Dynamic Forecasting and Temporal Feature Evolution of Stock Repurchases in Listed Companies Using Attention-Based Deep Temporal Networks

Xiang Ao 2,*,† , Jingxuan Zhang 1,* , Xinyu Zhao 1,* 1 School of Economics and Management, Beijing Jiaotong University, Beijing, China 2 School of Software Engineering, Beijing Jiaotong University, Beijing, China [email protected], {23241094, 23241229}@bjtu.edu.cn · 2026

Read Dynamic Forecasting and Temporal Feature Evolution of Stock Repurchases in Listed Companies Using Attention-Based Deep Temporal Networks by Xiang Ao 2,*,† , Jingxuan Zhang 1,* , Xinyu Zhao 1,* 1 School of Economics and Management, Beijing Jiaotong University, Beijing, China 2 School of Software Engineering, Beijing Jiaotong University, Beijing, China [email protected], {23241094, 23241229}@bjtu.edu.cn (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-09650.txt.

Introduction

Ao, Zhang, and Zhao build a dynamic early-warning system for stock repurchases, where static models miss the shifting financial patterns behind corporate decisions. Using Chinese A-share panel data from 2014 to 2024, they pair a Temporal Convolutional Network with an attention-based LSTM for tracking long- and short-term financial evolution. Rolling-window tests show the hybrid beating static baselines such as Logistic Regression and XGBoost. Explainable-AI analysis finds prolonged undervaluation as the long-term motive, with a sharp cash-flow rise as the decisive short-term trigger.

Catalog record

CREATORXiang Ao 2,*,† , Jingxuan Zhang 1,* , Xinyu Zhao 1,* 1 School of Economics and Management, Beijing Jiaotong University, Beijing, China 2 School of Software Engineering, Beijing Jiaotong University, Beijing, China [email protected], {23241094, 23241229}@bjtu.edu.cn
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE62234
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.09650

