Pith. sign in

REVIEW 3 major objections 5 minor 143 references

Counterfactual Analysis via Large Language Models

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A prompt-engineered large language model can construct loan counterfactuals with accuracy close to a machine-learning benchmark, and the paper argues this makes LLMs a well-suited tool for counterfactual analysis.

desk verdict A legitimate first attempt at LLM-based counterfactual ROI generation, but the two validation legs are both too weak to support the 'well-suited, capable' claim; worth a rigorous revision. read the letter →

arxiv 2608.05367 v1 pith:E723XTZG submitted 2026-08-05 cs.AI q-fin.GN

Zonghao Yang This is my paper · ORCID
classification cs.AIq-fin.GN
keywords counterfactualanalysislargelanguagemodelspromptengineeringonlinelendingreturnoninvestmentcausalreasoningforecastencompassing
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 sets out to establish that large language models are a workable tool for counterfactual analysis, not just for text prediction. It tests this on online lending: given a loan and its realized outcome under the original interest rate, the model predicts what the return on investment would have been under an alternative interest rate. Two pieces of evidence are offered. First, with careful prompting, the model's predictions of realized loan returns reach an out-of-sample R-squared of 2.84%, close to the 3.48% achieved by a gradient-boosted regression benchmark, and a forecast-encompassing test indicates the model adds predictive information beyond that benchmark. Second, the counterfactual outcomes it produces follow two directional monotonicity rules in 85.3% and 99.6% of cases, which the paper interprets as evidence of logical coherence.

What carries the argument

The machinery is a structured text prompt that stacks five ingredients: loan and borrower characteristics, platform-level averages, the observed outcome under the original interest rate (the anchor), a machine-learning counterfactual prediction, and an instruction to deliberate as a panel of four experts until consensus. The model's final answer is forced into JSON with status and duration fields, from which the ROI is recomputed. The validation machinery is a pair of directional rules: a charged-off loan stays charged off if the interest rate rises, and a fully paid loan stays fully paid if the interest rate falls; the paper treats agreement with these rules as evidence of coherent causal reasoning.

What would settle it

Run a null baseline that always predicts the observed loan status (and, say, the realized duration) for every counterfactual interest-rate change. If that baseline matches or exceeds the reported 85.3% and 99.6% agreement with the two rules, the agreement rates cannot establish that the LLM is reasoning about the counterfactual rather than repeating the outcome it was told.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a large language model can be turned into a counterfactual reasoner by how it is prompted. Asked to play a panel of credit-risk experts and given the loan and borrower details, platform-level statistics, the actual outcome under the original rate, and a machine-learning prediction under the alternative rate, the model returns a loan status and duration for the hypothetical scenario. Only 53.8% of its counterfactual status-and-duration predictions exactly match a fitted mixture-cure survival model, so the model is not simply echoing the machine prediction. The paper argues that predictive parity on realized outcomes together with strong adherence to the two monotonicity rules supports the conclusion that large language models capture the causal relationship between interest rate and loan outcome well enough to build sensible counterfactuals.

Load-bearing premise

The load-bearing premise is that agreement with the two directional rules demonstrates genuine counterfactual reasoning; if a trivial model that simply repeats the observed loan status also passes those rules, this evidence no longer distinguishes causal understanding from plausible text generation.

Editorial extensions

If this is right

  • With the right prompt scaffold, an LLM can serve as a practical substitute for a fitted regression or survival model when generating counterfactual loan outcomes under alternative interest-rate schemes.
  • Prompt engineering matters as much as fine-tuning: switching the model's role from borrower to credit expert moves classification from near-random to useful levels, and adding chain-of-thought or tree-of-thought instructions further improves prediction.
  • Combining an LLM with a machine-learning forecast yields better predictions than either alone, because the LLM revises the machine prediction using reasoning expressed in text.
  • The paper's prompting principles—role-play, deliberate reasoning, and anchoring on observed outcomes—are presented as transferable to counterfactual questions in other decision domains, not just lending.

Reading between the lines

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

  • Going beyond the paper, the two monotonicity rules check only the direction of loan status; a natural next test would compare the reported agreement rates against a trivial baseline that repeats the observed loan status for every counterfactual.
  • Going beyond the paper, the stochasticity of LLM responses noted in the paper could be used to output a distribution of counterfactual outcomes rather than a single point, giving lenders an uncertainty range around the counterfactual ROI.
  • Going beyond the paper, the same prompt scaffold could be validated on a simulated repayment process where counterfactual outcomes are actually known, which would isolate genuine causal reasoning from plausible text generation.
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

3 major / 5 minor

Summary. The paper investigates whether GPT-3.5 can be used for counterfactual analysis in online lending. It first evaluates GPT-3.5's predictive performance for realized loan ROI under seven prompt designs, reporting that prompt engineering raises R² from 1.97% to 2.84%, compared with 3.48% for gradient-boosted regression. It then generates counterfactual ROIs under alternative interest rates from Gopal et al. (2024), compares the results with a mixture-cure survival model, and validates the LLM outputs using two monotonicity rules. The paper concludes that LLMs are a novel, well-suited, and capable approach for counterfactual analysis.

