Financial time series are characterized by strong noise, frequent fluctuations, and dense local trend switching, making raw-sequence prediction difficult. This paper investigates stock time-series prediction from the perspective of pre-training sample construction and proposes Multi-step Trend Structure Extraction (MTSE). MTSE identifies local highs and lows in price and related indicator series under different step sizes and combinations, transforming continuous observations into trend-event representations. The occurrence and confirmation positions of events are distinguished, confirmed price-turning events are used as sample anchors, and the latest confirmed events in volume, RSI, and BIAS available before the decision point are aligned within a unified decision table. Using ChiNext Index data at one-minute, 15-minute, 60-minute, daily, and weekly frequencies, we compare a Raw sliding-window pipeline with an MTSE reconstruction pipeline under CatBoost, Transformer, and CNN models. Paired statistical tests, layered ablations, and net-of-cost backtesting are used for evaluation. The results show that MTSE changes sample size, label-switching frequency, and trend-segment organization. Relative to Raw inputs, MTSE improves F1, AUC, and MCC in most models and time granularities. Ablation results indicate that the gains are associated with the joint organization of event filtering, structural information, and multi-indicator alignment. In net-of-cost backtesting, the MTSE pipeline generally provides better risk–return profiles. These findings suggest that pre-training trend-structure reconstruction can improve model learning and financial usability in the ChiNext Index setting.
This study proposes a fraud detection framework for FinTech that integrates tabular deep learning with relational graph reasoning. We address a key limitation of state-of-the-art tabular fraud models: their assumption that transactions are independent, despite evidence that modern fraud often arises from coordinated infrastructure reuse (e.g., shared devices, email domains, accounts, and addresses). Using the IEEE-CIS Fraud Detection dataset (≈590,540 e-commerce transactions), we develop a hybrid architecture that combines TabNet, chosen for its built-in interpretability on tabular data, with a GraphSAGE encoder trained on a leakage-free heterogeneous entity graph constructed strictly from historical training transactions. Unseen entities in validation and test splits are mapped to dedicated unknown nodes, mirroring deployment conditions and preventing optimistic bias. Class imbalance is handled using square-root class weighting, avoiding synthetic oversampling methods that could distort real infrastructure topology. Empirically, the standalone TabNet baseline achieves ROC–AUC = 0.759 and PR–AUC = 0.149 on a forward-in-time test split. Incorporating relational learning via the proposed TabNet–GraphSAGE fusion improves performance to ROC–AUC = 0.791 and PR–AUC = 0.206, representing a 38% relative gain in PR–AUC and substantially enhancing minority fraud retrieval. The graph encoder alone attains meaningful predictive power, and ablation results confirm that the largest gains emerge from learned message passing and fusion, indicating complementarity between behavioural and infrastructural representations. Overall, results demonstrate that fraud detection is fundamentally a hybrid problem: combining local transaction anomalies with global relational structure yields more operationally relevant ranking quality and improved robustness in high-volume FinTech environments.
This paper introduces a novel sentiment-driven framework for modeling asset return volatility and for comparing FIGARCH-based specifications by integrating machine learning and time-series econometric techniques. Leveraging a dataset of over 347,000 asset-specific news headlines and financial data spanning from March 2022 to January 2025, we first develop a Random Forest classifier to assess the likelihood that daily news headlines will significantly affect asset returns, either positively or negatively, across four key time frames: Open-to-Close, Open-to-Open, Close-to-Close, and Close-to-Open. The Random Forest model generates impact probability scores for each of the four time frames, which are then incorporated as exogenous variables into a GARCH model to enhance its predictive power over return volatility. In parallel, we utilize the FinBERT model to extract average sentiment scores from headlines and construct sentiment-based exogenous variables. These features are incorporated into an extended FIGARCH(1, 1) model to examine their predictive contributions to return volatility. Empirical results across 37 assets from stock, cryptocurrency, and commodity markets demonstrate that the inclusion of sentiment and impact probability features improves model performance (measured by AIC, BIC, and log-likelihood), particularly in time frames sensitive to market openings and overnight news. Our findings underscore the value of combining textual sentiment indicators produced by the FinBERT model, which is specifically trained on financial data, alongside directional time frame-based impact probabilities computed by the Random Forest model with advanced volatility modeling to enhance return and volatility models and support informed model selection for asset returns.
