Semantic Faithfulness and Entropy Production Measures to Tame Your LLM Demons and Manage Hallucinations

2025-12-08T15:12:35Z

Evaluating faithfulness of Large Language Models (LLMs) to a given task is a complex challenge. We propose two new unsupervised metrics for faithfulness evaluation using insights from information theory and thermodynamics. Our approach treats an LLM as a bipartite information engine where hidden layers act as a Maxwell demon controlling transformations of context $C $ into answer $A$ via prompt $Q$. We model Question-Context-Answer (QCA) triplets as probability distributions over shared topics. Topic transformations from $C$ to $Q$ and $A$ are modeled as transition matrices ${\bf Q}$ and ${\bf A}$ encoding the query goal and actual result, respectively. Our semantic faithfulness (SF) metric quantifies faithfulness for any given QCA triplet by the Kullback-Leibler (KL) divergence between these matrices. Both matrices are inferred simultaneously via convex optimization of this KL divergence, and the final SF metric is obtained by mapping the minimal divergence onto the unit interval [0,1], where higher scores indicate greater faithfulness. Furthermore, we propose a thermodynamics-based semantic entropy production (SEP) metric in answer generation, and show that high faithfulness generally implies low entropy production. The SF and SEP metrics can be used jointly or separately for LLM evaluation and hallucination control. We demonstrate our framework on LLM summarization of corporate SEC 10-K filings.

BondBERT: What we learn when assigning sentiment in the bond market

2025-12-08T09:21:22Z

Bond markets respond differently to macroeconomic news compared to equity markets, yet most sentiment models are trained primarily on general financial or equity news data. However, bond prices often move in the opposite direction to economic optimism, making general or equity-based sentiment tools potentially misleading. We introduce BondBERT, a transformer-based language model fine-tuned on bond-specific news. BondBERT can act as the perception and reasoning component of a financial decision-support agent, providing sentiment signals that integrate with forecasting models. We propose a generalisable framework for adapting transformers to low-volatility, domain-inverse sentiment tasks by compiling and cleaning 30,000 UK bond market articles (2018-2025). BondBERT's sentiment predictions are compared against FinBERT, FinGPT, and Instruct-FinGPT using event-based correlation, up/down accuracy analyses, and LSTM forecasting across ten UK sovereign bonds. We find that BondBERT consistently produces positive correlations with bond returns, and achieves higher alignment and forecasting accuracy than the three baseline models. These results demonstrate that domain-specific sentiment adaptation better captures fixed income dynamics, bridging a gap between NLP advances and bond market analytics.

DeepSVM: Learning Stochastic Volatility Models with Physics-Informed Deep Operator Networks

2025-12-08T04:53:23Z

Real-time calibration of stochastic volatility models (SVMs) is computationally bottlenecked by the need to repeatedly solve coupled partial differential equations (PDEs). In this work, we propose DeepSVM, a physics-informed Deep Operator Network (PI-DeepONet) designed to learn the solution operator of the Heston model across its entire parameter space. Unlike standard data-driven deep learning (DL) approaches, DeepSVM requires no labelled training data. Rather, we employ a hard-constrained ansatz that enforces terminal payoffs and static no-arbitrage conditions by design. Furthermore, we use Residual-based Adaptive Refinement (RAR) to stabilize training in difficult regions subject to high gradients. Overall, DeepSVM achieves a final training loss of $10^{-5}$ and predicts highly accurate option prices across a range of typical market dynamics. While pricing accuracy is high, we find that the model's derivatives (Greeks) exhibit noise in the at-the-money (ATM) regime, highlighting the specific need for higher-order regularization in physics-informed operator learning.

