Research

A Financial Brain Scan of the LLM, with H. Chen, A. Didisheim, Mo Pourmohammadi, and Hanqing Tian

We use sparse autoencoders (SAEs) to “brain scan” large language models, mapping internal activations to interpretable concepts and enabling targeted “steering” of specific features. We show that prompting models to adopt specific preferences activates many unintended concepts, complicating causal inference in agent simulations. In contrast, SAEs permit smooth, more isolated steering of target concepts. Furthermore, our scans reveal a stark disconnect between the model’s internal feature contributions and its self-generated explanations, showing that these explanations are unreliable guides to its reasoning. Finally, SAEs provide a lightweight intervention for existing models, producing interpretable embeddings and allowing researchers to attenuate specific economic biases

Presentations:  AFA 2027 (upcoming), EFA 2026, Toronto Asset Pricingm Investments Workshop 2026, Stanford Institute for Theoretical Economics (SITE) 2026, Oxford University and the University of Technology Sydney.

Out of the (Black)Box: AI as Conditional Probability (2026), with H. Chen and A. Didisheim

We explore the economic significance and interpretability of the distribution of conditional probabilities behind LLM's text generation. Using a dataset of news and returns, we find that conditional probabilities are interpretable and correlate with model accuracy. Conversely, measures of declared confidence used in the literature are opaque, structurally biased, unstable, and more model-dependent, indicating that LLMs cannot assess their own confidence. Using conditional probabilities, we analyze LLM biases and provide insights into the internal mechanisms driving model decisions. Our results indicate that conditional probabilities provide a reliable and transparent reflection of LLM beliefs, particularly for economic 

Presentations:  AFA 2026, CEIBS Inaugural CAC Academic Conference 2026, WFA 2025, NTU AI for Finance Summer School, AI & Big Data in Finance Research Forum, FIRN-UQ Asset Management Meeting, INSEAD Finance Symposium, Barcuh College, University of Bristol, University of Manchester, University of Warwick, and University of Oxford 

AI in Finance and Information Overload (2025), with A.Balogh, A. Didisheim and Hanqing Tian

Artificial intelligence is reshaping financial markets, yet the limits to its rationality remain underexplored. This paper documents information overload in Large Language Models applied to financial analysis. Using earnings forecasts from corporate calls and market reaction predictions from news, we show that predictive accuracy follows an inverted U-shaped pattern, where excessive context degrades performance. Larger LLMs mitigate this effect, increasing the optimal context length. Our findings underscore a fundamental limitation of AI-driven finance: more data is not always better, necessitating empirical tuning to determine the right amount of context for each task. 

AI’s predictable memory in financial analysis (2025), with A. Didisheim and M. Fraschini

Look-ahead bias in Large Language Models (LLMs) arises when information that would not have been available at the time of prediction is included in the training data and inflates prediction performance. This paper proposes a practical methodology to quantify look-ahead bias in financial applications. By prompting LLMs to retrieve historical stock returns without context, we construct a proxy to estimate memorization-driven predictability. We show that the bias varies predictably with data frequency, model size, and aggregation level: smaller models and finer data granularity exhibit negligible bias. Our results help researchers navigate the trade-off between statistical power and bias in LLMs.

AI’s predictable memory in financial analysis (Economics Letters, 2025), with A. Didisheim and M. Fraschini 

Look-ahead bias in Large Language Models (LLMs) arises when information that would not have been available at the time of prediction is included in the training data and inflates prediction performance. This paper proposes a practical methodology to quantify look-ahead bias in financial applications. By prompting LLMs to retrieve historical stock returns without context, we construct a proxy to estimate memorization-driven predictability. We show that the bias varies predictably with data frequency, model size, and aggregation level: smaller models and finer data granularity exhibit negligible bias. Our results help researchers navigate the trade-off between statistical power and bias in LLMs.