Significance. If the central claim were established, the paper would be a useful proof-of-concept for using LLMs as counterfactual generators in a high-stakes decision context, and it provides a concrete prompt-engineering recipe together with a comparison against an existing survival model. The paper is transparent about API costs and about the low absolute predictive performance, which is a strength. However, the evidence as presented is not load-bearing: the reported R² depends on a bias-correction step whose estimation sample is unclear, and the counterfactual coherence test is largely satisfied by a trivial baseline that repeats the observed outcome. The contribution is promising but currently unproven.

major comments (3)
  1. [Section 4.1, footnote 3, Table 1 Panel C] The reported R² values in Table 1, Panel C depend on a bias-correction regression (ROI = c0 + c1 * ROI_hat + epsilon) whose estimation sample is not specified. If the coefficients c0 and c1 are estimated on the same 10,000 observations used to compute R², then the corrected predictions are in-sample linear fits rather than out-of-sample predictions; a linear regression of actual on predicted can absorb the sample covariance and turn a negative raw R² into a positive reported R². Please either fit the correction on the 2013 training sample or use cross-validation, and report the raw out-of-sample R² alongside the corrected value. The same ambiguity affects the 'same procedure' applied to ML predictions and therefore the headline comparison of 2.84% versus 3.48%.
  2. [Section 4.2.2, Figure 3] The coherence validation is not probative as presented. The prompt in Figure 3 explicitly gives the actual loan outcome under the original rate, and Section 3.2 reports that alternative rates differ from original rates by only 2.8 percentage points on average. Under these conditions, a trivial model that repeats the observed status for every counterfactual rate satisfies Rule 1 for every charged-off loan with an increased rate and Rule 2 for every fully-paid loan with a decreased rate, by construction. The paper does not report the fraction of counterfactuals for which this null baseline satisfies the two rules, nor any other baseline such as a constant-duration model. The reported 85.3% and 99.6% agreement can therefore reflect prompt-following or anchoring rather than causal reasoning. Moreover, the rules only restrict the sign of the status change; they place no constraint on duration or counterfactual ROI magnitude, so they cannot adjudicate the large discrepancies between the LLM and mixture-cure outcomes in Table 2 (SD 17.01 versus 10.09; mean 3.68% versus 4.34%).
  3. [Section 4.1, Table 1; Section 3.1] The evaluation protocol for prompt engineering is not fully specified. Seven prompt variants are compared and the best-performing prompt is selected using the same 10,000 observations on which R², F1, and AUC are reported, without an explicit validation split; this makes the reported 2.84% R² a selected maximum rather than an out-of-sample estimate. In addition, Section 3.1 states that the test sample is used 'for evaluation and fine-tuning,' and Footnote 4 reports fine-tuning on 665,185 samples; the paper should state explicitly whether the 10,000 evaluation observations are disjoint from the fine-tuning samples. If they overlap, the fine-tuning comparison and the ML benchmark are not out-of-sample.
minor comments (5)
  1. [Section 2] The phrase 'an suitable approach' should be 'a suitable approach'; also 'Key Words' should be 'Keywords' with standard formatting.
  2. [Section 4.1.2, Column (5)] The text says 'the AUC reaching 58.5%' for the chain-of-thought prompt, but Table 1 Panel B shows Column (4) has AUC 58.5% and Column (5) has AUC 62.0%; the text and table should be reconciled.
  3. [Equation (1)] Equation (1) writes the ML coefficient as (1 - φ), imposing a sum-to-one constraint on the two forecast coefficients; a standard forecast-encompassing test should estimate unrestricted coefficients and test the relevant restriction. Please clarify or re-estimate.
  4. [Section 4.2.1] The paper notes that LLM responses are stochastic, but all Table 1 and Table 2 numbers appear to be based on a single run; report means and standard deviations across multiple runs for the main 10,000-sample results.
  5. [Table 2] The abbreviation 'GCPP' is used without definition; spell out 'Gopal et al. (2024)' or define the abbreviation in the table note, and clarify whether '0.2% lower' refers to percentage points.

Circularity Check

2 steps flagged · score 6.0 of 10

Reported R² is computed after an in-sample linear recalibration, and the counterfactual-coherence test is passed by construction by a status-preserving baseline; the central 'suitable for counterfactual analysis' claim is therefore only partially supported by independent evidence.

  1. fitted input called prediction [Section 4.1.1, footnote 3; Table 1 Panel C]
    "The raw predictions by the large language model yield a large negative out-of-sample R-squared. We conjecture that this may be due to an embedded bias in the language description within the pre-training corpus. To correct such a bias, we regress the actual ROI on the predicted ROI, ROI=c 0 +c 1 ˆROI+ϵ. Then, we form the bias-corrected prediction by ˜ROI= ˆc0 + ˆc1 ˆROI. The reported R-squared is based on the bias-corrected predictions. We perform the same procedure on ML predictions to ensure a fair comparison."

    The R² values presented as GPT's predictive performance are not out-of-sample predictions: they are computed after fitting a linear regression of actual ROI on predicted ROI, and the paper does not state that the coefficients are estimated on a separate sample. Any predictor with nonzero covariance with the target can be linearly recalibrated to yield positive in-sample R², so the claimed 1.97%–2.84% (approaching gradient-boosted regression's 3.48%) is partly manufactured by the fitted correction. The comparison to ML is made after applying the same procedure, but the absolute 'predictive power' evidence for the suitability claim is an artifact of this fitting step unless a holdout is used.