Order-Flow Filtration and Directional Association with Short-Horizon Returns

2025-12-08T04:09:43Z

Electronic markets generate dense order flow with many transient orders, which degrade directional signals derived from the limit order book (LOB). We study whether simple structural filters on order lifetime, modification count, and modification timing sharpen the association between order book imbalance (OBI) and short-horizon returns in BankNifty index futures, where unfiltered OBI is already known to be a strong short-horizon directional indicator. The efficacy of each filter is evaluated using a three-step diagnostic ladder: contemporaneous correlations, linear association between discretised regimes, and Hawkes event-time excitation between OBI and return regimes. Our results indicate that filtration of the aggregate order flow produces only modest changes relative to the unfiltered benchmark. By contrast, when filters are applied on the parent orders of executed trades, the resulting OBI series exhibits systematically stronger directional association. Motivated by recent regulatory initiatives to curb noisy order flow, we treat the association between OBI and short-horizon returns as a policy-relevant diagnostic of market quality. We then compare unfiltered and filtered OBI series, using tick-by-tick data from the National Stock Exchange of India, to infer how structural filters on the order flow affect OBI-return dynamics in an emerging market setting.

Unveiling Hedge Funds: Topic Modeling and Sentiment Correlation with Fund Performance

2025-12-07T01:52:51Z

The hedge fund industry presents significant challenges for investors due to its opacity and limited disclosure requirements. This pioneering study introduces two major innovations in financial text analysis. First, we apply topic modeling to hedge fund documents-an unexplored domain for automated text analysis-using a unique dataset of over 35,000 documents from 1,125 hedge fund managers. We compared three state-of-the-art methods: Latent Dirichlet Allocation (LDA), Top2Vec, and BERTopic. Our findings reveal that LDA with 20 topics produces the most interpretable results for human users and demonstrates higher robustness in topic assignments when the number of topics varies, while Top2Vec shows superior classification performance. Second, we establish a novel quantitative framework linking document sentiment to fund performance, transforming qualitative information traditionally requiring expert interpretation into systematic investment signals. In sentiment analysis, contrary to expectations, the general-purpose DistilBERT outperforms the finance-specific FinBERT in generating sentiment scores, demonstrating superior adaptability to diverse linguistic patterns found in hedge fund documents that extend beyond specialized financial news text. Furthermore, sentiment scores derived using DistilBERT in combination with Top2Vec show stronger correlations with subsequent fund performance compared to other model combinations. These results demonstrate that automated topic modeling and sentiment analysis can effectively process hedge fund documents, providing investors with new data-driven decision support tools.

Exploiting Supply Chain Interdependencies for Stock Return Prediction: A Full-State Graph Convolutional LSTM

2025-12-07T01:43:39Z

Stock return prediction is fundamental to financial decision-making, yet traditional time series models fail to capture the complex interdependencies between companies in modern markets. We propose the Full-State Graph Convolutional LSTM (FS-GCLSTM), a novel temporal graph neural network that incorporates value-chain relationships to enhance stock return forecasting. Our approach features two key innovations: First, we represent inter-firm dependencies through value-chain networks, where nodes correspond to companies and edges capture supplier-customer relationships, enabling the model to leverage information beyond historical price data. Second, FS-GCLSTM applies graph convolutions to all LSTM components - current input features, previous hidden states, and cell states - ensuring that spatial information from the value-chain network influences every aspect of the temporal update mechanism. We evaluate FS-GCLSTM on Eurostoxx 600 and S&P 500 datasets using LSEG value-chain data. While not achieving the lowest traditional prediction errors, FS-GCLSTM consistently delivers superior portfolio performance, attaining the highest annualized returns, Sharpe ratios, and Sortino ratios across both markets. Performance gains are more pronounced in the denser Eurostoxx 600 network, and robustness tests confirm stability across different input sequence lengths, demonstrating the practical value of integrating value-chain data with temporal graph neural networks.

Market Reactions and Information Spillovers in Bank Mergers: A Multi-Method Analysis of the Japanese Banking Sector

2025-12-06T19:47:12Z

Major bank mergers and acquisitions (M&A) transform the financial market structure, but their valuation and spillover effects remain open to question. This study examines the market reaction to two M&A events: the 2005 creation of Mitsubishi UFJ Financial Group following the Financial Big Bang in Japan, and the 2018 merger involving Resona Holdings after the global financial crisis. The multi-method analysis in this research combines several distinct methods to explore these M&A events. An event study using the market model, the capital asset pricing model (CAPM), and the Fama-French three-factor model is implemented to estimate cumulative abnormal returns (CAR) for valuation purposes. Vector autoregression (VAR) models are used to test for Granger causality and map dynamic effects using impulse response functions (IRFs) to investigate spillovers. Propensity score matching (PSM) helps provide a causal estimate of the average treatment effect on the treated (ATT). The analysis detected a significant positive market reaction to the mergers. The findings also suggest the presence of prolonged positive spillovers to other banks, which may indicate a synergistic effect among Japanese banks. Combining these methods provides a unique perspective on M&A events in the Japanese banking sector, offering valuable insights for investors, managers, and regulators concerned with market efficiency and systemic stability

