A Financial Brain Scan of the LLM
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.
Ruth: We're looking at the paper: "A Financial Brain Scan of the LLM".
Warren: Here's the gist of the paper. Emerging techniques in computer science make it possible to “brain scan” large language models (LLMs), identify the plain-English concepts that guide their reasoning, and steer them while holding other factors constant This paper shows that this approach can map LLM-generated economic forecasts to concepts such as sentiment, technical analysis, and timing, and compute their relative importance without reducing performance.
Ruth: Pruning redundant reasoning tokens The method allows researchers to open and control an LLM by inserting an interpretable sparse representation within its architecture.
Warren: Through this representation, a researcher can identify the concepts the model uses to process a given input, expressed in plain English.
Ruth: The array of concepts is extremely large and ranges from optimism or financial risk to fondness for specific cuisines.
Title and authors: Ruth: The paper's title, "A Financial Brain Scan of the LLM," immediately tells us that this is about trying to scan or look into what's happening inside these models concerning finance.
Warren: It suggests they are moving beyond just looking at the final answer and are digging into the reasoning process itself to find those underlying financial concepts.
Iris: They are essentially trying to open up these black boxes by inserting an interpretable sparse representation right in the architecture of the model so researchers can see what's going on.
Ruth: That’s right, they introduce this sparse embedding which encodes all the information flowing through the model and then decodes it back into a representation that is much more readable.
Warren: It sounds like they are essentially giving us a way to open and control an LLM by inserting this layer, which is what page two of the paper talks about when they mention applying techniques from Cunningham et al. in this context.
Iris: They're focusing on the fact that the model's inner representation is dense, meaning it’s hard to interpret, so they are trying to make it sparse and meaningful for economic tasks like these ones (<ref:2508.21285#pg2>).
Ruth: And page one sets up a big problem by pointing out that the scale and density of LLMs make them uninterpretable black boxes for economics and finance, which is why they felt this approach was necessary (<ref:2508.21285#pg1>).
Warren: So, the core idea is using this sparse representation to map forecasts onto concepts like sentiment or timing, and then figuring out which concept actually drives the prediction most.
Iris: And the paper claims they can do that while maintaining performance, which is a big claim because usually when you add interpretability, you expect a hit to accuracy (<ref:2508.21285#pg1>).
Ruth: It seems they are showing that this approach isn't just for curiosity; it’s designed to be transparent and replicable for empirical research in the social sciences (<ref:2508.21285#pg1>).
The paper's summary: Ruth: To summarize what they did, this paper introduces a method where you insert an interpretable sparse representation into an already trained LLM's architecture to allow you to identify the plain-English concepts guiding their reasoning.
Warren: They build this by training a sparse autoencoder which is essentially an encoder-decoder structure that takes all the information flowing through the model and encodes it into this representation before decoding it back.
Iris: This autoencoder is trained specifically on the residual streams produced at a fixed layer in the language model, and it’s loss function encourages both accurate reconstruction and sparsity in that hidden representation (<ref:2508.21285#pg3>).
Ruth: So, the key takeaway here is that this sparse autoencoder approach provides both interpretability and performance by mapping any prompt through an LLM to a semantically rich sparse representation (<ref:2508.21285#pg4>).
Warren: And when they look at what's most important, they found that the most relevant feature groups for financial applications are sentiment, closely followed by market and finance concepts and technical analysis ideas (<ref:2508.21285#pg4>).
Iris: They also show how you can steer the model by identifying a relevant feature in this sparse representation, figuring out what perturbation would happen if that feature were more strongly activated, and adding that to the residual stream to get a steered probability distribution (<ref:2508.21285#pg3>).
Ruth: And they show this steering can be used practically to simulate agents with specific preferences without needing to retrain the base model by just repeating prompts, which is really useful for modeling political economy or psychology (<ref:2508.21285#pg4>).
The paper's improvements: Ruth: The paper points out that this sparse autoencoder method substantially outperforms some recent methods presented in other work, specifically Chen et al. in two thousand twenty-two (ref:Chen et al <ref:2508.21285#pg1>., two thousand twenty-four) <ref:2508.21285#pg1>.
Warren: And the results show that when you use this sparse autoencoder approach, any prompt passed through an LLM can be mapped to a semantically rich sparse representation that you can then apply in the framework from Gentzkow et al. in two thousand nineteen (ref:Gentzkow et al <ref:2508.21285#pg1>., two thousand nineteen) <ref:2508.21285#pg1>.
Iris: The paper highlights that they found that off-the-shelf LLMs are often overly optimistic for trading applications, and reducing positivity actually improves portfolio performance when you steer the model slightly negatively (ref:Figure six).
Ruth: That is a really practical improvement because it gives researchers a way to mitigate documented biases, like those seen in Fedyk et al. in two thousand twenty-four or even deliberately amplify bias for simulation purposes (Figure six).
Warren: And they also found that the dominant driver of short-horizon predictability isn't just any feature, but a small set of sentiment-like features that explains most of the signal even when you aggressively reduce dimensionality (ref:Table III).
Iris: They also show how steering can be used to simulate agents with configurable preferences across political economy and psychology where retraining is costly, which expands the applicability way beyond just finance (ref:Figure six).
Conclusion: Ruth: So, to wrap up, this paper on "A Financial Brain Scan of the LLM" presents a simple, scalable way to open and control large language models by inserting this sparse layer between the model's hidden states and its outputs.
Warren: The main points are that concept steering works when applied to firm-level news, showing that off-the-shelf LLMs lean toward optimism for trading tasks.
Iris: It also clarifies the drivers of short-horizon predictability by showing that a small set of sentiment-like features explains most of the signal even with aggressive dimensionality reduction (<ref:2508.21285#pg4>).
Ruth: This framework extends beyond finance to help us simulate agents with specific dispositions, preferences, and contextual sensitivities where retraining is too expensive for those kinds of controlled variations (ref:Figure six).
Warren: It’s a simple tool that allows you to correct biases by selecting an appropriate steering level for a given feature, or even deliberately amplify bias for research simulations.
