Teams and Belief Overreaction: Barahona, Cassella, Jansen & Pezone (2026)

What this is. The paper’s core results, the measurement framework for belief overreaction, the decomposition of the team effect into three channels, and the key estimating equations: enough to understand what was found and how, without reading the full 17-page article. To replicate or extend it, read the original at https://doi.org/10.1016/j.jfineco.2025.104219.

TL;DR

The paper addresses a fundamental question in behavioral finance: does moving from individual to team decision-making amplify or attenuate belief overreaction to recent asset returns? Using preregistered randomized experiments on the Labvanced platform with 1,512 Prolific participants, plus a within-subject field study of US equity mutual fund managers (1980-2018), the paper finds that two-person teams reduce individual overreaction by 30 to 55 percent. A quantitative decomposition, following the approach of Enke et al. (2023), partitions the lab team effect into three channels: internal reflection (the act of pre-team deliberation), self-selection (the tendency of the less-biased member to lead), and external screening (the group interaction itself). Self-selection accounts for roughly 70 percent of the reduction. LLM analysis of roughly 18,000 chat exchanges in the Group treatment corroborates this, and dynamic evidence shows that participants reduce their leadership role after making larger forecast errors. In the field, where overreaction is measured indirectly from funds’ sensitivity to past stock returns, mutual fund teams attenuate extrapolative overreaction by about 55 percent relative to the individual behavior of the same managers. This attenuation is associated with better investment performance.

Core results

Magnitudes and significance are as reported; \* / \*\* / \*\*\* = 10% / 5% / 1%. Standard errors in brackets, clustered at the team level (equivalent to individual-level for the Individual treatment). Locators reference the source PDF.

#	Result	Locator	Magnitude
R1	Teams reduce overreaction by ~30%: Group treatment dummy lowers the individual overreaction coefficient by 0.092, significant at 1%, robust across demographics, financial sophistication, and income fixed-effects controls	Table 2, p. 6	Group = -0.092*** [0.035] (Col 1, n = 704); stable at -0.099*** to -0.101*** in Cols 2-4; Individual mean beta = 0.311, SD = 0.424
R2	Self-selection explains ~70% of the team effect: the three-channel decomposition shows internal reflection contributes essentially nothing (-0.001, n.s.) and external screening contributes -0.025 to -0.031 (n.s.), while self-selection contributes -0.067 to -0.090 (significant in most specifications)	Table 3/Eq. (5), p. 8	Most conservative spec (Col 4): IR = -0.001 (n.s.); SS = -0.068* [0.039]; ES = -0.031 (n.s.); total team effect = -0.098 [0.035]
R3	Teams show a weaker recency effect: fitting the exponential return-extrapolation model separately for each treatment, the recency weight (1 minus the decay parameter) is about 2.4x smaller for Group participants than Individual participants	Fig. 3/Eq. (7), p. 11	Group lambda2 = 0.953 (recency = 0.047) vs Individual lambda2 = 0.886 (recency = 0.114); Group places more equal weight across all 40 past returns
R4	Self-selection is dynamic and driven by past errors: participants who led the team prediction in round t-1 are significantly less likely to lead in round t after the team made a large forecast error in round t-1	Table 4/Eq. (6), p. 10	MostVotes_{t-1} x |Error_{t-1}| coefficient = -4.440*** [0.645] (Votes, Col 1); -0.137*** [0.018] (MostVotes dummy, Col 3); n = 7,657 round observations
R5	Teams achieve higher prediction accuracy: Group participants generate lower mean-squared error, lower mean-absolute error, and earn a higher experimental bonus than Individual participants	Table 5, p. 11	Group vs Individual: MSE -157.722*** [30.394]; MAE -2.015*** [0.408]; Bonus +$0.059** [0.023]; full controls, n = 703
R6	Mutual fund teams attenuate extrapolative overreaction by ~55%: within-subject comparison of team overreaction and statistical counterfactual individual overreaction shows teams transmit only about 45% of extrapolative behavior, while contrarian (non-overreacting) behavior is fully transmitted	Table 8/Eq. (10-11), p. 15	IV sum delta0 + delta1 = 0.45 (Col 7); null of full transmission (= 1) rejected at IV p = 0.015 (Col 7) and IV p = 0.018 (Col 8); contrarian-only delta0 not significantly below 1 (IV p = 0.568)
R7	Lower absolute belief bias: the group treatment also reduces absolute sensitivity to past returns, establishing that the main result is closer to the rational benchmark rather than merely a change in coefficient sign	Table 5, p. 11	Group coefficient on absolute overreaction = -0.101*** [0.027]; Individual mean = 0.413
R8	Chat evidence supports self-selection: both LLM measurement approaches indicate that one partner commonly leads while the other withdraws, and their team-level measures are positively associated	Fig. 2, p. 9	Uncontested rounds mean = 16.37 (SD = 3.16); LLM score mean = 6.56 (SD = 0.53); Spearman correlation = 0.293***
R9	No significant average field difference across all teams: the within-team comparison of solo counterfactual and observed team trading sensitivity is not statistically significant in the pooled sample	Table 7, p. 13	Solo counterfactual mean = -0.011 (s.e. 0.022; t = -0.489); team mean = -0.034 (s.e. 0.028; t = -1.200); CF - TM = 0.023 (s.e. 0.026; t = 0.891), N = 308
R10	Contrarian behavior is retained in teams: among contrarian teams, solo and team sensitivities are statistically indistinguishable	Table 7, p. 13	Solo counterfactual mean = -0.224 (s.e. 0.029; t = -7.593); team mean = -0.178 (s.e. 0.043; t = -4.127); CF - TM = -0.046 (s.e. 0.034; t = -1.374), N = 143
R11	Extrapolative behavior is attenuated in teams: among extrapolative teams, mean sensitivity falls by about one-half between solo management and team management	Table 7, p. 13	Solo counterfactual mean = 0.184 (s.e. 0.024; t = 7.815); team mean = 0.090 (s.e. 0.035; t = 2.589); CF - TM = 0.094 (s.e. 0.039; t = 2.414), N = 165
R12	IV evidence does not reject full transmission of contrarian behavior: in the field IV specification, the coefficient on solo counterfactual sensitivity is statistically consistent with one	Table 8, p. 15	Test of H0 delta0 = 1: p = 0.568 (IV Col 5)
R13	Solo and team fund samples have similar observed characteristics: the descriptive statistics support the paper’s decision to use a within-subject design rather than rely on between-sample comparability	Table 6, p. 12	Solo vs team: mean TNA = 1017.82 vs 1200.26 million; mean stocks held = 89.88 vs 103.07 (median 52 vs 54); expense ratio = 1.25 vs 1.26 percentage points
R14	Extrapolative trading predicts weaker fund performance: the paper’s full-sample and recursive sorts show lower performance among extrapolators than contrarians across raw, benchmark-adjusted, CAPM, FF3, and FF5 measures	Fig. 4, p. 16	Panel A labels mean beta as -0.22 for contrarians and 0.15 for extrapolators; exact numerical performance bar values are not printed
R15	No evidence of naive one-period extrapolation: participants are not more likely to predict an identical return in consecutive periods, distinguishing the measured bias from simply repeating the latest return	text p. 11	Null reported in unreported results; no coefficient or test statistic is given

