How Costly Are Cultural Biases: D'Acunto, Ghosh & Rossi (2026)

What this is. The paper’s core results, the competing hypotheses it tests, and the regression specifications with their defining equations: enough to know what it found and how, without reading all 26 pages. To replicate or extend it, read the full source at the original.

TL;DR

Using lender-level panel data from Faircent, a peer-to-peer lending platform in India, D’Acunto, Ghosh, and Rossi compare the choices the same lenders make when unassisted and after observing suggestions from an automated robo-advising tool (Auto Invest). Unassisted Hindu lenders are 5.8 percentage points less likely to fund Muslim borrowers than Muslim lenders are, and lenders of all castes systematically under-lend to Shudra (lower-caste) borrowers relative to the platform population. These biases are costly: the disfavored borrowers default less and earn higher standardized returns on average. In the Auto Invest period, both biases are smaller and average lender-level total returns are 4.5 pp to 7.3 pp higher; the authors do not identify a general causal effect of adoption. Lenders rarely override robo-advised suggestions to previously-disfavored groups, which supports inaccurate statistical discrimination (biased beliefs) rather than taste-based discrimination, as in Becker (1957), as the dominant mechanism.

Core results

Magnitudes and significance are as reported; \*/\*\*/\*\*\* = 10%/5%/1%. Locators point into the source PDF.

#	Result	Locator	Magnitude
R1	Hindu lenders are 5.8 pp less likely to fund Muslim borrowers than Muslim lenders before robo-advising	Table 2, col 1, p.11	Hindu Lender = -0.058***, t=-3.52; baseline Muslim-borrower share is 18% for Muslim lenders and 12% for Hindu lenders
R2	The Hindu-Muslim lending gap is 4.5 pp smaller in the Auto Invest period	Table 2, col 1, p.11	Hindu Lender x Auto Invest = 0.045**, t=2.51; the Hindu-Muslim lending gap narrows by 4.5 pp but does not disappear
R3	Discrimination is costly: Muslim borrowers in Hindu lender portfolios default 2.4 pp less than Hindu borrowers	Table 3, col 1, p.17	Muslim Borrower = -0.024**, t=-2.02; in-group bias selects worse borrowers from the preferred group
R4	In the Auto Invest period, the estimated default reductions are 11.2 pp for Hindu borrowers and 7.3 pp for Muslim borrowers; the in-group estimate is larger	Table 3, col 2, p.17	Hindu Borrower x Auto Invest = -0.112***, t=-5.21; Muslim Borrower x Auto Invest = -0.073**, t=-2.49
R5	Before robo-advising, Muslim borrowers deliver higher standardized returns for Hindu lenders	Table 5, col 1, p.21	Muslim Borrower = 0.282***, t=6.25 (standardized return); confirms the financial cost of out-group bias
R6	In the Auto Invest period, Hindu borrowers’ returns are higher by 0.222 SD; the Muslim-borrower estimate is near zero and insignificant	Table 5, col 2, p.21	Hindu Borrower x Auto Invest = 0.222***, t=3.07; Muslim Borrower x Auto Invest = -0.012, t=-0.18
R7	Average lender-level total returns are 4.5 pp (in-group vs. out-group) and 7.3 pp (stereotypical discrimination) higher in the Auto Invest period	Fig. 11, p.23	Mean within-lender change in total returns, with loan returns value-weighted within lender and lenders equally weighted
R8	Shudra borrowers default 3.8 pp less than other borrowers before Auto Invest, confirming stereotypical discrimination is costly	Table 3, col 4, p.17	Shudra Borrower = -0.038***, t=-3.03; all-lender sample; lending to Shudra borrowers is higher in the Auto Invest period
R9	The out-group de-biasing estimate remains positive with lender and year fixed effects	Table 2, cols. 3-4, p.11	Hindu Lender x Auto Invest = 0.045** (t=2.07) with lender FE; 0.043** (t=1.96) with lender and year FE
R10	The de-biasing estimate is marginally significant for high-use lenders and near zero for low-use lenders	Table 2, cols. 5-6, p.11	Hindu Lender x Auto Invest = 0.009 (t=0.23) for <40% use and 0.052* (t=1.94) for >=40% use
R11	Out-group bias is larger in high-riot states; the de-biasing estimate is positive there, but is not significantly different across state groups	Fig. 4, p.13	Bias = 6.4 pp and de-biasing = 4.8 pp (p<0.01) in high-riot states; low-riot estimate small and insignificant; cross-group Wald chi-square = 0.18 (difference not significant)
R12	Out-group bias is larger where BJP vote share is high, and de-biasing differs across state groups	Fig. 4, p.13	Pre-adoption bias = 9.4 pp in high-vote-share states vs. 3.5 pp elsewhere; chi-square = 10.65 (p<0.01)
R13	The out-group bias is concentrated among lenders born after 1990	Fig. 4, p.13	Bias = 7.1 pp for post-1990 lenders and small, insignificant for older cohorts; chi-square for difference = 4.43 (p<0.01)
R14	Lending to Shudra borrowers is higher in the Auto Invest period	Fig. 6, p.15	Auto Invest coefficient = 0.011 (t=2.54), a 1.1 pp increase
R15	The Shudra lending response is strongest where caste is recognizable and local anti-caste crime is high	Fig. 6, p.15	For >70% caste recognizability, increase = 2.7 pp in high-crime states, less than half as large elsewhere
R16	The default gaps and differential default improvement disappear conditional on loan-risk controls	Table 3, col. 3, p.17	Muslim Borrower = -0.002 (t=-0.10); Hindu Borrower x Auto Invest = -0.072*** (t=-4.55); Muslim Borrower x Auto Invest = -0.070*** (t=-2.82)
R17	The authors find no systematic average change in lender portfolio diversification after Auto Invest	Text p.18; Table A.3, Online Appendix	No systematic change across the paper’s three diversification proxies: loan-size dispersion, loan-return dispersion, and payment dispersion
R18	Uncontrolled return gains are largest in the lower Hindu-borrower quantiles; with loan-risk controls the quantile estimates are insignificant, while the OLS estimate turns negative	Table 6, Panel A, p.21	Without risk controls: OLS 0.217*** (t=4.92), 25th percentile 0.250* (t=1.82), median 0.274*** (t=11.22), 75th percentile 0.109*** (t=11.66); with controls, OLS -0.160*** (t=-6.15), quantile estimates insignificant
R19	Purged lender-level value gains from removing cultural bias are 6% to 12% of pre-adoption capital	Text p.24	Rs.457 (6% of Rs.7,543) for in-group/out-group discrimination; Rs.862 (12% of Rs.7,091) for stereotypical discrimination
R20	Lenders rarely override robo-advised matches to groups they previously discriminated against, supporting biased beliefs over taste-based discrimination	Text pp.2, 24	The paper describes the vast majority as not overriding; no percentage is reported
R21	Raw default rates are lower after lenders use Auto Invest	Text p.8	Lender-level default rates: 33% before vs. 16% after adoption (p-value <0.1%); unadjusted descriptive comparison