Iris: While this technique is still in its early stages, it could evolve into a standard tool for building interpretable, field-ready LLMs across all the social sciences (<ref:2508.21285#pg4>).
MIT Sloan and NBER · University of Melbourne · Yale School of Management and Organization (SOM) · ESSEC Business School
q-fin.GN, cs.AI, cs.CE, econ.GN, q-fin.EC
Submitted: 2025-08-29
Updated: 2026-11-04
Comments: 82 pages
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 88/100
The gist: Emerging techniques in computer science make it possible to “brain scan” large language models (LLMs), identify the plain-English concepts that guide their reasoning…
Key concepts
- Sparse Auto-Encoder (SAE)
- A Sparse Auto-Encoder is a neural network trained to compress and then reconstruct information in a very sparse way. It learns an interpretable representation of the input data by forcing the hidden layer to be sparse, meaning only a few features are activated at any given time. This creates a structured way to look inside the LLM's internal workings.
- Concept Steering
- Concept steering is a control mechanism that allows researchers to influence an LLM's output by identifying and adjusting specific concepts within its sparse representation. By perturbing the residual stream based on a targeted feature, users can manually regulate how strongly the model 'thinks' about that concept, enabling fine-grained control over its reasoning.
- Sentiment Steering
- Sentiment steering uses a steered feature linked to risk aversion or attention to wealth to simulate agents with configurable preferences. By adjusting this steering coefficient, researchers can observe monotonic shifts in behavior, such as making the model less optimistic for trading applications and improving portfolio performance.
Terminology
Summary
Emerging techniques in computer science make it possible to “brain scan” large language models (LLMs), identify the plain-English concepts that guide their reasoning, and steer them while holding other factors constant This paper shows that this approach can map LLM-generated economic forecasts to concepts such as sentiment, technical analysis, and timing, and compute their relative importance without reducing performance. p. 0 A Financial Brain Scan of the LLM
Hui Chen∗
, Antoine Didisheim†
, Mohammad (Mo) Pourmohammadi‡
,
Luciano Somoza§
, Hanqing Tian¶
February, 2026
Abstract
Emerging techniques in computer science make it possible to “brain scan” large language models (LLMs), identify the plain-English concepts that guide their reasoning,
and steer them while holding other factors constant. We show that this approach can
map LLM-generated economic forecasts to concepts such as sentiment, technical analysis, and timing, and compute their relative importance without reducing performance.Where the paper says it
Pruning redundant reasoning tokens
The method allows researchers to open and control an LLM by inserting an interpretable sparse representation within its architecture. Through this representation, a researcher can identify the concepts the model uses to process a given input, expressed in plain English. p. 1 The
technique allows to open and control an LLM by inserting an interpretable sparse representation within its architecture. Through this representation, a researcher can identify the
concepts the model uses to process a given input, expressed in plain English. Through this representation, a researcher can identify the
concepts the model uses to process a given input, expressed in plain English. The array of
concepts is extremely large and ranges from optimism or financial risk to fondness for specific cuisines. The array of
concepts is extremely large and ranges from optimism or financial risk to fondness for specific cuisines. Furthermore, by adjusting the code, we can manually regulate the intensity at
which the model “thinks” about any targeted concept while keeping everything else equal.Where the paper says it
Where the paper says it
Where the paper says it
How the training signal is built
The technique introduces a sparse embedding within an already trained LLM. All information flowing through the model is encoded into this representation and subsequently decoded. p. 2 The technique used in this paper
solves this problem by introducing a sparse embedding within an already trained LLM. All
information flowing through the model is encoded into this representation and subsequently
decoded. All
information flowing through the model is encoded into this representation and subsequently
decoded. This encoder-decoder structure, referred to as a Sparse Auto-Encoder (SAE),
is trained to be benign and not impact the LLMs’ quality. Second, the residual streams are
not sparse, and they occupy a high dimensional, entangled latent space in which individual
coordinates lack clear semantic interpretation. In the human brain, different regions often
specialize for distinct tasks such as language, vision, or motor control, making it easier to
associate structure with function. The decoder then attempts to reconstruct the original input from z:
xˆ = g(Wdz + bd), (4)
where Wd ∈ R
d×k and bd ∈ R
d are decoder parameters, and g(·) is usually chosen to be
the identity function or a sigmoid, depending on the domain of the input data (Goodfellow,
Bengio, Courville, and Bengio, 2016). The encoder and decoder are trained jointly to minimize the reconstruction error between
the input x and its reconstruction xˆ:
LAE(x) = ∥x − xˆ∥
2
2
| {z }
Reconstruction loss
.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the gains come from
The sparse autoencoder approach provides both interpretability and performance. Using sparse autoencoders, any prompt passed through an LLM can be mapped to a semantically rich sparse representation, which can then be applied in step (i) of the Gentzkow et al. p. 3 The methodology proposed in this paper provides both interpretability and performance. Using sparse auto-encoders, any prompt passed through an LLM can be mapped to a semantically rich sparse representation, which can then be applied in step (i) of the Gentzkow
et al. Using sparse auto-encoders, any prompt passed through an LLM can be mapped to a semantically rich sparse representation, which can then be applied in step (i) of the Gentzkow
et al. (2019) framework. They
use the last layer of LLMs as embeddings and show that this representation outperforms all
previous textual methods in the finance literature. To compare our reduced sparse embeddings with the last-layer embeddings from the same LLM, we follow Chen et al. Comparing these models with the full feature model, which jointly incorporates
