Papers

Large language models (LLMs) are increasingly used in investment decision-making, yet prior work shows that they exhibit systematic, model-specific investment preferences. We study whether a model's overall investment stance can be calibrated to a specified direction and strength. We introduce an investment-bias dial, an inference-time intervention on a single neuron that continuously adjusts a model-level decision prior—its overall tendency toward buying or selling—without targeting specific firms or investment attributes. Using matched positive and negative evidence, we evaluate five open-weight LLMs and find that the dial produces monotonic changes in investment stance without modifying prompts or model parameters. At the response level, the dial shifts both investment decisions and the evidential emphasis of generated rationales under identical inputs. In an agentic retrieval setting, the dial also changes what information the model searches for, which evidence it selects, and which evidence is reflected in its final analysis. In a long-context evaluation, the dial maintains stable stance control as context length increases, whereas a matched system-prompt instruction progressively attenuates. We further show that changes in the dial propagate to security rankings and downstream portfolio composition in an exploratory backtest. Overall, our results show that an LLM's aggregate investment stance can be calibrated toward a specified target at inference time.

This paper asks whether firms' exposures to news about different types of AI risk are priced in U.S. stock returns. Using AI and risk keywords, I identify 7,787 Wall Street Journal articles from January 2016 to December 2025. I combine latent Dirichlet allocation (LDA) with the Domain Taxonomy in the MIT AI Risk Repository to construct four news-based systematic risk factors. I estimate betas to factor innovations and test pricing with univariate portfolio analysis and Fama-MacBeth regressions. Only the taxonomy-mapped Misinformation factor (D3) is robustly priced. Its high-minus-low beta portfolio earns monthly alphas of 0.49%-0.57%, and the estimated D3 price of risk is positive and statistically significant across beta-estimation windows, conventional factor and industry controls, alternative innovation models, and the pre-ChatGPT subsample. The other factors are not reliably priced, indicating that AI-risk pricing is domain-specific.

Complex systems are often organised into hierarchies whose internal interactions are stronger or faster than interactions across levels. When markets are treated as genuinely multilevel systems it becomes natural to represent them as systems with hierarchical causality. Here we show that reflexivity can be formulated within such a discrete hierarchical causal system; but one in which a higher-level actor state restricts the lower-level transition kernels that remain admissible. Then event dynamics can be separated from calendar embeddings: a set-valued actor-conditioned correspondence can be used to define the admissible family of event kernels, while joint state and waiting-time laws can be used to determine compatible timing to then natural demonstrate reflexivity. A selected event-state law need not determine a unique calendar embedding. Locally, uniqueness of the joint event or timing specification requires uniqueness of both the admissible event kernel and its compatible timing law. Reflexivity is thus the endogenous closure of a hierarchical constraint loop, while timing and projection can generate calendar-time memory or causal ambiguity even for Markov event dynamics.

Option-implied rates are often treated as frictionless because completed boxes deliver riskless payoffs. I show that this interpretation requires option- and benchmark-side implementation wedges to offset. Using SPX and RUT options from 2012-2025, I find a 2-3 bp unannualized increase in the option funding basis when maturity first crosses December 31. The effect behaves as a fixed price wedge, strengthens in the mid-2010s, survives alternative contracts and benchmarks, appears independently in government-bond CIP, and is reproduced in an independently constructed option panel. Put-call parity can identify discount rates precisely without establishing their economic purity.

Investors interpret social disclosures from a risk perspective, yet relevant information can reach them through channels that differ sharply in regulatory enforcement and materiality: SEC filings, sustainability reports, or financial reports. We analyse how social disclosure via each channel relates to idiosyncratic risk. Studying S&P 1,500 constituents, we distinguish between initiated and continued disclosure along the three disclosure channels. We find first-time disclosure of social issues via SEC filings is related to increased idiosyncratic firm risk, highlighting that unexpected information is published. Continuous disclosure of social issues is related to lower idiosyncratic risk for sustainability and financial reports, which is in line with the literature. The SEC filing effect is robust for downside idiosyncratic risk measures, the separation of social disclosure into human capital, product liabilities, and stakeholder engagement, and for propensity score matching. Our findings suggest that the risk impacts of sustainability disclosure depend on the newness of information and disclosure channels.

We develop a model of interbank networks with random liquidity shocks. Networks of dilutable debt—e.g., long-term, unsecured—facilitate efficient liquidity transfers: Shocked banks pledge interbank claims as collateral for new senior debt, diluting existing debt. Unlike with non-dilutable debt, indebtedness and connectedness are sources of stability, not fragility. Dilution is thus a “backdoor bail-in” that reallocates losses absent a resolution authority, trigger security, or ex post renegotiation. We uncover a class of networks, “exponential networks,” that implement optimal contingent transfers via plain debt. Yet exponential networks are not pairwise stable, whereas some core–periphery networks are, rationalizing observed interbank structures and their under-insurance against crises.