We revisit the international conditional Capital Asset Pricing Model (ICCAPM) using a Long Short-Term Memory (LSTM) model to estimate high-dimensional, time-varying covariances with fewer parametric restrictions than standard covariance estimators. Using international equity data, we find a positive and significant price of world covariance risk, in contrast to conventional covariance estimators. The LSTM approach delivers superior out-of-sample covariance forecasts and improves portfolio performance, yielding higher Sharpe ratios. Results also indicate increasing global market integration over time. These findings highlight the importance of flexible covariance estimation for empirical tests of international asset-pricing models.
This paper addresses the gap between improved forecasting accuracy from machine learning and deep learning models and their limited translation into realized portfolio performance under transaction-cost constraints. To bridge this gap, the study proposes a root mean square error (RMSE)-triggered rebalancing framework that links forecast reliability with portfolio adjustment decisions. Using daily data from VN-100 stock sample, the study first evaluates N-BEATS against a conventional ARIMA benchmark for return forecasting. The results show that N-BEATS provides more accurate and stable forecasts than ARIMA, and its predictions are therefore incorporated into Mean-Variance with Forecasting (MVF) and Mean Absolute Deviation with Forecasting (MADF) portfolio optimization models. The proposed framework uses RMSE as a predefined threshold-based trigger: portfolios are re-optimized only when realized forecast errors exceed the selected threshold. Portfolio performance is evaluated under transaction-cost settings using multiple return, risk, turnover, and drawdown metrics. The empirical findings indicate that RMSE-triggered rebalancing can improve net portfolio performance by reducing unnecessary portfolio adjustments while remaining responsive to material forecast deviations. MVF benefits most from moderate RMSE thresholds that balance return responsiveness with transaction-cost control, whereas MADF shows greater robustness under less frequent rebalancing. The study is limited by its focus on RMSE, a single proportional transaction-cost assumption, and a restricted asset universe. Future research should examine alternative error measures, richer cost structures, and broader asset classes. The study contributes by using RMSE as both a diagnostic of forecast reliability and an operational trigger for cost-aware rebalancing.
The rapid digitalization of financial services has enhanced transaction speed and accessibility but also amplified exposure to fraud activities that undermine institutional integrity and consumer trust. Effective fraud detection requires analytical frameworks that are not only accurate and adaptive but also interpretable and compliant with financial regulations. This study develops a scalable, explainable machine learning framework for detecting fraud financial transactions using a hybrid ensemble of Extreme Gradient Boosting (XGBoost) and Deep Neural Networks (DNN). The proposed model integrates transactional, temporal, and identity-linked features to capture behavioral and contextual patterns in large-scale, high-frequency data. To address the extreme class imbalance inherent in fraud detection, we evaluate multiple strategies including focal loss, class weighting, and threshold optimization. Model transparency is enhanced through SHAP-based interpretability analysis, providing granular insights into the feature interactions driving fraud risk. Empirical evaluation on a real-world transaction dataset demonstrates that the hybrid ensemble achieves superior detection accuracy and recall relative to baseline models while maintaining explainability suitable for regulated financial environments. The results highlight the potential of combining interpretable machine learning with adaptive ensemble learning to enhance resilience and trustworthiness in modern financial risk management systems.
