Pith. sign in

REVIEW 5 major objections 8 minor 14 references

InvestAlign: Overcoming Data Scarcity in Aligning Large Language Models with Investor Decision-Making Processes under Herd Behavior

T0 review · 5 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper argues that synthetic data from a solvable investment model can replace scarce real-user data when aligning language models to human investors.

desk verdict A practical empirical alignment result that leans on a shaky theoretical convergence proof; the 44–61% MSE reductions are real, but Eq. 28 needs either a proper proof or a demotion to observation. read the letter →

arxiv 2507.06528 v1 pith:V6MY3UX7 submitted 2025-07-09 cs.CL cs.AIcs.ETcs.LG

Huisheng Wang , Zhuoshi Pan , Hangjing Zhang , Mingxiao Liu , Hanqing Gao , H. Vicky Zhao This is my paper
classification cs.CLcs.AIcs.ETcs.LG
keywords LLMalignmentsupervisedfine-tuningherdbehavioroptimalinvestmentbehavioralfinancesynthetictrainingdatascarcityconvergencerate
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes that the shortage of real-user data for aligning language models with investor behavior can be sidestepped by generating supervised fine-tuning (SFT) data from the theoretical solution of a simpler, similar optimal-investment problem. It studies a two-agent portfolio-choice setting with herd behavior, in which one agent's decisions influence another's; for the simplest case (absolute herd behavior, unilateral influence) a closed-form optimal decision rule exists. The authors argue that fine-tuning on this theoretically generated data converges faster than fine-tuning on collected user decisions, and that the resulting InvestAgent tracks real-user decisions more closely than the base model, both in the simple problem and in two harder variants where the theoretical solution is computationally heavy. If correct, this offers a data-efficient route to behavioral alignment in finance without collecting large amounts of personal investment data.

What carries the argument

The engine is P3's closed-form optimal decision rule, Eq. (11), which expresses A1's optimal investment $\hat{P}_1(t)$ as a function of risk-aversion coefficients $\alpha_1, \alpha_2$, influence coefficient $\theta_1$, market parameters $r, v, \sigma$, and a numerically computed constant $\eta$. The paper uses this rule to generate $10^3$ training samples over varied investment attributes. The convergence argument compares cross-entropy gradient norms for theoretical versus real-user data; assuming real-user decisions equal the theoretical solution plus i.i.d. uniform noise, the theoretical-data loss has a larger gradient norm, which in a locally convex loss landscape implies faster parameter convergence.

What would settle it

Using the paper's collected P3 data, compute the residuals $\tilde{P}_1^i(t) - \hat{P}_1^i(t)$ for each participant and time step; if they are visibly non-uniform, serially correlated, or shifted away from zero, the assumed noise model in Section 4.2 is violated, and one can directly test whether the theoretical-data gradient norm still exceeds the real-data gradient norm when the empirical residual distribution is used instead of $U(-\varepsilon, \varepsilon)$.

Watch

Extended reading notes

Core claim

The central claim is that supervised fine-tuning on SFT datasets built from the closed-form solution of the simple problem P3 (absolute herd behavior with unilateral influence) produces InvestAgents whose investment decisions align with real-user decisions better than models before fine-tuning, in P3 and also in the complex problems P1 (relative herd behavior, unilateral influence) and P2 (absolute herd behavior, mutual influence). Across the tested models, the paper reports overall mean squared error reductions of 45.59–61.26% for P3, 44.52–56.69% for P1, and 45.77–60.92% for P2 relative to the unfine-tuned baselines. The faster-convergence claim is stated as the gradient-norm inequality $\|\nabla \hat{L}(w)\| > \|\nabla \tilde{L}(w)\|$ in Eq. (28), and is supported by measured gradient norms and training-loss curves.