all 17 groups, allows us to quantify the contribution of individual financial concepts to LLM
performance in portfolio applications. Unsurprisingly, the most important feature group is Sentiment, closely followed by Market/Finance and Technical Analysis concepts.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Concept Steering
Steering can be understood as follows: 1. Identify a relevant feature in the sparse representation to steer the LLM toward. p. 6 For this
paper’s interpretability goal, representation learning with autoencoders can summarize risk
exposures while remaining economically transparent. This approach applies that idea by replacing dense, opaque LLM embeddings with a sparse, interpretable feature set, then using
those features to steer the model’s output. To address this, we seek to discover a sparse and semantically meaningful representation
of the residual stream. Specifically, we aim to identify a transformation:
r
(l)
7→ z
(l) ∈ R
k
where each coordinate of z
(l)
corresponds to an interpretable feature or concept. With m total blocks,
the final residual stream r
(m)
is passed to an output module that maps it to a probability
distribution over the vocabulary. This distribution represents the model’s prediction for the
next word:
P(Next Word | Input Text) = Block(Output)(r
(m)
). We show that this method can make LLMs
transparent for economic analysis, allowing researchers to identify the concepts driving predictions without reducing performance. Second, we find that these interpretable features
enable precise control over model behavior (such as sentiment or risk aversion), providing a
practical approach to correct biases, simulate preferences, and design controlled experiments.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Applied Steering and Bias Correction
Steering features linked as risk aversion and attention to wealth can be used to simulate agents with configurable preferences without retraining the base model. Repeated prompting yields monotonic shifts in behavior consistent with the intended steering direction, providing a practical mechanism to simulate agents with configurable preferences without retraining the base model. p. 6 Repeated
prompting yields monotonic shifts in behavior consistent with the intended steering direction:
stronger risk aversion decreases equity allocations and increases the share of safe choices,
whereas stronger attention to wealth produces the opposite effect. This provides a practical
mechanism to simulate agents with configurable preferences without retraining the base
model. and international
markets, reporting sizable gains over bag-of-words and other NLP baselines. Building on
that work, this paper retains predictive power but replaces dense embeddings with a sparse,
interpretable feature set, showing that most of the signal loads on refined sentiment. Second, we find that these interpretable features
enable precise control over model behavior (such as sentiment or risk aversion), providing a
practical approach to correct biases, simulate preferences, and design controlled experiments. Together, these findings offer a lightweight, replicable method for turning LLMs from opaque
systems into useful tools for empirical research in economics and the social sciences.Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
This paper presents a simple, scalable approach to opening and controlling large language models by inserting a sparse, interpretable layer between the model’s hidden states and its outputs. Concept steering performs as intended when applied to firm-level news, showing that off-the-shelf LLMs are overly optimistic for trading applications. p. 5 Steering and Bias Correction
Having shown how sparse autoencoders increase the transparency of LLMs, we next
highlight the second main advantage: they allow us to force the model to incorporate a
specific concept, with a chosen intensity, when processing an input. We start by building a long-short trading strategy based on prompt-extracted sentiment
from news in the spirit of Lopez-Lira and Tang (2023a) and Chen et al. and international
markets, reporting sizable gains over bag-of-words and other NLP baselines. Building on
that work, this paper retains predictive power but replaces dense embeddings with a sparse,
interpretable feature set, showing that most of the signal loads on refined sentiment. Repeated
prompting yields monotonic shifts in behavior consistent with the intended steering direction:
stronger risk aversion decreases equity allocations and increases the share of safe choices,
whereas stronger attention to wealth produces the opposite effect. This provides a practical
mechanism to simulate agents with configurable preferences without retraining the base
model. Furthermore, by adjusting the code, we can manually regulate the intensity at
which the model “thinks” about any targeted concept while keeping everything else equal. The capability of observing an LLM’s “mental” process and manually steering it towards
any direction are absent in current financial research and can be valuable for applications
across all social sciences.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
References
Aguilera, Ana M Manuel Escabias, and Mariano J Valderrama, 2006 Using principal components for estimating logistic regression with high-dimensional multicollinear data, Computational Statistics & Data Analysis 50, 1905–1924.
Aher Gati Rosa I Arriaga, and Adam Tauman Kalai, 2023 Using large language models to simulate multiple humans and replicate human subject studies, in Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research (PMLR). p. 7 Horton (2023) argues that LLMs can serve as “homo silicus” in
economic simulations. Aher, Arriaga, and Kalai (2023) formalize “Turing Experiments” and
show that recent models replicate classic findings in behavioral economics and psychology.Where the paper says it
Argyle Lisa P Ethan C Busby Nancy Fulda Joshua R Gubler Christopher Rytting, and David Wingate, 2023 Out of one, many: Using language models to simulate human samples, Political Analysis 31, 337–351. p. 7 Aher, Arriaga, and Kalai (2023) formalize “Turing Experiments” and
show that recent models replicate classic findings in behavioral economics and psychology. Argyle, Busby, Fulda, Gubler, Rytting, and Wingate (2023) demonstrate that conditioning
LLMs on demographics can reproduce group-level survey response distributions.Where the paper says it
Bybee Leland Bryan Kelly Asaf Manela, and Dacheng Xiu, 2024 Business news and business cycles, The Journal of Finance 79, 3105–3147. p. 6 The