In the aftermath of the proposal of the Black-Scholes-Merton model in 1973, a considerable number of research groups issued substantial criticisms of the model. An initial observation revealed a discontinuity in stock returns, a phenomenon referred to as the jump diffusion model. A multitude of efforts have been made to develop alternative models that can adequately capture the leptokurtic nature of returns, as evidenced by data observed in financial markets. A substantial proportion of these models presuppose that volatility manifests as a stochastic process; a notable illustration of this is the Heston model, which was proposed in 1993. Over the course of several decades, researchers have proposed a variety of modifications to the BS model. Among these modifications, there have been proposals to incorporate fractional and conformable derivatives into the solution of the partial differential equation. In contrast, the extant literature concerning the modification of the traditional Heston model is limited. Consequently, the present study exclusively focuses on the traditional Heston model and its modifications, specifically the Heston model augmented by He and Zhu (HZ), as well as the novel model developed within this research. The incorporation of conformable derivatives from Khalil et al. and Anderson et al. into this model has yielded the Conformable Heston model, a derivative that significantly outperforms the approximation of European call options’ price. Furthermore, an empirical analysis was conducted, encompassing two base samples and a third out-of-sample. The findings of our study suggest that the Heston Conformable model exhibits a significantly superior ability to approximate the market price of the contracts in the three samples when compared with the traditional Heston model and with the model proposed by HZ. In the empirical analysis section, the stability and positivity of the Conformable Heston model on Mexican stock contracts was tested. It has been demonstrated that the magnitude of error is negligible. This is substantiated by the observation that the median value is equivalent to 0.00, the maximum absolute error is approximately 0.33, and the number of observations that fall outside the 95th percentile is minimal. Furthermore, in Section 5, we employed the Diebold-Mariano test with the intention of validating the results. To this end, a comparison was made between the He and Zhu model and the Heston Conformable model. To this end, we employed samples of option contracts from international companies, in addition to out-of-sample data. The findings of this investigation indicate that the General Conformable model is to be preferred for every maturity date, as evidenced by the results of the Diebold-Mariano test, with significance levels tending towards 0.99. It is important to note, however, that these significance levels must be interpreted with caution, particularly in the context of maturities for which observations are scarce. In summary, the Conformable Heston model, and the conformable parameters are positive and stable, and the price of the model that meets the criteria is in close agreement with the prevailing market price. A further investigation was conducted by examining the option contracts of international companies, which resulted in the calculation of the mean square error of implied volatility. This analysis also yielded the conclusion that the Conformable Heston model provides the lowest value.
The emergence of cryptocurrency perpetual futures has created unique opportunities for automated market making, yet existing approaches often ignore the funding rate mechanism that transfers payments between long and short position holders every eight hours. This paper develops a reinforcement learning framework that jointly optimizes market making spread decisions and funding rate capture, formulating the problem as a Markov Decision Process with state features tailored to perpetual futures, including funding rates, inventory, realized volatility, and an order-flow imbalance proxy. This state design is a foundational methodological contribution because it embeds crypto-native contract mechanics directly into the agent’s information set rather than treating funding as an exogenous overlay. The empirical evaluation uses public Binance Futures BTCUSDT perpetual data, comprising 29,606 hourly observations from 31 December 2022 to 18 May 2026 and 3,701 funding-rate records. After replacing the synthetic proof of concept with historical data, the best reinforcement learning policy remains only marginally positive under the conservative Sharpe definition. However, a performance-first adaptive market making variant with volatility filtering, fee-aware quoting, and momentum-based inventory targeting achieves an annualized return of 24.63%, Sharpe ratio of 1.49, zero-risk-free Sharpe ratio of 1.87, and maximum drawdown of 6.39% on the final historical holdout. These findings indicate that the proposed state representation is useful for studying perpetual-futures control, but profitability depends critically on explicit fee, regime, and inventory filters. The paper therefore contributes a reproducible real-data benchmark and identifies the risk controls, execution assumptions, latency constraints, and multi-venue extensions required before reinforcement learning market making can be considered production-ready.
This article explores the extent to which established sector classifications continue to provide meaningful insights into the financial underpinnings of firms. Utilizing a contemporary cross-sectional sample of 500 unique S&P 500 firms, the study constructs a parsimonious set of financial-statement-based ratios capturing profitability, leverage, liquidity, interest coverage, efficiency, and cash-generation capacity. These indicators are analyzed through a combination of non-parametric inference, supervised machine learning, and unsupervised clustering.The findings indicate that sector membership accounts for only a portion of the cross-sectional variation in firms' financial ratios. Supervised-learning models have been shown to achieve a maximum validation accuracy of 49.3%, with the broader set of imbalance-sensitive metrics reported across models likewise confirming only partial recoverability of sector labels from accounting data. Concurrently, the unsupervised clustering process yielded nine economically interpretable groups that transcend conventional sector boundaries, thereby reflecting distinct combinations of profitability, leverage, liquidity, and cash-generation patterns. The findings of this study indicate that these accounting-based clusters exhibit lower within-group dispersion than sector labels in most of the ratios examined. An exploratory temporal comparison suggests only moderate persistence of group membership across subsequent annual reporting periods.The findings indicate that conventional sector designations persist in their usefulness, yet they prove inadequate in fully encapsulating the accounting structures of firms. Consequently, accounting-based clustering offers an economically interpretable complementary perspective on firm similarity, especially when the objective is to compare firms based on financial structure rather than nominal industry membership alone. As the analysis is primarily cross-sectional, future research may further examine the temporal stability of these accounting-based peer groups and test their portability across alternative firm universes and market settings.