Load-bearing premise

The proof assumes that real people's investment decisions are the mathematically optimal solution plus independent random noise spread evenly within fixed bounds, and that the range of possible human choices fits inside the range the model can produce; if human decisions contain systematic biases that the optimal-investment-plus-herding model does not capture, the faster-convergence inequality and the alignment benefit do not follow.

Editorial extensions

If this is right

  • If correct, LLM-based investment agents can be aligned to herd-affected human decision-making without collecting large proprietary datasets of individual investment choices.
  • The same recipe—solve a simple solvable analog, generate SFT samples, fine-tune—could in principle be applied to other decision problems in behavioral finance where stylized closed-form solutions exist.
  • Theoretical-data fine-tuning should be preferred to real-user data when the theoretical model is a good proxy, because it converges faster; adding a smaller amount of real-user data can further reduce error.
  • The trained InvestAgents reproduce two economic hypotheses: investment decisions converge as herd influence increases, and stronger mutual influence lowers expected total terminal wealth.
  • Fine-tuning on a general financial-domain dataset performs worse on this task than fine-tuning on the theoretically generated herd-behavior dataset.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • (Editorial inference) The faster-convergence proof relies on the idealization that human noise is uniform and independent; a natural next test is whether inequality (28) still holds when residuals are skewed, heteroskedastic, or autocorrelated.
  • (Editorial inference) The method's usefulness will be governed by how far a solvable toy problem can deviate from the target problem before the learned behavior stops transferring; comparing InvestAlign against a model fine-tuned on a size-matched random sample of real-user data would sharpen this.
  • (Editorial inference) Extending the approach to other behavioral biases such as loss aversion or overconfidence would require closed-form solutions of the same style for those biases, or a hybrid dataset blending theory with a thin layer of real data.
  • (Editorial inference) Because the paper measures alignment as closeness to the average human decision, a complementary test would be whether InvestAgent reproduces the full distribution of human choices, including the spread and individual heterogeneity, not just the mean.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 8 minor

Summary. This paper proposes InvestAlign, a framework for data-scarce supervised fine-tuning (SFT) of LLMs in behavioral finance. Given a complex optimal investment problem, InvestAlign identifies a similar and simple problem (P3: absolute herd behavior with unilateral influence) whose closed-form theoretical solution (Eq. (11), from the authors' prior work) is used to generate synthetic SFT datasets, and fine-tunes base LLMs (GPT-3.5, Qwen-2, Llama-3.1, GLM-4) into 'InvestAgents.' The paper reports that (i) the theoretical P3 solution is statistically consistent with decisions collected from 119 human participants (Section 3.3, Appendix A.6); (ii) InvestAgents reduce overall MSE relative to pre-SFT LLMs by 44–61% on P3 and on the complex problems P1 (relative herd, unilateral influence) and P2 (absolute herd, mutual influence) (Table 1); and (iii) fine-tuning on theoretical data yields faster parameter convergence than fine-tuning on real-user data, supported by a gradient-norm inequality (Eq. (28)) and by experiments in Figures 3–4. The paper also includes ablations, FinGPT baselines, mixed-data robustness checks, and validation of economic hypotheses H1/H2.

Significance. The empirical alignment results, if they hold up, are valuable: they offer a concrete, reproducible route to LLM alignment in a domain where real-user data are costly and privacy-sensitive, and they demonstrate genuine transfer from a solvable proxy problem to harder problems. The paper ships public code, reports consistent MSE reductions across three base models and three problems, and includes several substantiating analyses (hyperparameter ablations in A.9, FinGPT baselines in A.11, mixed-data robustness in A.10, and economic-principle checks in A.12) that go beyond a single headline number. The theoretical faster-convergence claim is the paper's weakest component: the proof of Eqs. (27)–(28) is a one-line assertion that is demonstrably invalid, and the noise model on which it rests is unvalidated. The observed gradient-norm difference in Figures 3–4 may still hold empirically in the tested settings, but it does not establish the advertised theorem.