  2. self definitional [Section 4.2.1 (Figure 3 prompt) and Section 4.2.2]
    "we also tell the LLM the actual loan outcome under the original interest rate set by LendingClub. This last piece of information serves as an anchor to link observed outcomes to unobserved counterfactuals. ... 1) A charged-off loan remains charged-off if the interest rate was increased, and 2) A fully-paid loan remains fully paid if the interest rate was decreased. ... We analyze all counterfactuals generated by the LLM and find that 85.3% abide by rule 1) and 99.6% abide by rule 2)."

    The coherence test is constructed from the same outcome that is inserted into the prompt. Because the prompt explicitly states the realized status (e.g., 'The borrower fully paid the loan in 36 months under the original interest rate of 14.99%'), a baseline that repeats the observed status for every hypothetical rate satisfies rule 1 for charged-off loans with higher counterfactual rates and rule 2 for fully-paid loans with lower counterfactual rates. With alternative rates only 2.8 percentage points away on average and no null baseline reported, the high rule-adherence rates are largely forced by the prompt's anchor and cannot by themselves demonstrate 'logical coherence and causal reasoning'; the validation metric is definitionally satisfied by a trivial response to the input.

full rationale

The paper has some independent content: the prediction exercise uses real LendingClub data, XGBoost and gradient-boosted benchmarks, an encompassing test, and the counterfactual section compares LLM output with the mixture-cure model; the 53.8% status/duration agreement shows the LLM is not merely copying ML predictions, and GPT-4o results provide an external sanity check. However, the two legs of the central claim are not fully self-contained. The reported R² is obtained after fitting a bias-correction regression on (apparently) the same evaluation sample, so the headline comparison '1.97% to 2.84%, closely approaching 3.48%' is partly a fitted-input artifact rather than an out-of-sample predictive result. The monotonicity validation in §4.2.2 uses rules that any status-preserving response to the prompt would satisfy in the relevant cases, so the 85.3%/99.6% coherence evidence cannot bear the weight placed on it. There is also a self-citation component (Gopal et al. 2024 includes the present author), but it is used mainly as a source of alternative rates and as a benchmark, not as the proof of the suitability claim; therefore it is not scored as a separate load-bearing circular step. Overall, the central suitability claim is partially circular/statistically forced, hence score 6 rather than 0 or 2.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central evaluation and validation rest on several modeling choices: the bias-correction regression (whose fitting sample is unclear), the assumption that prediction quality is a gate for counterfactual quality, the two ad hoc logical rules used to score counterfactual sensibility, and the adoption of the alternative interest rates from the author's own prior model. No new physical or mathematical entities are introduced.

free parameters (3)
  • Bias-correction coefficients (c0, c1) for ROI predictions = not reported
    Used to transform raw LLM and ML ROI predictions into reported R² values; if fit on the same test sample, the reported R² is an in-sample fit rather than an out-of-sample prediction.
  • Discount rate for ROI calculation = 2% annual
    Assumed input from prior practice; not fitted to data but affects all ROI values.
  • Prompt selection among seven variants = Column (7), four-expert tree-of-thought
    The best prompt is selected by comparing R² on the same 10,000-loan test set, without a separate validation set; this selection is a hand-chosen choice after inspecting results.
assumptions (4)
  • domain assumption Predictive accuracy on observed outcomes is a necessary condition for counterfactual validity (Section 4: 'necessary but not sufficient').
    The paper explicitly relies on this to justify using prediction performance as evidence of counterfactual suitability.
  • ad hoc to paper The two monotonicity rules (charged-off stays charged-off under rate increase; fully-paid stays fully-paid under rate decrease) are valid criteria for counterfactual sensibility (Section 4.2.2).
    The rules are introduced ad hoc for this context and only check status direction, not duration or ROI magnitude; no null baseline is provided.
  • domain assumption GPT-3.5's textual reasoning reflects genuine causal reasoning rather than plausible-sounding text (Section 4.2.1, Figure 4).
    The paper infers 'logical coherence and causal reasoning' from qualitative responses without quantitative validation.
  • domain assumption The alternative interest rates from Gopal et al. (2024) are a valid set of counterfactual interventions (Section 3.2).
    The rates are taken as given from the author's own prior working paper; the validity of the counterfactual scenario depends on this choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Counterfactual Analysis via Large Language Models." pith.science (2026). https://pith.science/paper/E723XTZG

@misc{pith2026260805367,
  author       = {Pith},
  title        = {Pith review of: Counterfactual Analysis via Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E723XTZG}},
  note         = {Machine review of arXiv:2608.05367}
}
read the original abstract

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

Figures

Figures reproduced from arXiv: 2608.05367 by the authors.