Applying concepts related to zero-shot meta-learning and pre-training of foundation models, we develop a meta reinforcement learning approach (denoted MetaRL) that is pre-trained on thousands of goals-based wealth management (GBWM) problems. Each GBWM problem involves a multiple year scenario over which the investor looks to optimally choose an investment portfolio each year and choose to fulfill all, some, or none of the different financial goals that arise each year. These choices seek to maximize the expected total investor utility obtained from the fulfilled financial goals. By eliminating separate training and optimization for each new investor problem, the MetaRL model in inference mode produces near-optimal dynamic investment portfolio and goal-fulfilling strategies for a new GBWM problem within a few hundredths of a second. This delivers expected utilities that are, on average, 97.8% of the optimal expected utilities (determined via Dynamic Programming). These results are remarkably robust to capital market regime changes, even when training uses only one capital market regime. Further, the MetaRL approach can enable solving problems with larger state spaces where Dynamic Programming becomes computationally infeasible.
We develop a rotation-invariant neural network that provides the global minimum-variance portfolio by jointly learning how to lag-transform historical returns and marginal volatilities and how to regularise the eigenvalues of large equity covariance matrices. This explicit mathematical mapping offers clear interpretability of each module's role, so the model cannot be regarded as a pure black box. The architecture mirrors the analytical form of the global minimum-variance solution yet remains agnostic to dimension, so a single model can be calibrated on panels of a few hundred stocks and applied, without retraining, to one thousand US equities, a crosssectional jump that indicates robust generalization capability. The loss function is the future short-term realized minimum variance and is optimized end-to-end on real returns. In out-of-sample tests from January 2000 to December 2024, the estimator delivers systematically lower realized volatility, smaller maximum drawdowns, and higher Sharpe ratios than the best competitors, including state-of-the-art non-linear shrinkage, and these advantages persist across both short and long evaluation horizons despite the model's short-term training focus. Furthermore, although the model is trained end-to-end to produce an unconstrained minimum-variance portfolio, we show that its learned covariance representation can be used in general optimizers under long-only constraints with virtually no loss in its performance advantage over competing estimators. These advantages persist when the strategy is executed under a highly realistic implementation framework that models market orders at the auctions, empirical slippage, exchange fees, and financing charges for leverage, and they remain stable during episodes of acute market stress.
We test whether long memory in interest rate futures reflects intrinsic dynamics or calendar-time artifact. Using tick-level data from 11 futures contracts, we compare multifractal scaling properties under calendar-time and information clocks. Kernel two-sample tests on 37-dimensional MFDFA feature vectors assess whether these distributions differ across clock types.Information clocks reduce volatility clustering by 90.5% on average, independent of liquidity (R2 = 0.02). The shift is structural: marginal returns are indistinguishable across clocks, but joint distributions with lagged volatility differ significantly (96% vs. 2.3% detection rate). Forecast gains concentrate in illiquid, irregularly traded contracts; liquid benchmarks show no improvement.Most long memory documented in interest rate futures is a sampling artifact, not an intrinsic property. Calendar-time volatility models require re-specification, not recalibration: the dependence structure changes, not only parameter scale.