major comments (5)
  1. [§4.2 / Appendix A.8, Eqs. (27)–(28)] The claimed proof of the faster-convergence result is invalid. The paper asserts that 'Because f_P1(t)(·) is monotonically decreasing, from (7) and (21), we can prove that Î < Ĩ < 1', but monotonicity of the model density does not determine the comparison of the two overlap integrals Î = ∫_{P̂} f̂ f_model dx and Ĩ = ∫_{P̃} f̃ f_model dx. A discrete counterexample that satisfies all of the paper's assumptions (f̂ decreasing, f̃ = f̂ convolved with symmetric uniform noise, f_model decreasing, supports nested with P̂ ⊂ P̃) is: f̂ on {1,2} with probabilities (0.8, 0.2); noise n ~ U(−1,1) so f̃ on {0,1,2,3} has probabilities (4/15, 1/3, 1/3, 1/15); f_model on {0,1,2,3} = (0.3, 0.3, 0.2, 0.2). Then Î = 0.8·0.3 + 0.2·0.2 = 0.28 > Ĩ = 4/15·0.3 + 1/3·0.3 + 1/3·0.2 + 1/15·0.2 = 0.26, which by the paper's own Eqs. (22)–(23) gives ||∇Lhat(w)|| < ||∇Ltilde(w)||, the reverse of Eq. (28). The faster-convergence theorem is therefore not established even under the paper's stated assumptions; the empirical evidence in Figures 3–4 is a separate matter.
  2. [§4.2] The convergence analysis rests on assumptions that are neither validated against the collected data nor stated with their evidentiary status: (i) real-user decisions equal the theoretical optimal solution plus i.i.d. U(−ε, ε) noise, with ε never estimated from the 119 participants' data; (ii) the support nesting P̂1(t) ⊂ P̃1(t) ⊂ P1(t); (iii) monotonicity of the LLM output density f_P1(t)(·); and (iv) the constancy of η when α1 and θ1 vary slightly (Appendix A.7). Under assumption (i), the comparison in Eq. (28) is between a clean signal and the same signal corrupted by noise, so the conclusion is substantially entailed by the modelling choice rather than by a property of investor behavior. The statistical validation in Section 3.3 and Appendix A.6 (difference t = −1.075; correlation t = −0.843 against 0.85) tests only whether the theoretical solution is close to real-user data in mean and correlation; it does not test the uniform-noise, independence, support-nesting, or monotonicity assumptions. The paper should either validate these assumptions empirically or re-frame the convergence claim as an empirical observation.
  3. [§4.2, Figures 3–4] The empirical comparison of gradient norms is confounded by a dataset-size and coverage mismatch. The theoretical SFT dataset contains 10³ samples (10 α1 values × 10 θ1 values × 10 trials) with uniform attribute coverage, whereas the real-user SFT dataset contains at most 119 samples concentrated in a few attribute classes (Section 3.2). The paper's second assumption ('sample size sufficiently large') is therefore false for the real-user dataset. The observed higher gradient norm on theoretical data could be due to the larger, more uniformly covered dataset rather than to the absence of label noise. A matched-sample comparison (e.g., subsampling the theoretical dataset to match the real-user class distribution and size, or upsampling the real-user data) is needed to support the convergence-rate claim.
  4. [§5.2, Table 1 (P2 rows)] The reported overall MSE for P2 is computed over the union P2 ∪ P3, not over the P2 data alone. Because P3 is exactly the problem on which InvestAgents were fine-tuned, including P3 data in the P2 metric can inflate the apparent cross-task improvement, and the statement that InvestAlign 'reduce[s] the overall MSEs by ... in P2' is not supported by a P2-only measurement. The authors should report the MSE restricted to the 44 P2 participants (or to the attribute classes actually covered by them) and clarify why the union was used.
  5. [§5.1, Table 1 and Abstract] The headline claim of 'significantly closer alignment' is supported only by point estimates of overall MSE reductions (45.59%–61.26%), with no confidence intervals or significance tests. Since the real-user samples are small (119, 80, and 44 participants for P3, P1, and P2) and the LLM responses are averaged over 10 seeds, bootstrap confidence intervals on the MSE reductions or a paired test over attribute classes should be reported to justify the word 'significantly'.
minor comments (8)
  1. [Appendix A.6] The null hypothesis 'mean correlation coefficient does not differ from 0.85' (t-statistic −0.843) uses a threshold (0.85) whose justification is not given; reporting the estimated mean and confidence interval of {ρi} would be more informative.
  2. [§5.2, Table 1] The text says GPT-3.5 was also evaluated on P2, but Table 1 contains no GPT-3.5 row for P2; the discrepancy should be resolved.
  3. [§5.1, Eqs. (9)–(10)] The overall MSE averages over attribute classes without weighting by the number of participants in each class, so classes with a single participant count as much as classes with many; a participant-weighted average or a per-class coverage report would be more robust.
  4. [§3.1] The symbol T denotes both the time horizon and the time set T = [0, T], and in Eqs. (6) and (26) P1(t) is used for both the decision variable and its support range; distinct notation would remove ambiguity.
  5. [Appendix A.4] GLM-4 results are shown in Figure 6 but no GLM-4 entries appear in Table 1; adding them would let readers compare all four models on the same metric.
  6. [§4.2, Eq. (21)] The derivation of Eq. (21) identifies f_P1(t)(x) with Sigmoid(z) without defining the mapping from the decision value x to the sigmoid input z; this step needs an explicit statement.
  7. [Appendix A.5] The section heading 'Comparison of Real-User Data and Pre-SFT LLMs' Investment Decision on P1 and P1' appears to contain a typo; it should likely read 'P1 and P2'.
  8. [§4.2, citation (Chen and Xiang, 2018)] The claim that 'a larger gradient norm leads to faster convergence' cites a survey of unstable gradients; this is a heuristic rather than a convergence-rate theorem, and even if Eq. (28) were valid, additional assumptions on the loss landscape would be needed to conclude a faster convergence rate.