Figure 1
Figure 1. Interest Rate Distribution 0.05 0.10 0.15 0.20 0.25 0.30 Interest Rate 0 10 20 30 40 50 60 70 80 Density LendingClub Gopal et al. (2024) [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Loan Prediction Prompts and Responses Prompt: Four experts in credit risk assessment collaborate to predict the loan outcome of an unsecured personal loan. They will be provided with the loan specifics, the borrower's credit profile, LendingClub's average statistics, and machine learning predictions. Based on the information, they need to provide an unbiased prediction for the loan outcome: Will the borrower fully p… view at source ↗
Figure 3
Figure 3. Counterfactual Construction based on LLM: Prompt [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Counterfactual Construction based on LLM: Response (GPT-3.5) [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

143 extracted references · 73 canonical work pages

  1. [1]

    Fintech and banking:

    Thakor, Anjan V , journal=. Fintech and banking:. 2020 , publisher=

    2020

  2. [2]

    Instance-based credit risk assessment for investment decisions in

    Guo, Yanhong and Zhou, Wenjun and Luo, Chunyu and Liu, Chuanren and Xiong, Hui , journal=. Instance-based credit risk assessment for investment decisions in. 2016 , publisher=

    2016

  3. [3]

    A new integrated similarity measure for enhancing instance-based credit assessment in

    Guo, Yanhong and Jiang, Shuai and Qiao, Han and Chen, Feiting and Li, Yaocong , journal=. A new integrated similarity measure for enhancing instance-based credit assessment in. 2021 , publisher=

    2021

  4. [4]

    A multi-objective instance-based decision support system for investment recommendation in

    Babaei, Golnoosh and Bamdad, Shahrooz , journal=. A multi-objective instance-based decision support system for investment recommendation in. 2020 , publisher=

    2020

  5. [5]

    Parametric portfolio policies: Exploiting characteristics in the cross-section of equity returns , author=. Rev. Financial Stud. , volume=. 2009 , publisher=

    2009

  6. [6]

    European Journal of Operational Research , volume=

    Benchmarking state-of-the-art classification algorithms for credit scoring: An update of research , author=. European Journal of Operational Research , volume=. 2015 , publisher=

    2015

  7. [7]

    Opportunities and challenges in online marketplace lending , type =

  8. [8]

    Chen, Tianqi and Guestrin, Carlos , booktitle=

Show all 143 references
  1. [9]

    22nd ACM SIGKDD Internat

    Proc. 22nd ACM SIGKDD Internat. Conf. Knowledge Discovery Data Mining

  2. [10]

    Expert Syst

    A boosted decision tree approach using Bayesian hyper-parameter optimization for credit scoring , author=. Expert Syst. Appl. , volume=. 2017 , publisher=

    2017

  3. [11]

    Expert Syst

    Risk assessment in social lending via random forests , author=. Expert Syst. Appl. , volume=. 2015 , publisher=

    2015

  4. [12]

    Data-driven optimization of

    Byanjankar, Ajay and Mezei, J. Data-driven optimization of. Intell. Syst. Account. Finance Manag. , volume=. 2021 , publisher=

    2021

  5. [13]

    IEEE PAMI , volume=

    The random subspace method for constructing decision forests , author=. IEEE PAMI , volume=. 1998 , publisher=

    1998

  6. [14]

    Chawla, Nitesh V and Bowyer, Kevin W and Hall, Lawrence O and Kegelmeyer, W Philip , journal=

  7. [15]

    Management Sci

    A generalized approach to portfolio optimization: Improving performance by constraining portfolio norms , author=. Management Sci. , volume=. 2009 , publisher=

    2009

  8. [16]

    Optimal versus naive diversification: How inefficient is the 1/

    DeMiguel, Victor and Garlappi, Lorenzo and Uppal, Raman , journal=. Optimal versus naive diversification: How inefficient is the 1/. 2009 , publisher=

    2009

  9. [17]

    Shrinking the cross-section , author=. J. Financial Econom. , volume=. 2020 , publisher=

    2020

  10. [18]

    Application of instance-based entropy fuzzy support vector machine in

    Cho, Poongjin and Chang, Woojin and Song, Jae Wook , journal=. Application of instance-based entropy fuzzy support vector machine in. 2019 , publisher=

    2019

  11. [19]

    The basic neoclassical model , author=

    Production, growth and business cycles: I. The basic neoclassical model , author=. Journal of Monetary Economics , volume=. 1988 , publisher=

    1988

  12. [20]

    1992 , publisher=

    Microeconomic analysis , author=. 1992 , publisher=

    1992

  13. [21]

    The democratization of personal consumer loans?

    Herzenstein, Michal and Andrews, Rick L and Dholakia, Utpal M and Lyandres, Evgeny , year=. The democratization of personal consumer loans?