This paper presents numerical solvers for generative and hybrid option pricing models that unify econometric and diffusion-based approaches. These models are formulated as systems of continuous partial differential equations (PDEs), with stochastic volatility updated at discrete reset dates according to generalized autoregressive conditional heteroskedasticity (GARCH) dynamics. In contrast to approaches that estimate volatility from option prices using the Black–Scholes model or Monte Carlo simulations, this method simplifies option pricing under stochastic volatility by exogenously supplying and updating latent intraday volatility using discrete return data. We then develop and analyze numerical techniques for solving the resulting system of parabolic partial differential equations, which feature time-varying diffusion coefficients governed by the stochastic volatility paths inferred from the return data. Convergence and stability analyses of the numerical schemes attest to the option pricing accuracy of the proposed framework using available discrete implied volatility samples without compromising its computational accuracy. Several sets of numerical tests using SPX data are presented to illustrate our approach and demonstrate its superiority over other empirically well-tested pricing methods.
This paper develops a data-driven framework combining fractional cointegration and structural break detection to examine long-run interdependence and market leadership among international tobacco equities. Using weekly data from May 2008 to October 2024 for Philip Morris International, Altria, British American Tobacco, Imperial Brands, and Japan Tobacco, the study applies the Fractionally Cointegrated Vector Autoregressive (FCVAR) model of Johansen and Nielsen (2012) integrated with the Bai–Perron multiple-break methodology. The empirical analysis supports the presence of a single fractionally cointegrated equilibrium relationship characterized by long memory and regime-dependent persistence. Three model-implied regime shifts, which align closely in timing with major regulatory and ESG-related events, such as the 2012 WHO-FCTC harmonization, the 2015 expansion of FDA regulation, and the 2021 post-COVID ESG rotation—mark distinct equilibrium regimes in the global tobacco market. Within these regimes, adjustment dynamics indicate recurrent long-run leadership by Japan Tobacco, more heterogeneous leadership patterns for Altria, and a clearly regime-dependent role for Philip Morris International. The integration of fractional modeling and break detection provides a robust data-science approach to disentangling persistence from structural change, offering new insights into leadership cycles, systemic risk, and sectoral resilience. These findings underscore how regulated and ESG-sensitive industries evolve through adaptive equilibrium processes, contributing to the broader literature on long-memory econometrics and data-driven financial analytics.
Pairs trading is a well-studied statistical arbitrage strategy including the identification of asset pairs exhibiting correlated changes in their historical prices. This statistical arbitrage strategy focuses on benefiting from non-permanent divergent behaviour of price, and it forecasts that the price relationship will revert to its usual and normal correlation. In this paper, we explore how more recent time-series-based deep learning techniques can be utilized in pairs trading, where cointegrated asset pairs are taken into account. We propose deep-learning and more traditional machine learning-based methods to predict the fluctuation of daily idiosyncratic residual terms between assets and their factor approximations. In our analysis, we focused on seven models: LSTM as a fundamental time-series method to capture interrelationships in dataset, Informer, Autoformer, iTransformer, Scaleformer, and Chronos as transformer-based deep time series methods, and AdaBoost, which is an ensemble learning-based machine learning method. We have assessed the performance of methods comprehensively over S&P 500 and cryptocurrency assets data starting from 2012 to 2020 respectively, and used a traditional statistical arbitrage-based relative value method as a baseline. All of our proposed learning-based methods turned out to be profitable strategies, obtaining higher Sharpe ratios and average returns by outperforming the baseline relative value method. Nevertheless, deep learning-based methods had a lower volume than the baseline, so when transaction costs are taken into account they showed better performance. Deep learning-based methods maximum drawdown was also lower than the traditional statistical arbitrage strategy. As a result, we show the benefits of time series-based deep learning methods in pairs trading across distinct asset classes.
High-frequency returns are of great significance to financial risk management and investment. This study analyzes the dynamics of the high-frequency return distributions. The calculations show that the high-frequency return distribution sequence has a small intrinsic dimension and exhibits time-varying characteristics. We find that all subdistribution sequences included multifractal structures, and the global Rényi index (GRI) series showed that the multifractal features changed over time. We also constructed a benchmark distribution sequence, in which each distribution was sampled from a normal population and the mean and standard deviation of the original series were maintained. The calculations show that the benchmark sequence also includes multifractal features, and the correlation dimension and GRI are smaller than those of the original sequence. However, the calculations show that there is a high correlation between the dimensional series of the benchmark sequence and the original sequence, suggesting that the mean and standard deviation are important parameters affecting fractal dynamics. In particular, we use the UMAP algorithm to show the low-dimensional representation of the distribution sequence, and find that the small-dimensional subsequences include earthworm-like clusters, while the subsequences with large-dimensions include cloud-like clusters. This study provides a new way to analyze the distribution of high-frequency returns, and explores the characteristics of distribution sequences, which is helpful for understanding the dynamics of distributions.