Is AI reasoning useful in finance? (Finance Research Letter 2026), with A. Didisheim, M. Fraschini and Hanqing Tian

Whether Large Language Models (LLMs) will result in a marginal productivity increase or a technological revolution largely depends on their ability to reason. LLMs with reasoning capabilities outperform vanilla ones on math and coding. However, it remains unclear whether such emergent abilities translate into improved economic insights. We evaluate state-of-the-art general-purpose reasoning-enhanced LLMs by OpenAI and DeepSeek on standard financial tasks: news sentiment and earnings direction prediction. Reasoning-enhanced models fail to demonstrate a significant advantage, while model size does. These findings indicate that improved reasoning does not necessarily translate into enhanced economic intuition, questioning their cost-effectiveness and practical utility in finance. Only finance-specific reasoning models yield a relatively modest increase in performance.

CBDC and Banks: Threat or Opportunity? (2025), with M. Fraschini

We study how banks react to the introduction of a Central Bank Digital Currency (CBDC) when households have heterogeneous preferences. We find that banks increase their deposit interest rates in response to a CBDC, even when the CBDC pays no interest rate. However, when the central bank provides funding to offset the loss in deposits, banks optimally push households towards the CBDC by reducing deposit interest rates. This allows them to liquidate reserves, reduce their cost of funding, and increase their profits. We calibrate the model to provide quantitative estimates of these mechanisms.

Presentations:  CEPR Fintech and Digital Currencies RPN Workshop 2024, BS Gillmore Centre Conference on DeFi & Digital Currencies 2024, 4th Sailing the Macro Workshop 2024, EEA 2024, ArmEA 2024, AFA 2024, Bank of England, Swiss Finance Institute Research Days 2021.

The Monetary Entanglement between CBDC and Central Bank Policies (2023),  with M. Fraschini and T. Terracciano

Using a banking model, we show that the effects of introducing a Central Bank Digital Currency (CBDC) depend on the ongoing monetary policy. We derive the conditions for a neutral introduction without central bank pass-through funding and find that they do not always hold with quantitative easing, as bank lending shrinks if demand for CBDC is above a certain threshold. Moreover, we find that commercial banks optimally liquidate their excess reserves to accommodate households’ demand for CBDC. Consequently, households will replace banks on the liability side of the central bank balance sheet, making quantitative tightening difficult to implement.

This paper previously circulated as "Central Bank Digital Currency and Quantitative Easing"

Honors: Finalist for the 2022 ECB Young Economist Prize

Presentations: EEA 2023, Bank of England, ECB Forum on Central Banking (Poster Session, June 2022), The Future.s of Money - Paris (June 2022), 26th Spring Meeting of Young Economists (SMYE, May 2022), ASSA 2022 Virtual Annual Meeting (AEA Poster Session, January 2022), Day-Ahead Workshop on Financial Regulation at the University of Zurich (October 2021), Swiss Finance Institute Research Days (June 2021), 14th Financial Risks International Forum (March 2021), Finance BB seminar at the University of Geneva (January 2021)

Media: Financial Times (see Outreach)

The End of the Crypto-Diversification Myth (2022), with A.Didisheim and M. Fraschini

Cryptocurrencies and equities have exhibited a high and positive correlation since March 2020. Without obvious fundamental drivers, we theoretically show that trading flows by retail investors can drive this correlation. Using a unique dataset of investor-level holdings from a bank offering trading accounts and cryptocurrency wallets, we show that retail investors tend to trade equities and cryptocurrencies simultaneously in the same direction. This behavior became prominent in March 2020. We provide suggestive evidence showing that stocks preferred by crypto-traders exhibit a stronger correlation with Bitcoin, especially when the cross-asset retail volume is high.

Presentations: AFA 2024, MFA Annual meeting 2023, ToDeFi 2023 (Best PhD paper award), 5th UWA Blockchain and Cryptocurrency conference, New Zealand Finance Meeting 2022, CB&DC Job Market Candidates Workshop, NYU Stern (PhD brownbag), Swiss Finance Institute, HEC Lausanne.

Media: Financial Times, VoxEU (see Outreach)