  14. [22]

    Evaluating credit risk and loan performance in online

    Emekter, Riza and Tu, Yanbin and Jirasakuldech, Benjamas and Lu, Min , journal=. Evaluating credit risk and loan performance in online. 2015 , publisher=

    2015

  15. [23]

    Heterogeneous ensemble for default prediction of

    Li, Wei and Ding, Shuai and Chen, Yi and Yang, Shanlin , journal=. Heterogeneous ensemble for default prediction of. 2018 , publisher=

    2018

  16. [24]

    A novel behavioral scoring model for estimating probability of default over time in

    Wang, Zhao and Jiang, Cuiqing and Ding, Yong and Lyu, Xiaozhong and Liu, Yao , journal=. A novel behavioral scoring model for estimating probability of default over time in. 2018 , publisher=

    2018

  17. [25]

    Cost-sensitive boosted tree for loan evaluation in

    Xia, Yufei and Liu, Chuanzhe and Liu, Nana , journal=. Cost-sensitive boosted tree for loan evaluation in. 2017 , publisher=

    2017

  18. [26]

    Management Sci

    Screening peers softly: Inferring the quality of small borrowers , author=. Management Sci. , volume=. 2016 , publisher=

    2016

  19. [27]

    Neural Learning of Online Consumer Credit Risk

    Wang, Di and Wu, Qi and Zhang, Wen , year=. Neural Learning of Online Consumer Credit Risk

  20. [28]

    Mitigating adverse selection in

    Weiss, Gregor NF and Pelger, Katharina and Horsch, Andreas , journal=. Mitigating adverse selection in

  21. [29]

    Social interactions in

    Herrero-Lopez, Sergio , booktitle=. Social interactions in

  22. [30]

    Herding behavior in online

    Lee, Eunkyoung and Lee, Byungtae , journal=. Herding behavior in online. 2012 , publisher=

    2012

  23. [31]

    Data-driven robust credit portfolio optimization for investment decisions in

    Chi, Guotai and Ding, Shijie and Peng, Xiankun , journal=. Data-driven robust credit portfolio optimization for investment decisions in. 2019 , publisher=

    2019

  24. [32]

    http://www.nber.org/papers/w16855

    Freedman, Seth M and Jin, Ginger Zhe. Learning by Doing with Asymmetric Information: Evidence from Prosper.com. 2011. doi:10.3386/w16855 , URL = "http://www.nber.org/papers/w16855", abstract =

    2011 doi

  25. [33]

    Market mechanisms in online

    Wei, Zaiyan and Lin, Mingfeng , journal=. Market mechanisms in online. 2017 , publisher=

    2017

  26. [34]

    Determinants of default in

    Serrano-Cinca, Carlos and Guti. Determinants of default in. PLoS One , volume=. 2015 , publisher=

    2015

  27. [35]

    Crossing the next frontier: The role of

    Drummer, Daniel and Feuerriegel, Stefan and Neumann, Dirk , journal=. Crossing the next frontier: The role of. 2017 , publisher=

    2017

  28. [36]

    2016 , publisher=

    Deep learning , author=. 2016 , publisher=

    2016

  29. [37]

    Nature , volume=

    Deep learning , author=. Nature , volume=

  30. [38]

    Deepcredit:

    Yang, Zhi and Zhang, Yusi and Guo, Binghui and Zhao, Ben Y and Dai, Yafei , crossref=. Deepcredit:. Proc. 12th Int. AAAI Conf. Web Soc. Media , pages=

  31. [39]

    12th Int

    Proc. 12th Int. AAAI Conf. Web Soc. Media

  32. [40]

    Deep learning for credit scoring: Do or don't? , author=. Eur. J. Oper. Res. , volume=. 2021 , publisher=

    2021

  33. [41]

    Competition and Distortion: A Theory of Information Bias on the

    Wu, Zhenhua and Hu, Lin and Lin, Zhijie and Tan, Yong , journal=. Competition and Distortion: A Theory of Information Bias on the. 2021 , publisher=

    2021

  34. [42]

    Available at SSRN 1107307 , year=

    Love & loans: The effect of beauty and personal characteristics in credit markets , author=. Available at SSRN 1107307 , year=

  35. [43]

    What’s in a Picture?

    Pope, Devin G and Sydnor, Justin R , journal=. What’s in a Picture?. 2011 , publisher=

    2011

  36. [44]

    Trust and credit: The role of appearance in

    Duarte, Jefferson and Siegel, Stephan and Young, Lance , journal=. Trust and credit: The role of appearance in. 2012 , publisher=

    2012

  37. [45]

    Judging borrowers by the company they keep: Friendship networks and information asymmetry in online

    Lin, Mingfeng and Prabhala, Nagpurnanand R and Viswanathan, Siva , journal=. Judging borrowers by the company they keep: Friendship networks and information asymmetry in online. 2013 , publisher=

    2013

  38. [46]

    Group membership, relationship banking and loan default risk: The case of online social lending , author=. Bank. Fin. Rev. , volume=

  39. [47]

    Friendships in online

    Liu, De and Brass, Daniel J and Lu, Yong and Chen, Dongyu , journal=. Friendships in online. 2015 , publisher=

    2015

  40. [48]

    A new aspect on

    Ma, Lin and Zhao, Xi and Zhou, Zhili and Liu, Yuanyuan , journal=. A new aspect on. 2018 , publisher=

    2018

  41. [49]

    Loan default prediction by combining soft information extracted from descriptive text in online

    Jiang, Cuiqing and Wang, Zhao and Wang, Ruiya and Ding, Yong , journal=. Loan default prediction by combining soft information extracted from descriptive text in online. 2018 , publisher=

    2018

  42. [50]

    Cheap talk?