This study examines the role of interdependencies in forecasting sovereign yield spreads in emerging markets using Graph Neural Networks (GNNs). Sovereign yield spreads reflect economic conditions and investor sentiment, making accurate predictions crucial for investors, policymakers, and financial institutions. Traditional forecasting models often treat sovereign risks in isolation, failing to account for financial spillovers and cross-country linkages. By structuring sovereign bonds within a graph-based framework, this study explicitly models these interdependencies to improve predictive accuracy. Using macroeconomic indicators such as GDP, inflation, and foreign exchange reserves, countries are represented as nodes in a financial network, with edges capturing key economic relationships. A Graph Convolutional Network (GCN) is trained to predict sovereign yield spreads, and its performance is benchmarked against a structurally identical feed-forward neural network, where the only difference is the use of graph convolution layers or dense layers. The results show that the GCN model consistently outperforms the feed-forward model, particularly in predicting extreme yield spread movements, demonstrating the importance of accounting for financial interdependencies. Our findings underline the potential of GNNs as a powerful tool in forecasting sovereign yield spreads in emerging markets. Considering the economic impact of these spreads, GNNs could present significant benefits for financial sector stakeholders.
This research delves into the predictive modeling of credit card delinquency by harnessing both credit and debit data, offering a nuanced perspective on consumer financial behavior. The study introduces a novel hierarchical Bayesian regression model that significantly surpasses traditional machine learning algorithms in predictive accuracy. By integrating behavioral aspects of financial decision-making, the model provides a profound understanding of the factors influencing delinquency, such as payment timing and repayment ability.We found that the combination of credit and debit data allows for a more comprehensive assessment of a cardholder's financial behavior and risk potential. The model effectively captures individual variations in financial behavior, making it possible to predict delinquency with higher precision. This approach not only enhances the predictive power but also aids in understanding the underlying patterns of financial behavior that lead to credit risk.The practical implications of this research are substantial for financial institutions, which can leverage these insights to refine risk assessment processes and develop targeted strategies for managing credit risk. The findings advocate for a more informed approach to credit scoring that considers broader behavioral factors, offering a strategic advantage in the competitive financial services market.
Wallstreetbets (WSB) is the perfect echo chamber to study retail investor behaviour and markets. We introduce a direct measure of individual stock attention and the concept of forum-wide attention herding. We fine-tune a large language model to classify investor sentiment. We find that WSB sentiment is inversely related to the VIX. In general, more individual stock attention leads to more stock purchases, and sentiment is a contrarian predictor of future returns. However, when attention herds on a stock with high user engagement, trades peak but there is no reversal in returns. Finally, our monthly attention herding portfolio generates sizable alphas.
The Decentralized Finance (DeFi) ecosystem has experienced over $10 billion in direct losses due to crime events. Beyond these immediate losses, such events often trigger broader market reactions, including price declines, trading activity changes, and reductions in market capitalization. Decentralized Autonomous Organizations (DAOs) govern DeFi applications through tradable governance assets that function like corporate shares for voting and decision-making. Leveraging DeFi’s granular trading data, we conduct an event study on 22 crime events between 2020 and 2022 to assess their economic impact on governance asset prices, trading volumes, and market capitalization. Using a dynamic difference-in-differences (DiD) framework with counterfactual governance assets, we aim for causal inference of intraday temporal effects. Our results show that 55% of crime events lead to significant negative price impacts, with an average decline of about 14%. Additionally, 68% of crime events lead to increased governance asset trading volume. Based on these impacts, we estimate indirect economic losses of over $1.3 billion in DAO market capitalization, far exceeding direct victim costs and accounting for 74% of total losses. Our study provides valuable insights into how crime events shape market dynamics and affect DAOs. Moreover, our methodological approach is reproducible and applicable beyond DAOs, offering a framework to assess the indirect economic impact on other cryptoassets.
