Selected work

Representative Papers

RepICL: Reusable In-Context Prediction Across Heterogeneous Representation Spaces

Oct 05, 2026

This study addresses the need for repeated training of few-shot predictors across heterogeneous representation spaces by proposing RepICL, a meta-trained in-context learner incorporating episodic whitening normalization to achieve a "learn once, reuse many times" prediction paradigm. Furthermore, this work constructs RepShiftBench, the first benchmark demonstrating that a shared few-shot prediction process can generalize to unseen representation spaces, and reveals that episodic whitening serves as a critical inductive bias for performance enhancement. Across twelve benchmark settings, the proposed method consistently outperforms logistic regression and existing baselines, substantially improving few-shot classification accuracy in both cross-dataset and cross-modal scenarios.

0 citationsRead paper

DIVINE: Simple Cross-Market Stock Pretraining via Diverse Indicator Reconstruction

Oct 02, 2026

This study addresses the challenges of noisy supervision signals and misalignment with return prediction in financial time series pre-training by proposing a cross-market pre-training framework. Methodologically, technical indicators are leveraged as stable supervision sources to reconstruct OHLCV data, while joint multi-market pre-training is employed to learn generalizable representations. Subsequently, only a lightweight Transformer encoder is transferred to downstream stock ranking tasks. The findings reveal that supervision signal design and market diversity are more critical than model scale. Remarkably, with merely 0.05M parameters, the proposed framework achieves superior average portfolio performance across six major markets, rivaling large-scale foundation models while maintaining both high efficiency and robustness.

0 citationsRead paper

MAPLE: Efficient and Diverse Multi-Alpha Generation for Portfolio Construction

Jul 27, 2026

This work addresses the limitation of conventional deep learning approaches for stock ranking, which typically produce a single alpha signal and lack explicit control over correlations among multiple alphas, resulting in insufficient portfolio diversity. The authors propose MAPLE, a novel framework that, within a single training run, jointly incorporates a unified capacity-scaled prediction head, an extreme-rank weighted listwise loss, and an explicit diversity regularizer to enable controllable generation of multiple alpha signals with desired correlation structures—all within a single model. Notably, MAPLE achieves this without increasing architectural complexity and is compatible with various backbone networks. Evaluated across four major equity markets in the U.S., China, and Japan, MAPLE significantly outperforms nine baselines, achieving up to 55× fewer parameters and 2.5× faster training while improving Sharpe ratios by 10–23% and Calmar ratios by 17–43%.

0 citationsRead paper

BoostLLM: Boosting-inspired LLM Fine-tuning for Few-shot Tabular Classification

May 07, 2026

This work addresses the underperformance of large language models (LLMs) compared to gradient-boosted trees like XGBoost in few-shot tabular classification tasks. It introduces, for the first time, a boosting-inspired paradigm into LLM fine-tuning by proposing a parameter-efficient training framework. The method iteratively trains lightweight adapters as weak learners through multiple rounds of residual optimization and incorporates structured inductive bias by fusing decision tree paths with original tabular features to form a dual-view input representation. This enables an adaptive transition from path-guided to feature-driven representations. Experimental results demonstrate that the proposed approach significantly outperforms standard fine-tuning across diverse LLMs and tabular datasets, matches or even surpasses XGBoost under varying sample sizes, and achieves superior performance to GPT-4o using a 4B-parameter model.

0 citationsRead paper

Integrating Inductive Biases in Transformers via Distillation for Financial Time Series Forecasting

Mar 17, 2026

Financial time series exhibit non-stationarity and regime-switching behavior, which challenge conventional Transformers due to their implicit stationarity assumptions. To address this limitation, this work proposes the TIPS framework, which dynamically integrates multiple inductive biases—causality, locality, and periodicity—into the Transformer architecture for the first time. Specifically, bias-specific teacher models generate attention masks, and a regime-aware knowledge distillation strategy is devised to enable robust forecasting in non-stationary markets. Empirical evaluations demonstrate that TIPS outperforms strong ensemble baselines by 55% in annualized returns, 9% in Sharpe ratio, and 16% in Calmar ratio across four major stock markets, while requiring only 38% of the baseline’s inference computation.