Overall (paper’s conclusion). Cultural biases lead lenders to select borrowers from preferred social groups who systematically underperform. The evidence is most consistent with inaccurate statistical discrimination: biased ex-ante beliefs about borrower quality correlated with ethnicity and caste, not a conscious taste for discrimination. Auto Invest use is associated with smaller estimates of both types of bias and higher returns. The authors estimate purged value gains of approximately 6% to 12% of average pre-adoption capital, concentrated in loans to previously favored in-group borrowers who were over-selected relative to their quality.

Theory / model

The paper has no formal economic model. It tests three competing hypotheses about the nature of the discrimination it documents, each with distinct predictions for lenders’ behavior when assisted vs. unassisted:

Hypothesis 1 (taste-based discrimination, Becker (1957)): lenders are willing to pay a utility cost to avoid transacting with disfavored groups. Prediction: lenders should frequently override robo-advised matches to disfavored-group borrowers, even when their economic incentives are aligned with the tool.

Hypothesis 2 (accurate statistical discrimination): disfavored groups are truly riskier on average, so the favored-group lending pattern reflects correct Bayesian updating on observable signals. Prediction: in-group borrowers should outperform (lower default, higher returns) relative to out-group borrowers matched by the platform’s unbiased screening.

Hypothesis 3 (inaccurate statistical discrimination, biased beliefs): lenders hold systematically incorrect ex-ante beliefs about out-group borrower quality, even when identical objective risk information is available. Predictions: (a) disfavored borrowers should outperform when lenders are unassisted; (b) performance should converge after robo-advising; (c) lenders should not override robo-advice, because their financial incentives do not justify doing so.