    Xu, Jennifer J and Chau, Michael , journal=. Cheap talk?. 2018 , publisher=

    2018

  43. [51]

    The use of profit scoring as an alternative to credit scoring systems in

    Serrano-Cinca, Carlos and Guti. The use of profit scoring as an alternative to credit scoring systems in. Dec. Sup. Syst. , volume=. 2016 , publisher=

    2016

  44. [52]

    Mining semantic soft factors for credit risk evaluation in

    Wang, Zhao and Jiang, Cuiqing and Zhao, Huimin and Ding, Yong , journal=. Mining semantic soft factors for credit risk evaluation in. 2020 , publisher=

    2020

  45. [53]

    Research commentary—too big to fail: Large samples and the p-value problem , author=. Inform. Systems Res. , volume=. 2013 , publisher=

    2013

  46. [54]

    Information Systems Research , volume=

    Crowds, lending, machine, and bias , author=. Information Systems Research , volume=. 2021 , publisher=

    2021

  47. [55]

    Regularization and variable selection via the elastic net , author=. J. R. Stat. Soc., Ser. , volume=. 2005 , publisher=

    2005

  48. [56]

    Portfolio selection , author=. J. Finance , volume=. 1952 , publisher=

    1952

  49. [57]

    How active is your fund manager?

    Cremers, KJ Martijn and Petajisto, Antti , journal=. How active is your fund manager?. 2009 , publisher=

    2009

  50. [58]

    Private equity performance: Returns, persistence, and capital flows , author=. J. Finance , volume=. 2005 , publisher=

    2005

  51. [59]

    Investments , author =

  52. [60]

    Changes in US family finances from 2016 to 2019: Evidence from the Survey of Consumer Finances , author=. Fed. Reserve Bulletin , volume=

    2016

  53. [61]

    Noise sensitivity of portfolio selection under various risk measures , author=. J. Bank. Financ. , volume=. 2007 , publisher=

    2007

  54. [62]

    Optimal regulation of

    Nemoto, Naoko and Storey, David J and Huang, Bihong , institution =. Optimal regulation of

  55. [63]

    Prosocial compliance in

    Du, Ninghua and Li, Lingfang and Lu, Tian and Lu, Xianghua , journal=. Prosocial compliance in. 2020 , publisher=

    2020

  56. [64]

    Financial Times , url =

    Kate, Beioley and Nicholas, Megaw , year =. Financial Times , url =

  57. [65]

    Michaud, Richard O , journal=. The. 1989 , publisher=

    1989

  58. [66]

    Expected returns and expected growth in rents of commercial real estate , author=. Rev. Financial Stud. , volume=. 2010 , publisher=

    2010

  59. [67]

    Determinants of borrowers' default in

    Polena, Michal and Regner, Tobias , journal=. Determinants of borrowers' default in. 2018 , publisher=

    2018

  60. [68]

    Deep Learning Credit Risk Modeling , author=. J. Fixed Inc. , volume=. 2021 , publisher=

    2021

  61. [69]

    Credit lines and credit utilization , author=. J. Money Credit Bank. , volume=. 2006 , publisher=

    2006

  62. [70]

    Default risk and income fluctuations in emerging economies , author=. Amer. Econ. Rev. , volume=

  63. [71]

    Consumer-lending discrimination in the

    Bartlett, Robert and Morse, Adair and Stanton, Richard and Wallace, Nancy , journal=. Consumer-lending discrimination in the. 2022 , publisher=

    2022

  64. [72]

    Network formation and financial inclusion in

    Katsamakas, Evangelos and S. Network formation and financial inclusion in. Systems , volume=. 2022 , publisher=

    2022

  65. [73]

    , author=

    All Models are Wrong, but Many are Useful: Learning a Variable's Importance by Studying an Entire Class of Prediction Models Simultaneously. , author=. J. Mach. Learn. Res. , volume=

  66. [74]

    Financial inclusion and alternate credit scoring for the millennials:

    Agarwal, Sumit and Alok, Shashwat and Ghosh, Pulak and Gupta, Sudip , institution =. Financial inclusion and alternate credit scoring for the millennials:

  67. [75]

    The rise of finance companies and fintech lenders in small business lending , author=. Rev. Financial Stud. , volume=. 2022 , publisher=

    2022

  68. [76]

    2022 , publisher=

    Balyuk, Tetyana , journal=. 2022 , publisher=

    2022

  69. [77]

    Fintech lending , author=. Annu. Rev. Financial Econ. , volume=

  70. [78]

    The Fall of Online

    Huang, Hui and Wang, Meng Lu , journal=. The Fall of Online

  71. [79]

    The role of technology in mortgage lending , author=. Rev. Financial Stud. , volume=. 2019 , publisher=

    2019

  72. [80]

    Why do peer-to-peer (

    Klein, Galit and Shtudiner, Zeev and Zwilling, Moti , journal=. Why do peer-to-peer (. 2021 , publisher=

    2021

  73. [81]

    Empirical asset pricing via machine learning , author=. Rev. Financial Stud. , volume=. 2020 , publisher=

    2020

  74. [82]

    Kingma, Diederik P and Ba, Jimmy Lei , booktitle=. Adam:

  75. [83]

    Proc. Int. Conf. Learn. Representations

  76. [84]

    Who should buy long-term bonds? , author=. Am. Econ. Rev. , volume=. 2001 , publisher=

    2001

  77. [85]