Circularity Check

1 steps flagged · score 5.0 of 10

The advertised faster-convergence theorem is tautological: real-user data are defined as the theoretical solution plus noise, so the proof compares the theory with a noisy copy of itself; the empirical alignment results remain independently tested.

  1. self definitional [Section 4.2, Eqs. (26)-(28)]
    "Because the real-user data often have a bigger noise than the theoretical solution, we assume that P̃1(t) is P̂1(t) plus a white noise n(t) that independently and identically satisfies a uniform distribution U(−ε, ε), i.e, P̃1(t) = P̂1(t) + n(t). ... Because the investment decisions of the pre-SFT LLM can be arbitrary due to the randomness of model parameters, we have P̂1(t) ⊂ P̃1(t) ⊂ P1(t). Because fP1(t)(·) is monotonically decreasing, from (7) and (21), we can prove that Î < Ĩ < 1, and thus we have ∥∇L̂(w)∥ > ∥∇L̃(w)∥."

    The faster-convergence result compares the theoretical density f̂ with a target f̃ that is, by the paper's own construction, the convolution of f̂ with U(−ε,ε). 'Real-user data' is thereby defined as a noisy copy of the very theoretical solution whose superiority is being proved. The theorem therefore reduces to the tautology that a clean signal has a larger gradient norm than its own smoothed version; no property of the actually collected real-user distribution enters the proof. The conclusion is entailed by the definition of the comparator rather than by the data. The asserted inequality Î<Ĩ is also not implied by monotonicity of f_model alone, but that is a mathematical-validity issue distinct from the circularity.