The platform design eliminates channels that could rationalize statistical discrimination: interest rates are set by the platform algorithm, borrower risk profiles are directly observable, lenders and borrowers never interact, and Faircent’s screening ensures no unbanked borrowers enter the pool. Heterogeneity tests rule out monitoring advantages, social collateral, and peer effects as alternative explanations (Fisman et al. (2017), Fisman et al. (2020)). By contrast, Hjort (2014) detects strong taste-based discrimination in a setting that removes scope for inaccurate statistical discrimination; the paper argues that Faircent’s FinTech context is closer in spirit to that ideal but that inaccurate beliefs nonetheless dominate. The evidence consistently supports Hypothesis 3: disfavored borrowers outperform before robo-advising, performance converges after, and lenders rarely override robo-advised matches. The paper further shows that biases are stronger for lenders in areas with higher Hindu-Muslim inter-ethnic conflict, consistent with culturally shaped priors rather than rational updating.

Method

The paper estimates OLS regressions on lender-borrower-loan triads, generally including lender and year fixed effects and clustering standard errors by lender. Table 2 varies the controls and fixed effects across columns; the diversification regressions cluster by city, and the quantile regressions omit fixed effects. Adoption is not randomized, and the authors state that they do not estimate a general causal effect of adoption (text p.7). The approach builds on D’Acunto and Rossi (2020), who study robo-advising effects in a savings context.

For loan returns the paper additionally estimates quantile regressions (Eq. 5, p.22) to identify which part of the return distribution drives the improvement:

$$Q_\tau(\text{Returns}_{i,j}) = a(\tau) + \beta(\tau)\,\text{Auto Invest}_j + \mathbf{X}'_{i,j}\,\zeta(\tau) + \varepsilon_{i,j} \tag{5}$$

where $Q_\tau$ is the $\tau$-th quantile of standardized loan returns for loan $i$ of lender $j$, and $\mathbf{X}_{i,j}$ are loan risk controls. Coefficient $\hat{\beta}(\tau)$ measures how the $\tau$-th quantile shifts after the lender adopts Auto Invest. Table 6 reports quantile regressions without lender or year fixed effects, with standard errors clustered by lender; the Hindu/Muslim samples have 2,326/220 loan observations, and the low-/high-recognizability Shudra samples have 462/1,158 observations (Table 6, p.21).

The lender-level total return is computed as a value-weighted average across loans originated before and after Auto Invest adoption (Eq. 6, p.23):

$$\text{Lender Tot Ret}_{i,t} = 100 \times \frac{\sum_j \text{Amount Disbursed}_{i,j,t} \times \text{Loan Return}_{j,t}}{\sum_j \text{Amount Disbursed}_{i,j,t}} \tag{6}$$

The paper then calculates each lender’s post-minus-pre total-return change (Eq. 7, p.23):

$$\text{Change Lender Return}_i = \text{Lender Tot Ret}_{i,\text{POST}} - \text{Lender Tot Ret}_{i,\text{PRE}} \tag{7}$$

For the rupee-value measure, Eq. 8 (p.23) holds pre-adoption disbursements fixed and applies post-adoption returns, purging the change that would arise only from different investment amounts:

$$\begin{aligned} \text{Change Lender Value}_i ={}& [\text{Amt Disb}_{i,\text{POST}} \times \text{Ret}_{i,\text{POST}} - \text{Amt Disb}_{i,\text{PRE}} \times \text{Ret}_{i,\text{PRE}}] \\ &- [\text{Amt Disb}_{i,\text{POST}} \times \text{Ret}_{i,\text{POST}} - \text{Amt Disb}_{i,\text{PRE}} \times \text{Ret}_{i,\text{POST}}] \tag{8} \end{aligned}$$

This is equivalent to pre-period amount disbursed times the change in lender return, and underlies the rupee amounts in R19.

Empirical specifications

Primary specification for in-group vs. out-group discrimination (Eq. 1, p.10):

$$\text{Muslim Borrower}_{i,j,t} = \alpha + \beta\,\text{Auto Invest}_{j,t} + \gamma\,\text{Hindu Lender}_j + \delta\,(\text{Hindu Lender}_j \times \text{Auto Invest}_{j,t}) + \zeta\,\mathbf{x}_{i,t} + \eta_j + \eta_t + \varepsilon_{i,j,t} \tag{1}$$

where $\text{Muslim Borrower}_{i,j,t} = 1$ if borrower $i$ funded by lender $j$ in year $t$ is Muslim; $\text{Auto Invest}_{j,t} = 1$ if the lender has adopted the tool by year $t$; $\text{Hindu Lender}_j = 1$ if lender $j$ is Hindu; $\mathbf{x}_{i,t}$ are loan characteristics assigned by the platform (maturity, amount, interest rate); $\eta_j$ are lender fixed effects; $\eta_t$ are year fixed effects. Coefficient $\hat{\gamma}$ (R1) measures the in-group bias before adoption; $\hat{\delta}$ (R2) measures the de-biasing effect of robo-advising.

