

Portfolio value curves show steady growth across all three indices (NASDAQ‑100 final value ≈ 1.75× initial capital).

Daily return distribution is positively skewed with mean 0.11‑0.23 % and pronounced positive tail.

Maximum drawdown stays within 14.27‑16.17 %, with rapid recovery after short losses, indicating a favorable risk‑return balance.