full rationale

InvestAlign's empirical alignment claims are not circular: the SFT dataset is generated from the closed-form solution (11) cited to Wang and Zhao 2024b, and the alignment of InvestAgents with real-user decisions in P3, P1, and P2 is measured against separately collected questionnaire data using overall MSE. No parameter of the theoretical solution is fitted to those test data, and the P1/P2 evaluations use held-out complex-problem data, so those results have independent empirical content. The structural circularity is confined to the theoretical faster-convergence result advertised in the abstract and Section 4.2. There, the paper explicitly models real-user investment decisions as the theoretical optimal decision plus uniform noise, then 'proves' that training on the theoretical solution converges faster. That is not a comparison against independent real-user behavior; it is a comparison between a distribution and its own noise-smoothed version, so the advertised advantage is built into the input model. In addition, the key step Î<Ĩ in Eq. (27) does not follow from the stated monotonicity and nesting assumptions and can fail, which further weakens the proof independently of circularity. The self-citations to the authors' prior work supply the P3 solution and the η≈constant approximation; these are prior stated-assumption results rather than circular renaming, though they are load-bearing. The experimental gradient-norm and training-loss comparisons using actual real-user data provide some independent empirical support for the convergence claim, which is why the circularity is partial rather than total.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a chain of modeling assumptions: the optimal-investment model itself, a noise model for real-user data, simplified sigmoid/cross-entropy loss, support nesting, and an approximation on eta. The free parameters are sparse but include the ad hoc noise bound and data-derived attribute ranges. No new physical or theoretical entities are introduced. The convergence result carries the largest circularity burden because it is largely entailed by the noise assumption.

free parameters (2)
  • Noise bound epsilon in real-user model = unspecified
    Introduced ad hoc in Section 4.2 to model real-user decisions as theoretical solution plus uniform noise U(-eps, eps); no empirical estimate is provided, yet it drives the gradient-norm comparison.
  • Attribute ranges Salpha1 and Stheta1 = [0.09,0.38] and [0,1e-7]
    Determined from the 119-participant sample in Section 3.3 and used to define participant classes and test attributes; not independently justified as representative of broader investor populations.
assumptions (5)
  • domain assumption The Merton-style optimal investment model with exponential utility and herding penalty (Eq. 2) describes real investor decision-making.
    Central premise for using Eq. (11) as training labels; Section 3.3 provides validation with a mean correlation test against 0.85, but the model itself is taken from prior literature.
  • ad hoc to paper Real-user decisions equal the theoretical optimal solution plus i.i.d. uniform noise.
    Stated in Section 4.2 to make the convergence analysis tractable; this assumption is load-bearing for the faster-convergence claim and is not validated against the collected data.
  • ad hoc to paper The LLM output layer is a Sigmoid and the SFT loss is a continuous cross-entropy between densities of investment decisions.
    Assumed in Section 4.2 for mathematical tractability; acknowledged by the authors as a simplification of the real LLM output layer.
  • ad hoc to paper The supports nest as P1hat(t) subset P1tilde(t) subset P1(t), and the model density f_P1(t) is monotonically decreasing.
    This is used in Section 4.2 to prove Eq. (27)-(28); the nesting is asserted rather than derived from data or model properties.
  • ad hoc to paper The parameter eta in Eq. (11) remains approximately constant when alpha1 and theta1 vary slightly.
    Invoked in Appendix A.7 to derive the Pareto approximation f_P1hat(x) ~ c/x^2; this approximation is needed for the subsequent gradient-norm comparison.

how reviews work

0 comments
Cite this review

Pith. "Pith review of InvestAlign: Overcoming Data Scarcity in Aligning Large Language Models with Investor Decision-Making Processes under Herd Behavior." pith.science (2026). https://pith.science/paper/V6MY3UX7

@misc{pith2026250706528,
  author       = {Pith},
  title        = {Pith review of: InvestAlign: Overcoming Data Scarcity in Aligning Large Language Models with Investor Decision-Making Processes under Herd Behavior},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V6MY3UX7}},
  note         = {Machine review of arXiv:2507.06528}
}
read the original abstract

