Daily Summary for 2026-11-05
daily
Transcript
Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.
Ruth: It's the fifth of November, twenty twenty-six, and this is the day's research.
Warren: 3 new papers came out today.
Ruth: I'm Ruth.
Warren: And I'm Warren, here with Iris.
Ruth: We'll take the day in one pass, then pull out the papers we're staying with.
Warren: Start us off with A Financial Brain Scan of the LLM.
The day's research: Ruth: We need better ways to interpret what large language models are actually thinking.
Warren: This is crucial for economic forecasting and trading applications where reasoning matters immensely.
Iris: The key work is opening and controlling these LLMs by inserting an interpretable sparse representation in their architecture.
Ruth: That technique is detailed in A Financial Brain Scan of the LLM.
Warren: That method maps forecasts to concepts like sentiment or technical analysis, allowing manual regulation without retraining.
Iris: Reducing positivity in off-the-shelf models actually improves portfolio performance with the highest Sharpe ratio.
Ruth: This steering capability uses a Sparse Auto-Encoder trained to be benign.
Warren: The sparse embeddings show sentiment and finance concepts are the most important features for financial applications.
Iris: That concept steering mechanism proves effective when applied to firm-level news, showing a small set of features explains most signal.
Ruth: This suggests off-the-shelf LLMs are often overly optimistic for trading.
Warren: Theoretically, we have foundational work on risk measures with dynamic monetary risk measures being the lower envelope of dynamic convex ones.
Iris: Normalized dynamic star-shaped risk measures similarly relate to lower envelopes of normalized dynamic convex ones.
Ruth: This representation theorem links different types of risk measures under suitable conditions.
Warren: Specific time consistent dynamic star-shaped risk measures induced by g-expectations can be normalized static star-shaped if g is star-shaped in z.
Iris: This provides a specific structure for these dynamic measures.
Ruth: Sublinear Markov Semigroups Beyond Uniqueness establishes rigorous mathematical foundations for sublinear valuation rules and uncertainty structures within a Markovian framework.
Warren: A Financial Brain Scan of the LLM shows how to control large language models by inserting an interpretable sparse representation to steer their reasoning toward specific concepts.
Iris: Dynamic star-shaped risk measures and g-expectations investigates how dynamic monetary risk measures can be represented as lower envelopes of dynamic convex risk measures.
Lucky paper: 2508.21285: Ruth: So we're looking at A Financial Brain Scan of the LLM today, and the authors are from MIT Sloan, NBER, and Yale SOM.
Warren: They’re trying to figure out how to control large language models by inserting an interpretable sparse representation into their architecture.
Iris: It shows that they can map model-generated economic forecasts to concepts like sentiment or technical analysis, and they can regulate the intensity of thinking about a specific concept without retraining the model.
Ruth: That steering capability is built on a Sparse Auto-Encoder, which is trained to be benign so it doesn't degrade the LLM's quality.
Warren: The resulting sparse embeddings show that sentiment and finance concepts are the most important features for financial applications when using this approach in A Financial Brain Scan of the LLM.
Iris: They found that reducing positivity in off-the-shelf models actually improves portfolio performance with the highest Sharpe ratio, even with a modest negative steering level.
Ruth: That's interesting because it suggests those off-the-shelf models are often overly optimistic for trading applications.
Warren: Theoretically, we also have foundational work on risk measures where dynamic monetary risk measures can be represented as the lower envelope of dynamic convex risk measures.
Iris: That representation theorem links different types of risk measures together under suitable conditions, which is key for how these dynamic measures behave.
Ruth: The paper also touches on sublinear Markov semigroups beyond uniqueness, establishing rigorous mathematical foundations for valuation rules and uncertainty structures in a Markovian framework.
Warren: A Financial Brain Scan of the LLM specifically introduces a method where you identify a relevant feature in the sparse representation to steer the LLM toward it.
Iris: When you activate that feature, you use the Sparse Auto-Encoder's decoder to find out what perturbation would have been added to the residual stream if that feature were more strongly activated.
Ruth: That perturbation is then added back into the residual stream to get a steered probability distribution over the next tokens, which gives fine-grained control.
Warren: This concept steering mechanism helps clarify short-horizon predictability by showing that a small set of sentiment-like features explains most of the signal, even when you reduce dimensionality.
Iris: The results indicate that under negative steering, positive classifications are associated with higher next day average returns, which is consistent with what the method was designed to do.
Ruth: So they validated steering as a tool to both identify and mitigate LLM bias by showing it can correct biases documented in other studies.
Warren: We saw that the baseline Sharpe ratio was three point eight seven, but negatively steered portfolios systematically achieved ratios of four point zero seven, four point one one, and four point two eight in the comparison figures shown in A Financial Brain Scan of the LLM.