design also complements prompt-based methods that classify news directly (Lopez-Lira and
Tang, 2023b), and it follows best-practice evaluation to avoid training-cutoff and look-ahead
errors (Sarkar and Vafa, 2024). It also connects to Bybee, Kelly, Manela, and Xiu (2020) by
6Where the paper says it
Bybee Leland Bryan T Kelly Asaf Manela, and Dacheng Xiu, 2020 The structure of economic news, SSRN Working Paper No. p. 6 The
design also complements prompt-based methods that classify news directly (Lopez-Lira and
Tang, 2023b), and it follows best-practice evaluation to avoid training-cutoff and look-ahead
errors (Sarkar and Vafa, 2024). It also connects to Bybee, Kelly, Manela, and Xiu (2020) by
6Where the paper says it
Chen Hui Antoine Didisheim, and Luciano Somoza, 2024 Out of the (black) box: Ai as conditional probability, Available at SSRN. p. 0 A Financial Brain Scan of the LLM
Hui Chen∗
, Antoine Didisheim†
, Mohammad (Mo) Pourmohammadi‡
,
Luciano Somoza§
, Hanqing Tian¶
February, 2026
Abstract
Emerging techniques in computer science make it possible to “brain scan” large language models (LLMs), identify the plain-English concepts that guide their reasoning,
and steer them while holding other factors constant. We show that this approach can
map LLM-generated economic forecasts to concepts such as sentiment, technical analysis, and timing, and compute their relative importance without reducing performance.Where the paper says it
Chen Yifei Bryan T Kelly, and Dacheng Xiu, 2022 Expected returns and large language models, SSRN Working Paper No. p. 6 The most related study is Chen et al. (2022), who use contextualized news embeddings
from large language models to forecast next-day stock returns in U.Where the paper says it
Cheng Qiang Pengkai Lin, and Yue Zhao, 2024 Does generative ai facilitate investor trading? p. 1 Introduction
Research on artificial intelligence is growing rapidly. Large language models (LLMs) are
already part of financial analysis, research workflows, and trading strategies (Cheng, Lin,
and Zhao, 2024). First, LLM forecasts are biased toward positive sentiment. While this finding may not generalize to all models, it is consistent with evidence
5We follow the approach proposed by Jobson and Korkie (1981), incorporating the correction noted in
Memmel (2003), to test the significance of our results.Where the paper says it
Where the paper says it
Cunningham Hoagy Aidan Ewart Logan Riggs Robert Huben, and Lee Sharkey, 2023 Sparse autoencoders find highly interpretable features in language models, arXiv preprint arXiv:2309. p. 1 , Cunningham,
Ewart, Riggs, Huben, and Sharkey, 2023; Gao, 2024; Shi, 2025) to economic tasks. The
technique allows to open and control an LLM by inserting an interpretable sparse representation within its architecture.Where the paper says it
Didisheim Antoine Martina Fraschini, and Luciano Somoza, 2025a Ai’s predictable memory and its implications for finance, SSRN Working Paper No. p. 0 A Financial Brain Scan of the LLM
Hui Chen∗
, Antoine Didisheim†
, Mohammad (Mo) Pourmohammadi‡
,
Luciano Somoza§
, Hanqing Tian¶
February, 2026
Abstract
Emerging techniques in computer science make it possible to “brain scan” large language models (LLMs), identify the plain-English concepts that guide their reasoning,
and steer them while holding other factors constant. We show that this approach can
map LLM-generated economic forecasts to concepts such as sentiment, technical analysis, and timing, and compute their relative importance without reducing performance.Where the paper says it
Didisheim Antoine Martina Fraschini Luciano Somoza, and Hanqing Tian, 2025b Is AI reasoning useful in finance? p. 0 A Financial Brain Scan of the LLM
Hui Chen∗
, Antoine Didisheim†
, Mohammad (Mo) Pourmohammadi‡
,
Luciano Somoza§
, Hanqing Tian¶
February, 2026
Abstract
Emerging techniques in computer science make it possible to “brain scan” large language models (LLMs), identify the plain-English concepts that guide their reasoning,
and steer them while holding other factors constant. We show that this approach can
map LLM-generated economic forecasts to concepts such as sentiment, technical analysis, and timing, and compute their relative importance without reducing performance. Thus, it is essential to develop and apply methods that increase transparency and align
model behavior with the researchers’ objectives. This paper applies an emerging technique in computer science (see, e.Where the paper says it
Where the paper says it
Didisheim Antoine Shikun Barry Ke Bryan T Kelly, and Semyon Malamud, 2024 Apt or “aipt”? p. 4 We repeat the prediction exercise—training logistic regressions on a rolling window and building a long-short strategy from
the forecasts—using subsets of the most important features. Consistent with the notion of
“Virtue of Complexity” Kelly, Malamud, and Zhou (2024), adding features always increases
the Sharpe ratio. A Financial Brain Scan of the LLM
Hui Chen∗
, Antoine Didisheim†
, Mohammad (Mo) Pourmohammadi‡
,
Luciano Somoza§
, Hanqing Tian¶
February, 2026
Abstract
Emerging techniques in computer science make it possible to “brain scan” large language models (LLMs), identify the plain-English concepts that guide their reasoning,
and steer them while holding other factors constant. We show that this approach can
map LLM-generated economic forecasts to concepts such as sentiment, technical analysis, and timing, and compute their relative importance without reducing performance.Where the paper says it
Where the paper says it
et al., Vaswani, 2017 Attention is all you need, Advances in neural information processing systems 30. p. 7 g Vaswani
et al. , 2017; Radford et al.Where the paper says it
Fatahi Somayeh Julita Vassileva, and Chanchal K Roy, 2024 Comparing emotions in chatgpt answers and human answers to the coding questions on stack overflow, Frontiers in Artificial Intelligence 7, 1393903. p. 6 in the literature (Fatahi, Vassileva, and Roy, 2024). Second, steering can be used to correct
this bias.Where the paper says it
Fedyk Anastassia Ali Kakhbod Peiyao Li, and Ulrike Malmendier, 2024 Chatgpt and perception biases in investments: An experimental study, Available at SSRN 4787249. p. 1 First, LLMs’ scale and density make them uninterpretable black boxes, limiting their usefulness for research (Ludwig, Mullainathan, and
Rambachan, 2025). Second, LLMs contain hard-to-identify biases, for instance tilting output toward specific demographic preferences (Fedyk, Kakhbod, Li, and Malmendier, 2024).Where the paper says it
Gabaix Xavier Ralph SJ Koijen Robert Richmond, and Motohiro Yogo, 2024 Asset embeddings, Available at SSRN 4507511.