Aligning Large Language Models (LLMs) with investor decision-making processes under herd behavior is a critical challenge in behavioral finance, which grapples with a fundamental limitation: the scarcity of real-user data needed for Supervised Fine-Tuning (SFT). While SFT can bridge the gap between LLM outputs and human behavioral patterns, its reliance on massive authentic data imposes substantial collection costs and privacy risks. We propose InvestAlign, a novel framework that constructs high-quality SFT datasets by leveraging theoretical solutions to similar and simple optimal investment problems rather than complex scenarios. Our theoretical analysis demonstrates that training LLMs with InvestAlign-generated data achieves faster parameter convergence than using real-user data, suggesting superior learning efficiency. Furthermore, we develop InvestAgent, an LLM agent fine-tuned with InvestAlign, which demonstrates significantly closer alignment to real-user data than pre-SFT models in both simple and complex investment problems. This highlights our proposed InvestAlign as a promising approach with the potential to address complex optimal investment problems and align LLMs with investor decision-making processes under herd behavior. Our code is publicly available at https://github.com/thu-social-network-research-group/InvestAlign.

Figures

Figures reproduced from arXiv: 2507.06528 by the authors.

Figure 1
Figure 1. Overview of InvestAlign. scale real-user data in Supervised Fine-Tuning (SFT) (Zhang et al., 2023). Fine-tuned with spe￾cific training datasets, LLMs can better generate investor behavior in complex problems. However, it faces the following obstacles. Collecting real￾user data can be costly due to the wide variation in investment attributes like risk preference and herd degree (Abbot, 2017). Additionally, many inves… view at source ↗
Figure 2
Figure 2. Comparison of real-user data (P˜ 1), pre-SFT LLMs’ decision (P1), and theoretical solution (Pˆ 1) on P3. theoretical solution with the same investment at￾tribute as {Pˆi 1 (t)}t∈T , respectively. We calculate the difference di and correlation coefficient ρi be￾tween {P˜i 1 (t)}t∈T and {Pˆi 1 (t)}t∈T , and conduct t-tests on the means of the differences {di}i∈I and the correlation coefficients {ρi}i∈I (Shao, 2008), r… view at source ↗
Figure 3
Figure 3. Comparison of the gradient norms between using theoretical solution and real-user data. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (15 more)
Figure 4
Figure 4. Figure 4: Comparison of the logarithmic training losses between using theoretical solution and real-user data. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Comparison of real-user data (P˜ 1), pre-SFT LLMs’ investment decision (P1), InvestAgents’ investment decision (P SFT 1 ), and theoretical solution (Pˆ 1) on P3. Overall MSE Pre-SFT InvestAgent Reduction P3 GPT-3.5 4.44 1.72 -61.26% Qwen-2 3.97 2.16 -45.59% Llama-3.1 4…
Figure 6
Figure 6. Figure 6: The experimental results of GLM-4 corresponding to Figures 2–5. The experimental results of GLM-4 in [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Comparison of real-user data (P˜ 1) and pre-SFT LLMs’ investment decision (P1) on P1. A.6 Statistical Tests’ Results of the Differences and Correlation Coefficients We define the difference di and correlation coefficient ρi between {P˜i 1 (t)}t∈T and {Pˆi 1 (t)}t∈T as …
Figure 8
Figure 8. Figure 8: Validation of the hypothesis H1 in P1. 2 4 6 8 10 Time Step t 3 4 5 6 7 P 1 a n d P 2 P1 : 1 = 5 × 10 8 P1 : 1 = 9 × 10 8 P2 (a) Real-user data 2 4 6 8 10 Time Step t 2 3 4 5 6 7 8 9 P 1 a n d P 2 P1 : 1 = 5 × 10 8 P1 : 1 = 9 × 10 8 P2 (b) Qwen-2 2 4 6 8 10 Time Step t…
Figure 9
Figure 9. Figure 9: , we observe that as the influence coefficient θ1 increases, A1’s decision {P SFT 1 (t)}t∈T gradually converges toward A2’s optimal decision {Pˆ 2(t)}t∈T . For P2, we plot both A1’s decisions {P SFT 1 (t)}t∈T and A2’s decisions {P SFT 2 (t)}t∈T under different influenc…
Figure 10
Figure 10. Figure 10: Validation of the hypothesis H1 in P2 [PITH_FULL_IMAGE:figures/full_fig_p021_10.png]
Figure 11
Figure 11. Figure 11: , we observe that as the influence coefficient θ increases, the mean of the sum of the two agents’ terminal funds E[X1(T) + X2(T)] exhibits a monotonic decrease, which is consistent with and validates the hypothesis H2. 5.0 5.5 6.0 6.5 7.0 7.5 Influence Coefficient 1e…
Figure 12
Figure 12. Figure 12: Prompt for pre-SFT LLMs and InvestAgents in P3 [PITH_FULL_IMAGE:figures/full_fig_p024_12.png]
Figure 13
Figure 13. Figure 13: Prompt for SFT [PITH_FULL_IMAGE:figures/full_fig_p025_13.png]
Figure 14
Figure 14. Figure 14: Prompt for pre-SFT LLMs and InvestAgents in P1 [PITH_FULL_IMAGE:figures/full_fig_p027_14.png]
Figure 15
Figure 15. Figure 15: Prompt for pre-SFT LLMs and InvestAgents in P2 [PITH_FULL_IMAGE:figures/full_fig_p029_15.png]
Figure 16
Figure 16. Figure 16: Questionnaire for real-user data validation in [PITH_FULL_IMAGE:figures/full_fig_p030_16.png]
Figure 17
Figure 17. Figure 17: Questionnaire for real-user data validation in [PITH_FULL_IMAGE:figures/full_fig_p031_17.png]
Figure 18
Figure 18. Figure 18: Questionnaire for real-user data validation in [PITH_FULL_IMAGE:figures/full_fig_p032_18.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 11 canonical work pages