Optimal Investment, Consumption, and Insurance with Durable Goods under Stochastic Depreciation Risk

2025-12-06T13:08:43Z

We study an infinite-horizon optimal investment, consumption and insurance problem for an economic agent who consumes a perishable and a durable good. The agent trades in a risk-free asset, a risky asset, and a durable good whose price follows a correlated diffusion, while the stock of the durable good depreciates deterministically and is subject to insurable Poisson loss shocks. The agent can partially hedge these shocks via an insurance contract with loading and chooses optimal perishable consumption, portfolio holdings, and insurance coverage to maximise expected discounted CRRA utility. Exploiting the homogeneity of the problem, we reduce the Hamilton--Jacobi--Bellman equation to a static one-dimensional optimisation over constant portfolio shares and derive a semi-explicit optimal strategy. We then prove a verification theorem for the associated jump-diffusion wealth process with insurance, establishing the existence and optimality of this constant-fraction strategy under explicit transversality conditions for both risk-aversion regimes $0<γ<1$ and $γ>1$. Numerical experiments illustrate the impact of stochastic depreciation risk and insurance loading on the optimal allocation to financial assets, durable goods, and insurance coverage.

The Red Queen's Trap: Limits of Deep Evolution in High-Frequency Trading

2025-12-05T19:30:26Z

The integration of Deep Reinforcement Learning (DRL) and Evolutionary Computation (EC) is frequently hypothesized to be the "Holy Grail" of algorithmic trading, promising systems that adapt autonomously to non-stationary market regimes. This paper presents a rigorous post-mortem analysis of "Galaxy Empire," a hybrid framework coupling LSTM/Transformer-based perception with a genetic "Time-is-Life" survival mechanism. Deploying a population of 500 autonomous agents in a high-frequency cryptocurrency environment, we observed a catastrophic divergence between training metrics (Validation APY $>300\%$) and live performance (Capital Decay $>70\%$). We deconstruct this failure through a multi-disciplinary lens, identifying three critical failure modes: the overfitting of \textit{Aleatoric Uncertainty} in low-entropy time-series, the \textit{Survivor Bias} inherent in evolutionary selection under high variance, and the mathematical impossibility of overcoming microstructure friction without order-flow data. Our findings provide empirical evidence that increasing model complexity in the absence of information asymmetry exacerbates systemic fragility.

Predicting Price Movements in High-Frequency Financial Data with Spiking Neural Networks

2025-12-05T16:44:43Z

Modern high-frequency trading (HFT) environments are characterized by sudden price spikes that present both risk and opportunity, but conventional financial models often fail to capture the required fine temporal structure. Spiking Neural Networks (SNNs) offer a biologically inspired framework well-suited to these challenges due to their natural ability to process discrete events and preserve millisecond-scale timing. This work investigates the application of SNNs to high-frequency price-spike forecasting, enhancing performance via robust hyperparameter tuning with Bayesian Optimization (BO). This work converts high-frequency stock data into spike trains and evaluates three architectures: an established unsupervised STDP-trained SNN, a novel SNN with explicit inhibitory competition, and a supervised backpropagation network. BO was driven by a novel objective, Penalized Spike Accuracy (PSA), designed to ensure a network's predicted price spike rate aligns with the empirical rate of price events. Simulated trading demonstrated that models optimized with PSA consistently outperformed their Spike Accuracy (SA)-tuned counterparts and baselines. Specifically, the extended SNN model with PSA achieved the highest cumulative return (76.8%) in simple backtesting, significantly surpassing the supervised alternative (42.54% return). These results validate the potential of spiking networks, when robustly tuned with task-specific objectives, for effective price spike forecasting in HFT.