Shariah-compliant equity screening provides a transparent setting in which institutional rules determine who may own a stock. A binary label identifies current eligibility but not whether the feasible investor base is fragmented across standards or close to changing. We define this instability as classification uncertainty and formalize its investor-base consequence through permitted investor mass. In a 1999-2024 CRSP-Compustat panel of 13,188 securities classified under seven researcher-emulated Shariah rulebooks, screening-rule disagreement and proximity to active boundaries rank next-month screen-implied transitions. U.S. Fama-MacBeth diagnostics do not support an unconditional equal-weighted permission premium, and a September 2023 DJIM/S&P methodology change produces no robust matched repricing. The central event evidence uses 25 official Securities Commission Malaysia lists. The 410 inclusions already trading before the preceding review have positive but imprecise matched returns. Applying the pre-event turnover floor yields 295 inclusions with 1.76 percentage points over $[0,10]$ trading days ($p_{\mathrm{date}}=0.008$; $p_{\mathrm{wild}}=0.017$) and 2.25 points over $[0,20]$ ($p_{\mathrm{date}}=0.018$; $p_{\mathrm{wild}}=0.035$). Leave-one-date-out, first-inclusion-only, and mid-review placebo checks are supportive, although a joint 20-day pre-event test rejects. Ownership and demand-pressure diagnostics do not identify a unique marginal buyer or clean causal demand shock. The evidence supports treating classification risk as a portfolio-monitoring state. Official Shariah permission is associated with price effects in a recognized local market among sufficiently tradable securities; formal eligibility alone is insufficient.

Signed graphs provide an effective architecture for portraying a system in which cooperation and conflict coexist. Emerging from the concept of balance in psychological sciences, they have found applications across several domains. Financial markets are one such example that can be modeled using signed networks, where assets exhibit correlations in price movements. During periods of systemic risk, such a signed financial network shows a loss of balance, which has been consistently demonstrated. Here, we explore how this structural imbalance is distributed across scales within the financial network, revealing its mesoscopic origin. Adopting the framework of structural balance theory, we use a measure of polarization based on triadic motifs to investigate the distribution of structural imbalance across varying sectoral scales. We analyze the temporal evolution of global polarization and its sectoral constituents using longitudinal data derived from the S&P 500 index. By decomposing global polarization into intra-sectoral and inter-sectoral constituents, we show that structural imbalance arises predominantly from interactions between sectors rather than within them during periods marked by systemic risk. We employ randomization protocols to confirm that observed imbalance configurations are statistically significant and not artifacts of lower-order interactions. We derive a regression equation demonstrating that the variance in global polarization is well explained by macroeconomic variables, indicating that low levels of global polarization during economic crises are driven by compounding pressures from supply chain disruptions and inflation uncertainty. Collectively, these findings provide a quantitative framework for understanding how localized sectoral conflicts propagate across the financial network and contribute to large-scale structural instability during periods of economic crisis.

Open-weight language models from Chinese AI labs caught up on benchmarks relative to proprietary frontier models in recent months. Yet their reliability on real-world financial tasks remains largely untested. We updated the Financial Touchstone benchmark, which now has 2,967 question context-answer triplets across 495 international annual reports. We also apply a new set of models on the benchmark, expanding coverage from eleven to twenty models across ten providers, including recent open-weight models such as GLM 4.7, GLM 5, Kimi K2.6, and DeepSeek V3.2, as well as Alibaba's proprietary flagship Qwen3-Max. Anthropic's Claude Opus 4.6 achieves the highest accuracy (88.4%), while Google's Gemini 2.5 Pro maintains the lowest hallucination rate (0.08%). Notably, the open-weight Kimi K2.6 ranks third in accuracy, and the non-reasoning models GLM 5 and Mistral 3 rank fourth and fifth, challenging the assumption that reasoning architectures or proprietary weights are a prerequisite for strong financial comprehension. Information retrieval remains the primary bottleneck, accounting for 48.9% of all failures. We also document a new finding: geopolitical content filters in Chinese models refuse legitimate financial questions (0.08% of attempts), sometimes without clear reason, and the refusal behavior depends on the access route as much as on the model. The complete dataset and evaluation framework are publicly available.