Iris: And they also showed that steering can be used to deliberately amplify bias for research purposes, like simulating varying levels of risk aversion or home bias in human decision-making.
Ruth: That flexibility supports applications outside of finance, like political economy or psychology where controlled variation in dispositions is useful and retraining the base model would be costly.
Warren: They also looked at the clustering results from their analysis and found that the top three clusters—Sentiment, Finance/Markets, and Technical Analysis—captured fundamental concepts in finance.
Iris: However, they noted that temporal concepts were less expected, exhibiting a statistically significant Shapley Sharpe of zero point four one but one of the lowest individual Sharpe ratios at one point two seven.
Ruth: It seems the dominant driver for those temporal predictions is the directional content of news, whether it conveys positive or negative information about a company.
Warren: The steering technique also showed that under positive steering, negative classifications are associated with lower next day average returns, confirming the mechanism operates as intended.
Iris: In terms of what they found most important, they looked at the proximity of features and found that Sentiment, Finance/Markets, and Technical Analysis were in the top ten features by proximity.
Ruth: The paper concludes that this approach is a simple and scalable way to open and control large language models by inserting an interpretable sparse layer between hidden states and outputs.
Warren: While this technique is still in its infancy, concept steering could evolve into a standard tool for building interpretable, field-ready LLMs for research across all social sciences.
Lucky paper: 2305.02481: Ruth: So we’re looking at Dynamic star-shaped risk measures and g-expectations today.
Warren: This paper shows that dynamic monetary risk measures can be represented as the lower envelope of a family of dynamic convex risk measures.
Iris: And normalized dynamic star-shaped risk measures are the lower envelope of a family of normalized dynamic convex ones.
Ruth: That's a big structural link they are establishing between those two different types of risk measures.
Warren: They show that the translation between them only happens at a special position under some suitable conditions.
Iris: The representation for a dynamic monetary risk measure rho t X is given by the essential infimum over lambda in lambda of the essential supremum over Q in Qt of minus X Ft minus alpha t lambda Q.
Ruth: That formula looks complex, and it involves an integral over time and measures, which is quite technical.
Warren: It’s about finding that lower envelope using a family of convex functionals alpha t lambda defined over all probability measures equivalent to P on Ft.
Iris: The paper then looks at g-expectations and backward stochastic differential equations to define specific dynamic star-shaped risk measures induced by g expectations.
Ruth: That brings us to the point where they show that if the generator g is star-shaped in z, then rho g is a normalized static star-shaped risk measure.
Warren: That specific structure for these dynamic measures seems pretty attractive because it simplifies things significantly under those conditions.
Iris: They also analyze sensitivity and time consistency problems related to these measures, which are really important for real application.
Ruth: Sensitivity is defined by checking if for any t between zero and T, rho t satisfies an essential infimum of Y in A rho t over Qtilde is greater than minus infinity for some Qtilde equivalent to P on Ft.
Warren: And corollary three point two says that if a dynamic monetary risk measure has sensitivity, then there exists a family of sensitive dynamic convex or coherent measures such that rho t X equals the essential infimum over lambda in lambda of rho t lambda X for all X in L infinity T(P).
Iris: They also give us some concrete examples. Example four point one defines a generator g z as z over four I s less than or equal to one plus z squared I s greater than one, and that results in a normalized time consistent dynamic star-shaped risk measure on L infinity T(P).
Ruth: That example shows how the structure actually manifests when you plug in specific generators for the generator g.
Warren: Example four point three shows that if g z is just kappa z where kappa is positive and z is a real number, then the solution to backward stochastic differential equation sixteen results in a normalized time consistent dynamic convex risk measure on L infinity T(P).
Iris: It seems like the main conclusion of this paper boils down to those two representation theorems for the dynamic risk measures.
Ruth: So they’re essentially providing a rigorous way to build and characterize these complex dynamic risk measures using lower envelopes of simpler convex ones.
Warren: They also mention that there are still many interesting problems left to explore, like finding the representation for the time consistent dynamic risk measures or fully characterizing those induced by g expectations.
Iris: That sounds like they’ve mapped out a lot of territory but acknowledged there’s still deep work ahead in fully understanding every aspect of these star-shaped measures.
More episodes
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- The day in conversation for 2026-11-18
- Daily Summary for 2026-11-18
- 2601.05274-On the use of case estimate and transactional payment data in neural networks for individual loss reserving
- 2009.00868-Loss-Given-Default Modeling by Post-Last Passage Time Process
- 2606.30070-Financial Resilience Evaluation: From Conditional Expectations to Dynamic Convex Risk Measures
- 2212.12687-Measuring price impact and information content of trades in a time-varying setting
- Daily Summary for 2026-11-17
- 2401.08323-Dynamic portfolio selection under generalized disappointment aversion