A Unified AI System For Data Quality Control and DataOps Management in Regulated Environments

2025-12-05T09:33:43Z

In regulated domains such as finance, the integrity and governance of data pipelines are critical - yet existing systems treat data quality control (QC) as an isolated preprocessing step rather than a first-class system component. We present a unified AI-driven Data QC and DataOps Management framework that embeds rule-based, statistical, and AI-based QC methods into a continuous, governed layer spanning ingestion, model pipelines, and downstream applications. Our architecture integrates open-source tools with custom modules for profiling, audit logging, breach handling, configuration-driven policies, and dynamic remediation. We demonstrate deployment in a production-grade financial setup: handling streaming and tabular data across multiple asset classes and transaction streams, with configurable thresholds, cloud-native storage interfaces, and automated alerts. We show empirical gains in anomaly detection recall, reduction of manual remediation effort, and improved auditability and traceability in high-throughput data workflows. By treating QC as a system concern rather than an afterthought, our framework provides a foundation for trustworthy, scalable, and compliant AI pipelines in regulated environments.

Convolution-FFT for option pricing in the Heston model

2025-12-05T00:09:29Z

We propose a convolution-FFT method for pricing European options under the Heston model that leverages a continuously differentiable representation of the joint characteristic function. Unlike existing Fourier-based methods that rely on branch-cut adjustments or empirically tuned damping parameters, our approach yields a stable integrand even under large frequency oscillations. Crucially, we derive fully analytical error bounds that quantify both truncation error and discretization error in terms of model parameters and grid settings. To the best of our knowledge, this is the first work to provide such explicit, closed-form error estimates for an FFT-based convolution method specialized to the Heston model. Numerical experiments confirm the theoretical rates and illustrate robust, high-accuracy option pricing at modest computational cost.

Visibility-Graph Asymmetry as a Structural Indicator of Volatility Clustering

2025-12-03T20:34:19Z

Volatility clustering is one of the most robust stylized facts of financial markets, yet it is typically detected using moment-based diagnostics or parametric models such as GARCH. This paper shows that clustered volatility also leaves a clear imprint on the time-reversal symmetry of horizontal visibility graphs (HVGs) constructed on absolute returns in physical time. For each time point, we compute the maximal forward and backward visibility distances, $L^{+}(t)$ and $L^{-}(t)$, and use their empirical distributions to build a visibility-asymmetry fingerprint comprising the Kolmogorov--Smirnov distance, variance difference, entropy difference, and a ratio of extreme visibility spans. In a Monte Carlo study, these HVG asymmetry features sharply separate volatility-clustered GARCH(1,1) dynamics from i.i.d.\ Gaussian noise and from randomly shuffled GARCH series that preserve the marginal distribution but destroy temporal dependence; a simple linear classifier based on the fingerprint achieves about 90\% in-sample accuracy. Applying the method to daily S\&P500 data reveals a pronounced forward--backward imbalance, including a variance difference $Δ\mathrm{Var}$ that exceeds the simulated GARCH values by two orders of magnitude and vanishes after shuffling. Overall, the visibility-graph asymmetry fingerprint emerges as a simple, model-free, and geometrically interpretable indicator of volatility clustering and time irreversibility in financial time series.

Governance, productivity and economic development

2025-12-03T09:37:12Z

This paper explores the interplay between transfer policies, R\&D, corruption, and economic development using a general equilibrium model with heterogeneous agents and a government. The government collects taxes, redistributes fiscal revenues, and undertakes public investment (in R\&D, infrastructure, etc.). Corruption is modeled as a fraction of tax revenues that is siphoned off and removed from the economy. We first establish the existence of a political-economic equilibrium. Then, using an analytically tractable framework with two private agents, we examine the effects of corruption and evaluate the impact of various policies, including redistribution and innovation-led strategies.

A High-Level Framework for Practically Model-Independent Pricing

2025-12-02T17:32:27Z

We present a high-level framework that explains why, in practice, different pricing models calibrated to the same vanilla surface tend to produce similar valuations for exotic derivatives. Our approach acts as an overlay on the Monte Carlo infrastructure already used in banks, combining path reweighting with a conic optimisation layer without requiring any changes to existing code. This construction delivers narrow, practically model-independent price bands for exotics, reconciling front-office practice with the robust, model-independent ideas developed in the academic literature.