    Bond ladders and optimal portfolios , author=. Rev. Financial Stud. , volume=. 2011 , publisher=

    2011

  78. [86]

    and Yang, Zonghao , title=

    Gopal, Ram and Qiao, Xiao and Strub, Moris S. and Yang, Zonghao , title=. Inform. Systems Res. , volume=. 2024 , publisher=

    2024

  79. [87]

    2009 , publisher=

    Consumer credit models: Pricing, profit and portfolios , author=. 2009 , publisher=

    2009

  80. [88]

    Optimizing prices for consumer credit , author=. J. Revenue Pricing Manag. , volume=. 2013 , publisher=

    2013

  81. [89]

    Management Sci

    Risk-based loan pricing: Portfolio optimization approach with marginal risk contribution , author=. Management Sci. , volume=. 2020 , publisher=

    2020

  82. [90]

    J Financ

    Borrowing constraints, household debt, and racial discrimination in loan markets , author=. J Financ. Intermed. , volume=. 1993 , publisher=

    1993

  83. [91]

    The behavior of consumer loan rates during the 1990 credit slowdown , author=. J. Bank. Financ. , volume=. 1996 , publisher=

    1990

  84. [92]

    2023 , type =

    FinTech Lending with LowTech Pricing , author=. 2023 , type =

    2023

  85. [93]

    Designing win-win financial loan products for consumers and businesses , author=. J. Oper. Res. Soc. , volume=. 2005 , publisher=

    2005

  86. [94]

    Risk-based pricing of interest rates for consumer loans , author=. J. Monet. Econ. , volume=. 2006 , publisher=

    2006

  87. [95]

    FinTech borrowers: Lax screening or cream-skimming? , author=. Rev. Financial Stud. , volume=. 2021 , publisher=

    2021

  88. [96]

    , journal=

    Sharpe, William F. , journal=. Capital asset prices:. 1964 , publisher=

    1964

  89. [97]

    Predictably unequal?

    Fuster, Andreas and Goldsmith-Pinkham, Paul and Ramadorai, Tarun and Walther, Ansgar , journal=. Predictably unequal?. 2022 , publisher=

    2022

  90. [98]

    Goldstein, Itay and Jiang, Wei and Karolyi, G Andrew , journal=. To. 2019 , publisher=

    2019

  91. [99]

    On the pricing of corporate debt:

    Merton, Robert C , journal=. On the pricing of corporate debt:. 1974 , publisher=

    1974

  92. [100]

    Maximum likelihood estimates of the proportion of patients cured by cancer therapy , author=. J. R. Stat. Soc., Ser. , volume=. 1949 , publisher=

    1949

  93. [101]

    Mixture cure models in credit scoring: If and when borrowers default , author=. Eur. J. Oper. Res. , volume=. 2012 , publisher=

    2012

  94. [102]

    Operations Research , volume=

    Survival analysis methods for personal loan data , author=. Operations Research , volume=. 2002 , publisher=

    2002

  95. [103]

    Time to default in credit scoring using survival analysis:

    Dirick, Lore and Claeskens, Gerda and Baesens, Bart , journal=. Time to default in credit scoring using survival analysis:. 2017 , publisher=

    2017

  96. [104]

    Know Where to Invest:

    Wang, Zhao and Jiang, Cuiqing and Zhao, Huimin , journal=. Know Where to Invest:. 2022 , publisher=

    2022

  97. [105]

    Cai, Chao and Zou, Yubo and Peng, Yingwei and Zhang, Jiajia , journal=. smcure:. 2012 , publisher=

    2012

  98. [106]

    Biometrics , volume=

    The use of mixture models for the analysis of survival data with long-term survivors , author=. Biometrics , volume=. 1982 , publisher=

    1982

  99. [107]

    Understanding the implications of FinTech Peer-to-Peer (

    Najaf, Khakan and Subramaniam, Ravichandran K and Atayah, Osama F , journal=. Understanding the implications of FinTech Peer-to-Peer (. 2022 , publisher=

    2022

  100. [108]

    Marketplace lending: A new banking paradigm? , author=. Rev. Financial Stud. , volume=. 2019 , publisher=

    2019

  101. [109]

    Do fintech mortgage lenders fill the credit gap? evidence from natural disasters , author=. J. Financ. Quant. Anal. , volume=. 2023 , publisher=

    2023

  102. [110]

    Discrimination in mortgage lending , author=. Am. Econ. Rev. , volume=. 1978 , publisher=

    1978

  103. [111]

    Racial disparities in the auto loan market , author=. Rev. Financial Stud. , volume=. 2023 , publisher=

    2023

  104. [112]

    Discrimination in the small-business credit market , author=. Rev. Econ. Stat. , volume=. 2003 , publisher=

    2003

  105. [113]

    What drives racial and ethnic differences in high-cost mortgages?