  1. [2]

    FedSyn: Synthetic Data Generation using Federated Learning

    FedSyn: Synthetic data generation using fed- erated learning.arXiv preprint arXiv:2203.05931. Sushil Bikhchandani and Sunil Sharma. 2000. Herd behavior in financial markets.IMF Staff papers, 47(3):279–310. Daniil A Boiko, Robert MacKnight, and Gabe Gomes

    work page Pith review arXiv

  2. [3]

    A probability p of receiving $20, and a probability 1 - p of receiving nothing

    Your Investment Characteristics (1) At what probability (denoted by p) are the following two choices indifferent to you? A. A probability p of receiving $20, and a probability 1 - p of receiving nothing. B. Receiving $6. [________] (2) When making a decision, how much do you rely on the investment assistant? Please directly give an integer between 0 and 1...

    work page

  3. [8]

    Investment Decisions Now, you have 10 million dollars for investment and savings, and the investment assistant recommends the following investment proportions for the stock over the 10 years: [36.21%, 35.59%, 34.96%, 34.35%, 33.73%, 33.13%, 32.53%, 31.93%, 31.34%, 30.75%]. Considering the investment assistant’s recommendations, based on your own investmen...

    work page

  4. [10]

    You will establish a dedicated account to manage this retirement fund

    Task Description Starting from next year, you plan to use a portion of your savings (10 million dollars) to invest in a stock and a deposit as part of your personal retirement fund. You will establish a dedicated account to manage this retirement fund. This means you will make a one-time deposit of 10 million dollars into this account and will not deposit...

    work page

  5. [11]

    Investment Decisions Now, you have 10 million dollars for investment and savings, and the investment assistant recommends the following investment proportion changes for the stock over the 10 years, i.e., the difference of the investment proportions in the next year and the previous year: [-0.62%, -0.63%, -0.61%, -0.62%, -0.60%, -0.60%, -0.60%, -0.59%, -0...

    work page

  6. [12]