0 citationsRead paper

Recent publications

Latest Papers

RepICL: Reusable In-Context Prediction Across Heterogeneous Representation Spaces

Oct 05, 2026

This study addresses the need for repeated training of few-shot predictors across heterogeneous representation spaces by proposing RepICL, a meta-trained in-context learner incorporating episodic whitening normalization to achieve a "learn once, reuse many times" prediction paradigm. Furthermore, this work constructs RepShiftBench, the first benchmark demonstrating that a shared few-shot prediction process can generalize to unseen representation spaces, and reveals that episodic whitening serves as a critical inductive bias for performance enhancement. Across twelve benchmark settings, the proposed method consistently outperforms logistic regression and existing baselines, substantially improving few-shot classification accuracy in both cross-dataset and cross-modal scenarios.

0 citationsRead paper

DIVINE: Simple Cross-Market Stock Pretraining via Diverse Indicator Reconstruction

Oct 02, 2026

This study addresses the challenges of noisy supervision signals and misalignment with return prediction in financial time series pre-training by proposing a cross-market pre-training framework. Methodologically, technical indicators are leveraged as stable supervision sources to reconstruct OHLCV data, while joint multi-market pre-training is employed to learn generalizable representations. Subsequently, only a lightweight Transformer encoder is transferred to downstream stock ranking tasks. The findings reveal that supervision signal design and market diversity are more critical than model scale. Remarkably, with merely 0.05M parameters, the proposed framework achieves superior average portfolio performance across six major markets, rivaling large-scale foundation models while maintaining both high efficiency and robustness.

0 citationsRead paper

MAPLE: Efficient and Diverse Multi-Alpha Generation for Portfolio Construction

Jul 27, 2026

This work addresses the limitation of conventional deep learning approaches for stock ranking, which typically produce a single alpha signal and lack explicit control over correlations among multiple alphas, resulting in insufficient portfolio diversity. The authors propose MAPLE, a novel framework that, within a single training run, jointly incorporates a unified capacity-scaled prediction head, an extreme-rank weighted listwise loss, and an explicit diversity regularizer to enable controllable generation of multiple alpha signals with desired correlation structures—all within a single model. Notably, MAPLE achieves this without increasing architectural complexity and is compatible with various backbone networks. Evaluated across four major equity markets in the U.S., China, and Japan, MAPLE significantly outperforms nine baselines, achieving up to 55× fewer parameters and 2.5× faster training while improving Sharpe ratios by 10–23% and Calmar ratios by 17–43%.

0 citationsRead paper

BoostLLM: Boosting-inspired LLM Fine-tuning for Few-shot Tabular Classification

May 07, 2026

This work addresses the underperformance of large language models (LLMs) compared to gradient-boosted trees like XGBoost in few-shot tabular classification tasks. It introduces, for the first time, a boosting-inspired paradigm into LLM fine-tuning by proposing a parameter-efficient training framework. The method iteratively trains lightweight adapters as weak learners through multiple rounds of residual optimization and incorporates structured inductive bias by fusing decision tree paths with original tabular features to form a dual-view input representation. This enables an adaptive transition from path-guided to feature-driven representations. Experimental results demonstrate that the proposed approach significantly outperforms standard fine-tuning across diverse LLMs and tabular datasets, matches or even surpasses XGBoost under varying sample sizes, and achieves superior performance to GPT-4o using a 4B-parameter model.

0 citationsRead paper

Integrating Inductive Biases in Transformers via Distillation for Financial Time Series Forecasting

Mar 17, 2026

Financial time series exhibit non-stationarity and regime-switching behavior, which challenge conventional Transformers due to their implicit stationarity assumptions. To address this limitation, this work proposes the TIPS framework, which dynamically integrates multiple inductive biases—causality, locality, and periodicity—into the Transformer architecture for the first time. Specifically, bias-specific teacher models generate attention masks, and a regime-aware knowledge distillation strategy is devised to enable robust forecasting in non-stationary markets. Empirical evaluations demonstrate that TIPS outperforms strong ensemble baselines by 55% in annualized returns, 9% in Sharpe ratio, and 16% in Calmar ratio across four major stock markets, while requiring only 38% of the baseline’s inference computation.

0 citationsRead paper