Gao Leo et al. p. 1 This paper applies an emerging technique in computer science (see, e. , Cunningham,
Ewart, Riggs, Huben, and Sharkey, 2023; Gao, 2024; Shi, 2025) to economic tasks. tian1@unimelb. au
1
arXiv:2508.Where the paper says it
Where the paper says it
Gemini, 2025 Gemini: A family of highly capable multimodal models. p. 8 (2)
This architecture, shared by many state-of-the-art models, has demonstrated remarkable
performance across a wide range of NLP benchmarks (e. , OpenAI, 2024; Llama, 2024;
Gemini, 2025), yet it remains notoriously difficult to interpret.Where the paper says it
Gentzkow Matthew Bryan Kelly, and Matt Taddy, 2019 Text as data, Journal of Economic Literature 57, 535–574. p. 3 steering can detect and correct biases in canonical finance applications and how this technique
can be applied in broader settings to tune LLM’s preferences and biases. Interpretable LLMs’ embeddings
In their seminal review, Gentzkow, Kelly, and Taddy (2019) observe that a large fraction
of empirical applications of textual analysis in economics and finance can be summarized
in three broad steps: (i) constructing a numerical representation of the text; (ii) mapping
these representations to predicted values of an unknown outcome; and (iii) employing the
predicted values in subsequent descriptive or causal analyses.Where the paper says it
Gu Shihao Bryan T Kelly, and Dacheng Xiu, 2020 Empirical asset pricing via machine learning, Review of Financial Studies 33, 2223–2273. p. 6 Related Literature Modern asset-pricing research shows that flexible, high-dimensional
representations can improve prediction and interpretation. (2020) find that machine learning methods extract return-relevant structure from large predictor sets.Where the paper says it
Gu Shihao Bryan T Kelly, and Dacheng Xiu, 2021 Autoencoder asset pricing models, Journal of Econometrics 222, 429–450. p. 9 , Gu, Kelly, and Xiu, 2021). Formally, an autoencoder consists of two components: an encoder and a decoder.Where the paper says it
Horton John J. p. 7 Furthermore, our paper speaks to the fast-growing line of work that treats LLMs as
stand-ins for human subjects. Horton (2023) argues that LLMs can serve as “homo silicus” in
economic simulations. Horton (2023) argues that LLMs can serve as “homo silicus” in
economic simulations. Aher, Arriaga, and Kalai (2023) formalize “Turing Experiments” and
show that recent models replicate classic findings in behavioral economics and psychology. The method is
transparent, lightweight, and replicable for empirical research in the social sciences. ∗MIT Sloan and NBER.Where the paper says it
Where the paper says it
Where the paper says it
Jobson J Dave and Bob M Korkie, 1981 Performance hypothesis testing with the sharpe and treynor measures, Journal of Finance 889–908. p. 5 First, LLM forecasts are biased toward positive sentiment. While this finding may not generalize to all models, it is consistent with evidence
5We follow the approach proposed by Jobson and Korkie (1981), incorporating the correction noted in
Memmel (2003), to test the significance of our results.Where the paper says it
Kelly Bryan Semyon Malamud, and Kangying Zhou, 2024 The virtue of complexity in return prediction, The Journal of Finance 79, 459–503. p. 4 We repeat the prediction exercise—training logistic regressions on a rolling window and building a long-short strategy from
the forecasts—using subsets of the most important features. Consistent with the notion of
“Virtue of Complexity” Kelly, Malamud, and Zhou (2024), adding features always increases
the Sharpe ratio.Where the paper says it
Lieberum Tom et al. p. 4 Each feature in a sparse representation corresponds to a specific semantic
meaning. In this paper, we use the pre-trained and labeled sparse model released by Google
DeepMind (Lieberum et al. tian1@unimelb. au
1
arXiv:2508.Where the paper says it
Where the paper says it
Liu Xiao Zirui Wu Xueqing Wu Pan Lu Kai-Wei Chang, and Yansong Feng, 2024 Are llms capable of data-based statistical and causal reasoning? benchmarking advanced quantitative reasoning with data, arXiv preprint arXiv:2402. p. 0 tian1@unimelb. au
1
arXiv:2508.Where the paper says it
Llama, 2024 The llama 3 herd of models. p. 8 (2)
This architecture, shared by many state-of-the-art models, has demonstrated remarkable
performance across a wide range of NLP benchmarks (e. , OpenAI, 2024; Llama, 2024;
Gemini, 2025), yet it remains notoriously difficult to interpret.Where the paper says it
Lopez-Lira Alejandro, and Yuehua Tang, 2023a Can chatgpt forecast stock price movements? p. 2 We divide our analysis in two equally important parts. First, we focus on interpretability
and use the sparse embedding as input to a forecasting model, to understand what generates
the LLM-driven excess returns documented in recent papers (Chen, Kelly, and Xiu, 2022;
Lopez-Lira and Tang, 2023a; Chen, Didisheim, and Somoza, 2024). tian1@unimelb. au
1
arXiv:2508. tian1@unimelb. au
1
arXiv:2508.Where the paper says it
Where the paper says it
Where the paper says it
Lopez-Lira Alejandro, and Yuehua Tang, 2023b Can chatgpt forecast stock price movements? p. 6 Building on
that work, this paper retains predictive power but replaces dense embeddings with a sparse,
interpretable feature set, showing that most of the signal loads on refined sentiment. The
design also complements prompt-based methods that classify news directly (Lopez-Lira and
Tang, 2023b), and it follows best-practice evaluation to avoid training-cutoff and look-ahead
errors (Sarkar and Vafa, 2024).Where the paper says it
Loughran Tim Bill McDonald, 2011 When is a liability not a liability? p. 3 The first step (numerical representation) entails a trade-off between interpretability and
efficiency. Simpler methods, such as dictionary-based approaches and bag-of-words models,
provide highly transparent and interpretable representations (Tetlock, 2007; Loughran and
McDonald, 2011). steering can detect and correct biases in canonical finance applications and how this technique
can be applied in broader settings to tune LLM’s preferences and biases. Interpretable LLMs’ embeddings
In their seminal review, Gentzkow, Kelly, and Taddy (2019) observe that a large fraction
of empirical applications of textual analysis in economics and finance can be summarized
in three broad steps: (i) constructing a numerical representation of the text; (ii) mapping
these representations to predicted values of an unknown outcome; and (iii) employing the
predicted values in subsequent descriptive or causal analyses.Where the paper says it
Where the paper says it
Ludwig Jens Sendhil Mullainathan and Ashesh Rambachan, 2025 Large language models: An applied econometric framework, Technical report, National Bureau of Economic Research. p. 1 However, there are two main concerns
for researchers in economics and finance. First, LLMs’ scale and density make them uninterpretable black boxes, limiting their usefulness for research (Ludwig, Mullainathan, and
Rambachan, 2025).Where the paper says it
McQueen James B, 1967 Some methods of classification and analysis of multivariate observations, in Proc. of 5th Berkeley Symposium on Math. Stat. and Prob., 281–297.