The table reports OLS on the adopting-lender sample, with standard errors clustered by lender. Column 1 has no loan controls or fixed effects (113,284 observations); columns 2-4 add loan controls, lender fixed effects, and then year fixed effects in sequence (113,283 observations); columns 5-6 split by below/above 40% Auto Invest allocation (39,366 and 72,104 observations, respectively) (Table 2, p.11). The within-lender and within-year estimates are R9; the usage-intensity comparison is R10.

Stereotypical discrimination specification (Eq. 2, p.14):

$$\text{Shudra Borrower}_{i,j,t} = \alpha + \beta\,\text{Auto Invest}_{j,t} + \zeta\,\mathbf{x}_{i,t} + \eta_j + \eta_t + \varepsilon_{i,j,t} \tag{2}$$

This drops the Hindu Lender interaction because all castes, including Shudra lenders, discriminate against Shudra borrowers (stereotypical discrimination is not in-group favoritism but group-wide negative stereotyping). Caste recognizability (continuous probability from the Bhagavatula et al. (2017, 2018) matrimonial-registry algorithm) is used to test whether the $\hat{\beta}$ coefficient grows with how easily a borrower is identifiable as Shudra (Fig. 5-6, pp.14-15). The specification uses loan characteristics, lender and year fixed effects, and lender-clustered standard errors; Fig. 6 reports the overall estimate and subsamples by caste recognizability and state-level crimes against lower castes (Fig. 6, p.15).

Performance specifications (Eqs. 3-4, p.16 and p.19):

$$\text{Delinquent Loan}_{i,j,t} = \alpha + \gamma\,\text{Muslim Borrower}_{i,j} + \delta\,(\text{Muslim Borrower}_{i,j} \times \text{Auto Invest}_{j,t}) + \theta\,(\text{Hindu Borrower}_{i,j} \times \text{Auto Invest}_{j,t}) + \zeta\,\mathbf{x}_{i,t} + \eta_j + \eta_t + \varepsilon_{i,j,t} \tag{3}$$ $$\text{Loan Return}_{i,j,t} = \alpha + \gamma\,\text{Muslim Borrower}_{i,j} + \delta\,(\text{Muslim Borrower}_{i,j} \times \text{Auto Invest}_{j,t}) + \theta\,(\text{Hindu Borrower}_{i,j} \times \text{Auto Invest}_{j,t}) + \zeta\,\mathbf{x}_{i,t} + \eta_j + \eta_t + \varepsilon_{i,j,t} \tag{4}$$

where $\text{Delinquent Loan}_{i,j,t} = 1$ if the loan is closed delinquent (more than 90 days past due at closure) and $\text{Loan Return}_{i,j,t}$ is the standardized return. Coefficient $\hat{\gamma}$ tests whether disfavored borrowers outperformed before the tool was adopted (the cost of bias, R3 and R5); $\hat{\delta}$ and $\hat{\theta}$ capture the differential change in performance across borrower groups after adoption (R4 and R6). The default regressions include lender and year fixed effects and cluster standard errors at lender level. Table 3 uses 16,985 Hindu-lender triads for the religion tests and 6,821 caste-subsample triads; columns 3 and 6 add loan-risk controls (Table 3, p.17). The return regressions add lender portfolio risk, use lender and year fixed effects, and cluster by lender; the Table 5 sample has 2,134 observations for Hindu lenders and 859 for the caste analysis (Table 5, p.21). In the default falsification column, the Muslim/Hindu default gaps and differential improvement vanish conditional on loan risk (R16).

Return quantiles (Eq. 5, p.22). The controlled quantile regressions add loan amount, maturity, and annual interest rate; no lender or year fixed effects are included. Standard errors are clustered at lender level. Without risk controls, Hindu-borrower return gains are largest in the lower quantiles. With controls, the quantile estimates are insignificant, while the OLS estimate is negative and significant (Table 6, Panel A, p.21; R18).

Heterogeneity tests use cross-sectional proxies for the salience of cultural stereotypes: state-level Hindu-Muslim riots (Ticku (2015)), BJP vote shares (Bhavnani (2014)), and birth-cohort exposure to the rise of Hindu-Muslim conflict (Fig. 4, p.13). The bias is about twice as large for lenders residing in high-riot states (6.4 pp vs. small and insignificant elsewhere), and the de-biasing estimate is positive in high-riot states, but the estimated difference across state groups is not statistically significant (Wald chi-square = 0.18). The BJP-state comparison is 9.4 pp vs. 3.5 pp, and post-1990 lenders show a 7.1 pp bias while older-cohort estimates are small and insignificant (Fig. 4, p.13; R11-R13). The paper reports no randomized Auto Invest adoption and states that it does not estimate a general causal effect of adoption (text p.7).