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Dynamic Forecasting and Temporal Feature Evolution of Stock Repurchases in Listed Companies Using Attention-Based Deep Temporal Networks Xiang Ao 2,*,† , Jingxuan Zhang 1,* , Xinyu Zhao 1,* 1 School of Economics and Management, Beijing Jiaotong University, Beijing, China 2 School of Software Engineering, Beijing Jiaotong University, Beijing, China [email protected], {23241094, 23241229}@bjtu.edu.cn Abstract In recent years, stock repurchases have emerged as a critical financial strategy in capital markets for boosting investor confidence, optimizing capital structure, and signaling severe market undervaluation. Accurately anticipating and identifying firms’ potential repurchase intentions can not only capture significant excess returns (Alpha) for quantitative investment but also holds profound practical significance for constructing modernized financial risk supervision and intelligent decision-making systems. However, traditional empirical econometric models and static machine learning methods mostly rely on cross-sectional data assumptions, failing to adequately capture the non-linear temporal dependencies and lagging effects of evolving corporate financial conditions on repurchase decisions. To address this gap, this paper proposes a dynamic early warning system for stock repurchases that integrates deep temporal networks with economic theory. Based on large-scale, multidimensional financial and trading panel data of Chinese A-share listed companies from 2014 to 2024, this study reconstructs static economic proxy variables (e.g., Tobin’s Q, net operating cash flow, debt-to-asset ratio) into multivariate dynamic sliding window sequences. On this basis, we construct a deep prediction engine combining a Temporal Convolutional Network (TCN) and an Attention-based Long Short-Term Memory network (Attention-LSTM) to accurately capture the long- and short-term evolutionary patterns of multidimensional financial features. Rigorous rolling-window cross-validation results demonstrate that this deep temporal system outperforms traditional baseline models based on logistic regression and static XGBoost in terms of AUC, PR-AUC, and prediction accuracy across various lead times, effectively achieving intertemporal forward-looking early warnings for repurchase strategies. Furthermore, leveraging Explainable Artificial Intelligence (XAI) techniques, this study extracts the temporal attention weight evolution of core economic indicators 1 to 3 years prior to the decision. Micro-level algorithmic evidence reveals that “undervaluation” (e.g., the prolonged depression of Tobin’s Q) constitutes the long-term underlying motive driving repurchases, whereas a sharp improvement in “cash flow abundance” plays a decisive triggering role at the critical point of decision-making. This research not only provides an advanced deep learning algorithmic paradigm for intelligent financial supervision and cutting-edge quantitative investment but also supplements the “Undervaluation Hypothesis” and the “Free Cash Flow Hypothesis” in classical corporate finance with dynamic empirical evidence characterized by temporal depth. Keywords :Stock Repurchase ; Time Series Classification ; Deep Learning ; Attention Mechanism ; Explainable Artificial Intelligence (XAI) ; Undervaluation Hypothesis * * footnotetext: These authors contributed equally to this work. † † footnotetext: Corresponding author. 1 Introduction In modern capital markets fraught with uncertainty, rational capital allocation remains a core issue in contemporary corporate finance research. In recent years, driven by drastic fluctuations in global economic cycles and profound transformations in capital market structures, stock repurchases have transcended mere financial adjustment methods. They have evolved into critical tools for corporate management to signal severe undervaluation of intrinsic value to the market, boost investor confidence, and optimize capital structures [ 1 , 2 ] . For external quantitative investors and market regulators, the ability to anticipate and accurately identify the potential repurchase intentions of listed companies not only means capturing highly valuable excess return signals (Alpha), but also holds inestimable practical value for preventing market anomalies and perfecting intelligent risk early-warning systems. However, because corporate decision-making processes are intertwined with complex macroeconomic constraints and micro-financial trade-offs, how to leverage massive datasets to precisely forecast this major financial decision remains a significant challenge in the field of quantitative financial prediction. Throughout the development of capital markets, an extensive body of literature has attempted to reveal and predict corporate financial decisions. Early studies primarily relied on traditional econometric models (e.g., logistic regression and Probit models). These methods, with their strong statistical interpretability, laid the empirical foundation for the “Undervaluation Hypothesis” and the “Free Cash Flow Hypothesis” [ 3 , 4 , 5 ] . Currently, with the rapid advancement of computer science, an increasing number of researchers and financial institutions are abandoning traditional subjective experience or simple linear extrapolation, turning instead to machine learning and deep learning algorithms to process multidimensional financial indicators. Related studies have demonstrated the outstanding performance of algorithms such as eXtreme Gradient Boosting (XGBoost), Random Forest, and Support Vector Machines (SVM) in capturing the non-linear interactive relationships among corporate financial proxy variables (e.g., Tobin’s Q and net operating cash flow), thereby significantly improving the accuracy of financial forecasting [ 6 , 7 ] . However, despite the widespread application of static machine learning in financial forecasting, these traditional empirical architectures still suffer from insurmountable theoretical limitations when dealing with structured panel data. The core pain point lies in the fact that static models often treat data from different years as mutually independent “cross-sectional snapshots” stacked together, severely ignoring the autocorrelation and evolutionary logic of corporate financial features along the temporal dimension [ 8 ] . In reality, the strategic decisions of modern corporate boards are by no means based solely on a single year’s financial performance; rather, they are a comprehensive response to the firm’s operational status over several consecutive years (such as long-term valuation depression or continuous cash flow accumulation). Traditional static processing completely loses the highly valuable “lagging effects” and dynamic temporal memory during backpropagation or tree node splitting. To break through the limitations of static assumptions, deep temporal network architectures (such as Long Short-Term Memory networks (LSTM) and Temporal Convolutional Networks (TCN)) have been successively proposed in recent years. By introducing sliding time windows and complex gating/attention mechanisms, these architectures can parallelly extract state evolutions from historical time series, achieving massive success in stock index prediction and high-frequency trading [ 9 , 10 , 11 ] . However, the vast majority of existing empirical studies focus on the pattern recognition of high-frequency trading data. In fact, compared to the instantaneous capture of market sentiment by high-frequency data, low-frequency panel data based on annual or quarterly periods contains a firm’s more essential operational logic and strategic evolutionary trajectory. As a major capital allocation decision, a stock repurchase is often rooted in a financial cycle lasting several years (such as continuous cash flow accumulation or long-term valuation depression) rather than merely reacting to instantaneous market fluctuations. Therefore, evaluating the performance of network models with temporal depth on low-frequency micro-financial data not only reveals the structural evolution of corporate decision-making over long cycles but also fills the research gap regarding deep learning’s application in understanding and warning of complex corporate behaviors. Based on the aforementioned considerations of the long-term temporal dependencies of financial data and complex non-linear decision-making mechanisms, this paper proposes a deep temporal network framework incorporating an attention mechanism [ 12 ] to dynamically forecast the stock repurchase behaviors of Chinese A-share listed companies. Deviating from previous cross-sectional studies, this paper is no longer confined to analyzing isolated single-period data; instead, it reconstructs the multidimensional financial and trading indicators of large-scale listed companies from 2014 to 2024 into continuous three-dimensional (3D) temporal tensors. On this basis, we feed dynamic sliding windows spanning several consecutive years (e.g., three years as a complete cycle) into a deep prediction

Read the complete text