    Bayer, Patrick and Ferreira, Fernando and Ross, Stephen L , journal=. What drives racial and ethnic differences in high-cost mortgages?. 2018 , publisher=

    2018

  106. [114]

    The Janus effect of generative

    Susarla, Anjana and Gopal, Ram and Thatcher, Jason Bennett and Sarker, Suprateek , journal=. The Janus effect of generative. 2023 , publisher=

    2023

  107. [115]

    The promise and peril of generative

    Jo, A , journal=. The promise and peril of generative

  108. [116]

    Regression models and life-tables , author=. J. R. Stat. Soc. Ser. B Stat. Method. , volume=. 1972 , publisher=

    1972

  109. [117]

    Attention is all you need , author=. Adv. Neural Inf. Process. Syst. , volume=

  110. [118]

    A Systematic Survey of Prompt Engineering in Large Language Models:

    Sahoo, Pranab and Singh, Ayush Kumar and Saha, Sriparna and Jain, Vinija and Mondal, Samrat and Chadha, Aman , journal=. A Systematic Survey of Prompt Engineering in Large Language Models:

  111. [119]

    Tree of thoughts:

    Yao, Shunyu and Yu, Dian and Zhao, Jeffrey and Shafran, Izhak and Griffiths, Tom and Cao, Yuan and Narasimhan, Karthik , journal=. Tree of thoughts:

  112. [120]

    Nature , volume=

    Role play with large language models , author=. Nature , volume=. 2023 , publisher=

    2023

  113. [121]

    Advances in Neural Information Processing Systems , volume=

    Large language models are zero-shot reasoners , author=. Advances in Neural Information Processing Systems , volume=

  114. [122]

    ACM Transactions on Intelligent Systems and Technology , volume=

    A survey on evaluation of large language models , author=. ACM Transactions on Intelligent Systems and Technology , volume=. 2024 , publisher=

    2024

  115. [123]

    arXiv preprint arXiv:2405.00622 , year=

    Causal Evaluation of Language Models , author=. arXiv preprint arXiv:2405.00622 , year=

    arXiv

  116. [124]

    International Journal of Forecasting , volume=

    Forecasting loss given default for peer-to-peer loans via heterogeneous stacking ensemble approach , author=. International Journal of Forecasting , volume=

  117. [125]

    Wiley International Encyclopedia of Marketing , year=

    Social desirability bias , author=. Wiley International Encyclopedia of Marketing , year=

  118. [126]

    International Conference on Machine Learning , pages=

    Using large language models to simulate multiple humans and replicate human subject studies , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  119. [127]

    Review of Economic Studies , volume=

    Econometric evaluation of linear macro-economic models , author=. Review of Economic Studies , volume=. 1986 , publisher=

    1986

  120. [128]

    Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina , journal=. Bert:

  121. [129]

    1 + 1 > 2?

    Lu, Tian and Zhang, Yingjie , journal=. 1 + 1 > 2?. 2024 , volume=

    2024

  122. [130]

    The Impact of Credit Scoring on Consumer Lending , author=. RAND J. Econ. , year=

  123. [131]

    Quarterly Journal of Economics , year=

    Using Maimonides' rule to estimate the effect of class size on scholastic achievement , author=. Quarterly Journal of Economics , year=

  124. [132]

    Journal of Marketing Research , year=

    Personalization in Email Marketing: The Role of Non-Informative Advertising Content , author=. Journal of Marketing Research , year=

  125. [133]

    and Peterson, Eric D

    Guller, Ulrich and Hervey, Stephen and Purves, Harry and Muhlbaier, Lawrence H. and Peterson, Eric D. and Eubanks, Steve and Pietrobon, Ricardo , journal=. Laparoscopic versus open appendectomy:. 2004 , volume=

    2004

  126. [134]

    MIS Quarterly , volume=

    A tree-based approach for addressing self-selection in impact studies with big data , author=. MIS Quarterly , volume=. 2016 , publisher=

    2016

  127. [135]

    Quarterly Journal of Economics , volume=

    Human decisions and machine predictions , author=. Quarterly Journal of Economics , volume=. 2018 , publisher=

    2018

  128. [136]

    Moris and Yang, Zonghao , year=

    Gopal, Ram D and Qiao, Xiao and Strub, S. Moris and Yang, Zonghao , year=. Doing Well by Doing Good:

  129. [137]

    Information processing at successive stages of decision making:

    Levin, Irwin P and Huneke, Mary E and Jasper, John D , journal=. Information processing at successive stages of decision making:. 2000 , publisher=

    2000

  130. [138]

    Organizational Behavior and Human Decision Processes , volume=

    The role of epistemic motivation in individuals’ response to decision complexity , author=. Organizational Behavior and Human Decision Processes , volume=. 2013 , publisher=

    2013

  131. [139]

    The Twelfth International Conference on Learning Representations , year=

    Large Language Models as Optimizers , author=. The Twelfth International Conference on Learning Representations , year=

  132. [140]

    Bank Run, Interrupted: Modeling Deposit Withdrawals with Generative

    Kazinnik, Sophia , year=. Bank Run, Interrupted: Modeling Deposit Withdrawals with Generative

  133. [141]

    Nature Medicine , volume=

    Large language models in medicine , author=. Nature Medicine , volume=. 2023 , publisher=

    2023

  134. [142]

    Learning and Individual Differences , volume=

    Kasneci, Enkelejda and Se. Learning and Individual Differences , volume=. 2023 , publisher=

    2023

  135. [143]

    Marketing Science , year=

    Frontiers: Determining the Validity of Large Language Models for Automated Perceptual Analysis , author=. Marketing Science , year=

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.