Portfolio-diversification specification (Table 4, p.19). The paper estimates lender-level OLS regressions for the pre/post change in three portfolio-dispersion measures. For the in-group comparison, the estimating specification is:

$$\text{Diversification Portfolio}_{j,t} = \alpha + \gamma\,\text{Share Muslim Borrowers}_{j,t} + \delta\,(\text{Share Muslim Borrowers}_{j,t} \times \text{Auto Invest}_{j,t}) + \theta\,(\text{Share Hindu Borrowers}_{j,t} \times \text{Auto Invest}_{j,t}) + \varepsilon_{j,t}$$

There are no fixed effects; standard errors are clustered at the city level. The six columns use 2,586/399/2,707 observations for the in-group loan-size, loan-return, and payment-dispersion outcomes and 998/292/979 observations for the corresponding caste outcomes (Table 4, p.19). The paper reports no systematic average change in portfolio diversification across the three proxies (text p.18; Table A.3, Online Appendix; R17). The conclusion also notes that lenders rarely override Auto Invest matches to groups previously avoided, supporting inaccurate beliefs over taste-based preferences, but reports no override percentage (text pp.2, 24; R20).

Sample: lender-borrower-loan triads from the Faircent platform, January 2018 to March 2020. Main sample: 113,283 triads involving 2,818 unique Hindu and Muslim lenders and borrowers. Caste sub-sample: 62,831 triads for which Hindu varna (caste category) of the borrower can be inferred from the matrimonial registry. Loan maturity averages 22 months; median maturity 24 months; average loan amount is approximately Rs.130,000 (~$1,770); average annual interest rate is 24%. Standard errors are clustered at the lender level.

Datasets used

Dataset	Role in paper	Wiki page
Faircent P2P lending platform	Main panel: lender-borrower-loan triads; loan amounts, interest rates, maturity, delinquency; Auto Invest adoption and fund-allocation share; lender and borrower demographic characteristics (name, state, date of birth, occupation)	No page yet
Marriage registry (Bhagavatula et al. 2017, 2018)	Religion and caste inference for lenders and borrowers; 2.5 million individuals from online matrimonial agencies; used to assign religion and varna probabilities from surname, date of birth, and location	No page yet
Ticku (2015) Hindu-Muslim riots data	State-level count of large-scale riots between Hindus and Muslims (1980-2000); proxy for inter-ethnic conflict salience in heterogeneity tests	No page yet
Bhavnani (2014) BJP election data	Average BJP candidate vote shares across national and state elections (1977-2015) per Indian state; proxy for ideological salience of Hindu nationalism	No page yet
National Crime Records Bureau (NCRB 2019)	Crimes against Scheduled Castes per 100,000 inhabitants per Indian state (2018); proxy for salience of caste discrimination in stereotypical bias heterogeneity tests	No page yet

Sample: January 2018 to March 2020. Roughly 60% of loans issued in 2019 and 19% in the first three months of 2020. Median lender disburses funds to borrowers across 13 different Indian states; 90% of lenders serve borrowers in at least 5 different states.

When to read the full paper

Use the original if you are: studying the mechanism of discrimination (taste-based vs. inaccurate statistical) in financial markets; examining how robo-advising corrects culturally biased lending choices; studying in-group vs. out-group or stereotypical discrimination in a high-stakes economic setting; or extending the Faircent platform setting to other FinTech platforms or demographic groups. Table 2 (p.10) gives the main lending-bias results; Table 3 (p.17) the default-performance results; Tables 5-6 (pp.21-22) the return results; and Figs. 4, 6, and 8 the heterogeneity and mechanism tests.

Attribution and rights

Source: peer-reviewed, Journal of Financial Economics 175 (2026), article 104202. This distillation was extracted by an LLM on 2026-06-24 and is not human-verified or independently reproduced. The CC BY 4.0 licence permits mirroring; the verbatim PDF is not hosted in this batch.

Attribution (CC BY 4.0). D’Acunto, Francesco, Pulak Ghosh, and Alberto G. Rossi. “How costly are cultural biases? Evidence from FinTech.” Journal of Financial Economics 175 (2026) 104202. DOI: 10.1016/j.jfineco.2025.104202. © 2025 The Authors. Published by Elsevier B.V. Licensed under CC BY 4.0. This page is an adaptation by the Institute for Automated Research: core results extracted and re-expressed; changes were made.