    A probability p of receiving $20, and a probability 1 - p of receiving nothing

    Your Investment Characteristics (1) At what probability (denoted by p) are the following two choices indifferent to you? A. A probability p of receiving $20, and a probability 1 - p of receiving nothing. B. Receiving $6. [________] (2) When making a decision, how much do you rely on the investment assistant? Please directly give an integer between 0 and 1...

    work page

  7. [13]

    You will establish a dedicated account to manage this retirement fund

    Task Description Starting from next year, you plan to use a portion of your savings (10 million dollars) to invest in a stock and a deposit as part of your personal retirement fund. You will establish a dedicated account to manage this retirement fund. This means you will make a one-time deposit of 10 million dollars into this account and will not deposit...

    work page

  8. [14]

    A probability p1 of receiving $20, and a probability 1 - p1 of receiving nothing

    Your and Your Partner’s Investment Attributes (Completed by Two Participants Together) (1) At what probability (denoted by p1) are the following two choices indifferent to you? A. A probability p1 of receiving $20, and a probability 1 - p1 of receiving nothing. B. Receiving $6. At what probability (denoted by p2) are the following two choices indifferent ...

    work page

Show all 14 references
  1. [15]

    Investment Decisions (Completed by Two Participants Together) Now, you have 10 million dollars for investment and savings, and the investment assistant recommends the following investment proportions for the stock over the 10 years: [36.21%, 35.59%, 34.96%, 34.35%, 33.73%, 33....

  2. [1007]

    On a scale [0,10] , how much do you rely on the investment assistant when making decisions, where 10 represents the highest reliance level and 0 the lowest?

    Springer. Jun Shao. 2008.Mathematical statistics. Springer Sci- ence & Business Media. Jacob Steinhardt. 2023. Emergent deception and emer- gent optimization.Bounded Regret, 19:2023. Cheng Wang, Chuwen Wang, Wang Zhang, Shirong Zeng, Yu Zhao, Ronghui Ning, and Changjun Jiang. ...

    2008 arXiv

  3. [2022]

    Tucker Balch, Vamsi K Potluru, Deepak Paramanand, and Manuela Veloso

    Training a helpful and harmless assistant with reinforcement learning from human feedback.arXiv preprint arXiv:2204.05862. Tucker Balch, Vamsi K Potluru, Deepak Paramanand, and Manuela Veloso. 2024. Six levels of privacy: A framework for financial synthetic data.arXiv preprint...

    2024 arXiv

  4. [2023]

    Samuel R Bowman, Jeeyoon Hyun, Ethan Perez, Edwin Chen, Craig Pettit, Scott Heiner, Kamil˙e Lukoši¯ut˙e, Amanda Askell, Andy Jones, Anna Chen, et al

    Emergent autonomous scientific research ca- pabilities of large language models.arXiv preprint arXiv:2304.05332. Samuel R Bowman, Jeeyoon Hyun, Ethan Perez, Edwin Chen, Craig Pettit, Scott Heiner, Kamil˙e Lukoši¯ut˙e, Amanda Askell, Andy Jones, Anna Chen, et al. 2022. Measurin...

    2022 arXiv

  5. [2024]

    Robert C Merton

    Augmenting large language models with chem- istry tools.Nature Machine Intelligence, pages 1–11. Robert C Merton. 1969. Lifetime portfolio selection under uncertainty: The continuous-time case.The Review of Economics and Statistics, pages 247–257. Ethan Perez, Sam Ringer, Kami...

    1969 arXiv

  6. [2025]

    Neural Computing and Applications, pages 1–15

    Large language models in finance (FinLLMs). Neural Computing and Applications, pages 1–15. Jiaxiang Li, Siliang Zeng, Hoi-To Wai, Chenliang Li, Alfredo Garcia, and Mingyi Hong. 2024a. Getting more juice out of the sft data: Reward learning from human demonstration improves sft...

    2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.