Memmel Christoph, 2003 Performance hypothesis testing with the sharpe ratio, Available at SSRN 412588. p. 5 First, LLM forecasts are biased toward positive sentiment. While this finding may not generalize to all models, it is consistent with evidence
5We follow the approach proposed by Jobson and Korkie (1981), incorporating the correction noted in
Memmel (2003), to test the significance of our results.Where the paper says it
OpenAI 2024 Gpt-4 technical report. p. 8 (2)
This architecture, shared by many state-of-the-art models, has demonstrated remarkable
performance across a wide range of NLP benchmarks (e. , OpenAI, 2024; Llama, 2024;
Gemini, 2025), yet it remains notoriously difficult to interpret.Where the paper says it
Park Joon Sung Joseph C O’Brien Carrie J Cai Meredith Ringel Morris Percy Liang, and Michael S Bernstein, 2023 Generative agents: Interactive simulacra of human behavior, in Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST ’23) (ACM). p. 7 Argyle, Busby, Fulda, Gubler, Rytting, and Wingate (2023) demonstrate that conditioning
LLMs on demographics can reproduce group-level survey response distributions. In HCI,
Park, O’Brien, Cai, Morris, Liang, and Bernstein (2023) build multi-agent, memory-based
“generative agents” that exhibit believable individual and social behavior.Where the paper says it
Radford Alec et al. p. 7 g Vaswani
et al. , 2017; Radford et al. Sparse Autoencoders (SAEs)
Autoencoders are a class of neural network models designed for unsupervised dimensionality reduction. Like Principal Component Analysis (PCA), they aim to compress input
data into a lower-dimensional latent representation which minimizes the reconstruction error.Where the paper says it
Where the paper says it
Rousseeuw Peter J, 1987 Silhouettes: a graphical aid to the interpretation and validation of cluster analysis, Journal of computational and applied mathematics 20, 53–65. p. 4 4 The methodology produces 17 clusters with distinct
economic interpretation. This separation enables the construction of two model types and their corresponding
long–short portfolios: (i) models trained exclusively on features from a single group, and (ii)
models trained on all but one group, following the Shapley value framework (Gu, Kelly, and
2For both models, we use logistic regression, trained on a three-year rolling window with an additional
one-year validation set and re-estimated annually.Where the paper says it
Sarkar Suproteem K, 2025 Economic representations. p. 3 (2022) demonstrate that complex, opaque embeddings generated by modern machine learning methods (see, e. , Sarkar, 2025) can significantly outperform simpler
representations in important financial prediction tasks.Where the paper says it
Sarkar Suproteem K. p. 3 (2022) demonstrate that complex, opaque embeddings generated by modern machine learning methods (see, e. , Sarkar, 2025) can significantly outperform simpler
representations in important financial prediction tasks. Building on
that work, this paper retains predictive power but replaces dense embeddings with a sparse,
interpretable feature set, showing that most of the signal loads on refined sentiment. The
design also complements prompt-based methods that classify news directly (Lopez-Lira and
Tang, 2023b), and it follows best-practice evaluation to avoid training-cutoff and look-ahead
errors (Sarkar and Vafa, 2024).Where the paper says it
Where the paper says it
Shi Wei et al. p. 1 This paper applies an emerging technique in computer science (see, e. , Cunningham,
Ewart, Riggs, Huben, and Sharkey, 2023; Gao, 2024; Shi, 2025) to economic tasks. (2)
This architecture, shared by many state-of-the-art models, has demonstrated remarkable
performance across a wide range of NLP benchmarks (e. , OpenAI, 2024; Llama, 2024;
Gemini, 2025), yet it remains notoriously difficult to interpret.Where the paper says it
Where the paper says it
Tetlock Paul C, 2007 Giving content to investor sentiment: The role of media in the stock market, The Journal of finance 62, 1139–1168. p. 5 Comparing these models with the full feature model, which jointly incorporates
all 17 groups, allows us to quantify the contribution of individual financial concepts to LLM
performance in portfolio applications. Unsurprisingly, the most important feature group is Sentiment, closely followed by Market/Finance and Technical Analysis concepts.Where the paper says it
Vaswani Ashish et al. p. 7 This layered design lets the model gradually learn more
complex language patterns as information moves up through the layers. g Vaswani
et al. g Vaswani
et al. , 2017; Radford et al.Where the paper says it
Where the paper says it
Wang Liang Nan Yang Xiaolong Huang Linjun Yang Rangan Majumder, and Furu Wei, 2023 Improving text embeddings with large language models, arXiv preprint arXiv:41>.