Business users confronted with a moving metric need to know which part of their data moved and why. Existing data-explanation methods typically return predicates: conjunctions of attribute-value conditions that isolate responsible records. Predicates are exact and directly executable as filters, but they describe axis-aligned regions and may not compactly capture segments defined by combinations of continuous tendencies. This paper presents Self-Explaining Segment Trees (SEST), an architecture in which an explanation is a multivariate cluster in a feature subspace selected for relevance to a designated key performance indicator (KPI). SEST selects the subspace once per KPI using Shapley attributions over a decision-tree surrogate, recursively partitions the population while choosing the branching factor independently at each node through mixture-model silhouette search, and attaches to every node a dual explanation payload: standardized effect sizes over numeric features and type-dependent contribution profiles over user-designated dimensions. These explanations are computed from untransformed data so surfaced values retain source units and category labels. A stance layer reduces any depth of the tree to its extremal KPI-suppressing and KPI-amplifying segments. We establish termination and a node-count bound determined by the depth limit and minimum segment size, and characterize per-tree construction cost as quadratic in population size in the degenerate case and geometrically decaying across depth in the balanced case. This is an architecture and methodology paper; we report no predictive-accuracy or validation results and leave outcome validation to future work.

Financial question answering is typically evaluated by answer correctness, yet in SEC filings a plausible and even numerically correct answer can be grounded in the wrong evidence. Similar facts and disclosures recur across sections of a filing, across reporting periods of the same firm, and across comparable firms. FinRank targets this provenance-sensitive retrieval problem by requiring systems to identify evidence for the intended entity, reporting period, and disclosure context. The benchmark contains 1185 manually authored question-answer records over the 10-K and 10-Q filings of 22 companies. Each record includes a reference answer, gold supporting passages, and hand-curated hard negatives drawn from confusable passages within filings, across reporting periods, and across comparable firms. FinRank evaluates passage retrieval, reranking, and hard-negative discrimination as separately measured tasks. Baseline results demonstrate the difficulty of this setting: among the evaluated systems, even a 7B instruction-tuned embedder reaches only 44.8% Recall@10 on the pooled evidence corpus; sub-billion-parameter encoders gain at most 3.5 points over BM25, a finance-adapted embedder trails BM25 by 9.7 points, and pairwise accuracy falls by 13.0-20.5 percentage points when random negatives are replaced with the curated hard negatives. FinRank provides an evidence-first benchmark for developing financial question answering systems that are not only accurate but also grounded in the correct disclosure.

Counterfactual analysis aims to predict potential outcomes under hypothetical scenarios, offering valuable insights for decision-making. This paper investigates the application of large language models (LLMs), specifically the GPT-3.5 model, for counterfactual analysis. We focus on the online lending context, where the counterfactual return on investment (ROI) is crucial for evaluating different interest rate schemes. We begin by assessing the predictive performance of GPT and comparing it with advanced machine learning algorithms. The results show that prompt engineering can significantly enhance GPT's predictions, with the R-squared increasing from 1.97% to 2.84%, closely approaching the 3.48% achieved by gradient-boosted regression. Subsequently, we utilize GPT to generate counterfactual ROIs under a set of alternative interest rates. GPT exhibits logical coherence and causal reasoning in its responses. The findings underscore the potential of LLMs as effective tools for counterfactual analysis in online lending, suggesting broader applications for LLMs in various predictive and decision-making contexts.

The impact of web datasets on market prices has suggested the development of new sources of information, such as social media and web portals, indicating the possibility of an emergent phenomenon. We propose a defining perspective, termed open information, that adds to the existing types of public and private information. We demonstrate their existence and justify material significance for pricing. In this respect, we present statistical hypotheses to test for a web dataset, GDELT, integrated for carbon pricing, that is represented by EU Allowance spot prices. Tests are designed with VAR and GARCH-X formulations, and return forecasting. The outcomes cannot rule out the material existence of open information. The result is significant for providing a conceptual basis to integrate a vast number of web datasets as alternative data in investment decisions.

The objective of this paper is to provide a methodology for applying the DeTEcT framework to modelling token economies, to formalise the configuration of the simulation environment, and to introduce an event analysis framework. A token economy is an economic system that has a unique mechanism for controlling its monetary supply, and a medium, in the form of a token or currency, for the valuation of goods and services, the settlement of transactions, and the storage of value. We show the key decisions that must be made when modelling an economy with the DeTEcT framework and showcase some numerical methods that can be used in conjunction with the framework to perform economic simulations. We also propose a framework for analysing and measuring the impacts of events on an economy, while also developing a procedure to measure the significance of these impacts. Throughout the paper, we use Bitcoin as a case study to demonstrate how to apply the frameworks and tools we proposed here. We show how a model of Bitcoin token economy can be set up, and how to measure the impacts of Bitcoin's endogenous policies (i.e., BIPs) on the wealth distribution of its economic participants.

We ask a representative sample to write prompts seeking spending and investing advice from LLMs, then simulate the lifetime effects of following the advice under realistic asset and labor market conditions. Applying this method to GPT-5.2, we find following the advice would move respondents toward life cycle theory: broader participation in diversified equity funds, age-declining equity shares, and larger savings buffers. Recommendations vary systematically by gender, prior AI experience, and financial literacy. For gender, two-thirds of recommended equity-share differences arise from men and women writing different prompts (demand), while one-third arise from gender labels attached to otherwise identical prompts (supply).