Appendix A transparently reports the choices made to merge the 25 clusters into 17 groups and provides the final naming of each group
The choice of 25 clusters is guided by the maximization of the silhouette coefficient (Rousseeuw, 1987), consistent with standard practice in the literature. The results yielded the following classification: • 0, 3, 11, 22 → Punctuation and Symbols • 1, 18 → Healthcare • 2 → Sentiment • 4 → Finance/Markets • 6 → Technology • 7 → Temporal Concepts • 8 → Scientific Jargon • 9 → Finance/Corporate • 12 → Legal • 5, 13 → Technical Analysis ●. p. 5 Comparing these models with the full feature model, which jointly incorporates
all 17 groups, allows us to quantify the contribution of individual financial concepts to LLM
performance in portfolio applications. Unsurprisingly, the most important feature group is Sentiment, closely followed by Market/Finance and Technical Analysis concepts.Where the paper says it
Appendix B Neural Network Predictions
Table VI reports the return predictions and number of features. The Shapley Sharpe captures the incremental predictive power unique to a given cluster, in the sense that the logistic predictor cannot replicate the same information from correlated features in other clusters. p. 4 4 The methodology produces 17 clusters with distinct
economic interpretation. This separation enables the construction of two model types and their corresponding
long–short portfolios: (i) models trained exclusively on features from a single group, and (ii)
models trained on all but one group, following the Shapley value framework (Gu, Kelly, and
2For both models, we use logistic regression, trained on a three-year rolling window with an additional
one-year validation set and re-estimated annually. Comparing these models with the full feature model, which jointly incorporates
all 17 groups, allows us to quantify the contribution of individual financial concepts to LLM
performance in portfolio applications. Unsurprisingly, the most important feature group is Sentiment, closely followed by Market/Finance and Technical Analysis concepts. We show that this approach can
map LLM-generated economic forecasts to concepts such as sentiment, technical analysis, and timing, and compute their relative importance without reducing performance. We also show that models can be steered to be more or less risk-averse, optimistic,
or pessimistic, which allows researchers to correct or simulate biases. We show that this approach can
map LLM-generated economic forecasts to concepts such as sentiment, technical analysis, and timing, and compute their relative importance without reducing performance. We also show that models can be steered to be more or less risk-averse, optimistic,
or pessimistic, which allows researchers to correct or simulate biases. In the spirit
of Lopez-Lira and Tang (2023a), we use these forecasts to construct intra-day long–short
portfolios, going long (equally weighted) on stocks with positive sentiment and short on
those with negative sentiment. Interestingly, negatively steered predictions achieve a statistically significant higher annualized Sharpe ratio (4. In the spirit
of Lopez-Lira and Tang (2023a), we use these forecasts to construct intra-day long–short
portfolios, going long (equally weighted) on stocks with positive sentiment and short on
those with negative sentiment. Interestingly, negatively steered predictions achieve a statistically significant higher annualized Sharpe ratio (4. Interestingly, time-related features (e. ,
dates, years, timelines) exhibit the fourth-highest Shapley value but the lowest individual
Sharpe ratios, i. Interestingly, time-related features (e. ,
dates, years, timelines) exhibit the fourth-highest Shapley value but the lowest individual
Sharpe ratios, i. In the spirit
of Lopez-Lira and Tang (2023a), we use these forecasts to construct intra-day long–short
portfolios, going long (equally weighted) on stocks with positive sentiment and short on
those with negative sentiment. Interestingly, negatively steered predictions achieve a statistically significant higher annualized Sharpe ratio (4. In the spirit
of Lopez-Lira and Tang (2023a), we use these forecasts to construct intra-day long–short
portfolios, going long (equally weighted) on stocks with positive sentiment and short on
those with negative sentiment. Interestingly, negatively steered predictions achieve a statistically significant higher annualized Sharpe ratio (4. Comparing these models with the full feature model, which jointly incorporates
all 17 groups, allows us to quantify the contribution of individual financial concepts to LLM
performance in portfolio applications. Unsurprisingly, the most important feature group is Sentiment, closely followed by Market/Finance and Technical Analysis concepts. Comparing these models with the full feature model, which jointly incorporates
all 17 groups, allows us to quantify the contribution of individual financial concepts to LLM
performance in portfolio applications. Unsurprisingly, the most important feature group is Sentiment, closely followed by Market/Finance and Technical Analysis concepts. Features with large absolute loadings strongly
affect predictions, while those near zero have little impact. We repeat the prediction exercise—training logistic regressions on a rolling window and building a long-short strategy from
the forecasts—using subsets of the most important features. Features with large absolute loadings strongly
affect predictions, while those near zero have little impact. We repeat the prediction exercise—training logistic regressions on a rolling window and building a long-short strategy from
the forecasts—using subsets of the most important features. Interestingly, time-related features (e. ,
dates, years, timelines) exhibit the fourth-highest Shapley value but the lowest individual
Sharpe ratios, i. Interestingly, time-related features (e. ,
dates, years, timelines) exhibit the fourth-highest Shapley value but the lowest individual
Sharpe ratios, i. Interestingly, time-related features (e. ,
dates, years, timelines) exhibit the fourth-highest Shapley value but the lowest individual
Sharpe ratios, i. In lay terms, just like neuroscientists record which brain regions
“light up” when a person sees a face or hears music, here we look at which features in the
model activate when it processes certain words or contexts. In both cases, the system is not
labeled in advance but the researcher discovers meaning by observing consistent patterns of
activation.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Table III Return Predictions and Topics
The top three clusters—Sentiment, Finance/Markets, and Technical Analysis—capture fundamental concepts in finance and portfolio construction as reflected in news. The next most important cluster, however, is less expected: temporal concepts. This cluster exhibits a statistically significant Shapley Sharpe of 0.41, yet one of the lowest individual Sharpe ratios (1.27). The dominant driver of such predictions is the directional content of news—whether it conveys positive or negative information about a company. Clusters such as Sentiment, Finance/Markets, and Technical Analysis directly capture this dimension. The secondary determinant is the timeframe—whether the news is likely to affect markets in the short or long run. Timing features thus add value only in conjunction with sentiment-related features, which is precisely what Table III reflects.
Table IV Validating the Steering Technique
The proportion of positive classifications increases monotonically with the steering coefficient. When the model’s positivity node is activated, it is more likely to predict a positive market reaction. p. 5 On the sparse representation, we select a feature linked to “positivity’. For
each news item, we obtain a baseline LLM forecast and additional forecasts corresponding
to varying levels of positive steering. As expected, the proportion of positive classifications increases monotonically with stronger positive steering and the conditional returns
are consistent with the model’s induced interpretation of the same headline. In the spirit
of Lopez-Lira and Tang (2023a), we use these forecasts to construct intra-day long–short
portfolios, going long (equally weighted) on stocks with positive sentiment and short on
those with negative sentiment. Repeated
prompting yields monotonic shifts in behavior consistent with the intended steering direction:
stronger risk aversion decreases equity allocations and increases the share of safe choices,
whereas stronger attention to wealth produces the opposite effect. This provides a practical
mechanism to simulate agents with configurable preferences without retraining the base
model.Where the paper says it
Where the paper says it
Where the paper says it
Figure 6 Comparison of model performance metrics
The baseline Sharpe ratio equals 3.87, while negatively steered portfolios systematically achieve higher ratios (4. p. 5 In the spirit
of Lopez-Lira and Tang (2023a), we use these forecasts to construct intra-day long–short
portfolios, going long (equally weighted) on stocks with positive sentiment and short on
those with negative sentiment. Interestingly, negatively steered predictions achieve a statistically significant higher annualized Sharpe ratio (4. Second, steering can be used to correct
this bias. In principle, this bias-correction framework can be applied to any LLM bias or preference
that can be mapped onto the set of features in the sparse representation, either to attenuate
or to amplify such tendencies. steering can detect and correct biases in canonical finance applications and how this technique
can be applied in broader settings to tune LLM’s preferences and biases. Interpretable LLMs’ embeddings
In their seminal review, Gentzkow, Kelly, and Taddy (2019) observe that a large fraction
of empirical applications of textual analysis in economics and finance can be summarized
in three broad steps: (i) constructing a numerical representation of the text; (ii) mapping
these representations to predicted values of an unknown outcome; and (iii) employing the
predicted values in subsequent descriptive or causal analyses. In principle, this bias-correction framework can be applied to any LLM bias or preference
that can be mapped onto the set of features in the sparse representation, either to attenuate
or to amplify such tendencies. We illustrate this by steering features labeled as risk aversion
and attention to wealth, and subsequently prompting the model to choose between safe and
risky investments or to allocate a budget between Treasuries and an equity index.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Figure 7 Top 10 features by proximity - Topics 0–5
The choice of 25 clusters is guided by the maximization of the silhouette coefficient (Rousseeuw, 1987), consistent with standard practice in the literature. The results yielded the following classification: • 0, 3, 11, 22 → Punctuation and Symbols • 1, 18 → Healthcare • 2 → Sentiment • 4 → Finance/Markets • 6 → Technology. p. 7 , positivity or risk aversion-while holding other
latent factors fixed, enabling clean comparative statics in finance tasks and portable designs
for the social sciences. Methodology
The technology underlying generative models like chat-GPT and similar LLMs, is based
on generative architectures that model the probability distribution of text sequences.Where the paper says it
Figure 8 Top 10 features by proximity - Topics 6–11
The choice of 25 clusters is guided by the maximization of the silhouette coefficient (Rousseeuw, 1987), consistent with standard practice in the literature. The results yielded the following classification: • 6 → Technology • 7 → Temporal Concepts • 8 → Scientific Jargon • 9 → Finance/Corporate • 12 → Legal. p. 7 , positivity or risk aversion-while holding other
latent factors fixed, enabling clean comparative statics in finance tasks and portable designs
for the social sciences. Methodology
The technology underlying generative models like chat-GPT and similar LLMs, is based
on generative architectures that model the probability distribution of text sequences.Where the paper says it
Figure 9 Top 10 features by proximity - Topics 12–17
The choice of 25 clusters is guided by the maximization of the silhouette coefficient (Rousseeuw, 1987), consistent with standard practice in the literature. The results yielded the following classification: • 12 → Legal • 13 → Technical Analysis • 14 → Governance and Politics • 15 → Energy and Manufacturing. p. 0 A Financial Brain Scan of the LLM
Hui Chen∗
, Antoine Didisheim†
, Mohammad (Mo) Pourmohammadi‡
,
Luciano Somoza§
, Hanqing Tian¶
February, 2026
Abstract
Emerging techniques in computer science make it possible to “brain scan” large language models (LLMs), identify the plain-English concepts that guide their reasoning,
and steer them while holding other factors constant. We show that this approach can
map LLM-generated economic forecasts to concepts such as sentiment, technical analysis, and timing, and compute their relative importance without reducing performance.Where the paper says it
Figure 10 Top 10 features by proximity - Topics 18–23
The choice of 25 clusters is guided by the maximization of the silhouette coefficient (Rous…
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Sparse Autoencoder based feature extraction can map LLM-generated economic forecasts to concepts such as sentiment, technical analysis, and timing without reducing performance. This allows for identifying
the plain-English concepts that guide their reasoning
in LLMs. -
Steering mechanisms allow researchers to manually regulate the intensity at which the model
thinks
about any targeted concept while keeping everything else equal. This enables models to be steered to be more or less risk-averse, optimistic, or pessimistic, which allows researchers to correct or simulate biases. -
The method provides a
fine-grained, concept-level control mechanism grounded in interpretable internal features.
This control can be used tosimulate agents with configurable preferences
across diverse contexts like political economy and psychology.
Abstract
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 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.
Sources
- Sparse Autoencoders Find Highly Interpretable Features in Language Models
- Scaling and evaluating sparse autoencoders
- Assessing Look-Ahead Bias in Stock Return Predictions Generated By GPT Sentiment Analysis
- Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2
- Are LLMs Capable of Data-based Statistical and Causal Reasoning? Benchmarking Advanced Quantitative Reasoning with Data
- Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models
- Route Sparse Autoencoder to Interpret Large Language Models
- Improving Text Embeddings with Large Language Models
Related papers
- Amortized Neural Filtering for Calibrated Posteriors in Statistical Arbitrage
- Multi-Jurisdictional Legal Identity Assurance for Capability Gating: A Design-Science Proposal for Tiered, Reusable Identity Assurance of Natural, Juridical, and Machine Entities
- Decomposable Reward Modeling and Realistic Environment Design for Reinforcement Learning-Based Forex Trading
- Stablecoin Integration and Stress Transmission Through Crypto-Asset Service Providers: Evidence from Austria
- Tempting the Agent: The Economics of Reputation without Persistent Identity in AI Agent Markets
- Tokenize Everything, But Can You Sell It? RWA Liquidity Challenges and the Road Ahead