Overall (paper’s conclusion). Both in the lab and in the field, teams reduce belief overreaction relative to individuals. The dominant mechanism is self-selection: in two-person teams, the less-biased member tends to take on decision authority. This process is dynamic (driven by past forecast errors and feedback) and is confirmed by LLM-based analysis of chat exchanges. In the field, the attenuation of extrapolative trading by mutual fund teams is associated with better subsequent fund performance, while contrarian (non-overreacting) behavior is preserved.

Theory / model

The paper does not present a formal structural model. It motivates belief overreaction through representativeness and extrapolation accounts, including Barberis (2018) and Greenwood and Shleifer (2014), and tests whether teams attenuate the tendency to overweight past returns. In the lab task, returns follow the AR(1) process described on p. 4:

$$x_t = \rho x_{t-1} + \varepsilon_t, \qquad \varepsilon_t \sim \mathcal{N}(0, \sigma^2), \qquad \rho = 0.5, \quad \sigma = 20$$

The paper treats zero sensitivity to the latest return as the rational benchmark because past returns have weak predictive power. It predicts an individual coefficient above zero under overreaction and a negative group-treatment effect if teams reduce it (pp. 4-5).

The authors replicate widespread individual overreaction in expectations documented by Bordalo et al. (2020) before testing how teams change it.

The conceptual team effect is the difference between mean overreaction in the Group and Individual treatments. Equation (3) partitions that difference into internal reflection, self-selection, and external screening (p. 6):

$$\Delta\beta_G = \beta_G - \beta_I = (\beta_{IR} - \beta_I) + (\beta_{SS} - \beta_{IR}) + (\beta_G - \beta_{SS}) \tag{3}$$

Internal Reflection measures beliefs after participants know they are part of a team but before team interaction; Self-Selection uses votes to choose which participant’s forecast becomes the team forecast; External Screening is the residual difference between actual discussion and the voting treatment. The decomposition predicts that self-selection lowers team overreaction when less biased members tend to lead. The direction is theoretically ambiguous if the more biased members instead assume leadership (pp. 7-8).

Method

Randomized lab experiment. The Individual and Group arms randomly assign US Prolific participants to forecast alone or in two-person teams. Each participant sees 40 returns from the same hypothetical AR(1) process over 20 rounds. The two additional treatments isolate channels: Internal Reflection adds individual forecasts before discussion, while Self-Selection asks partners to allocate 100 votes across their independent forecasts. The final sample is 1,512 participants: 248 Individual, 456 Group, 405 Internal Reflection, and 403 Self-Selection (pp. 4, 7-8). Participants receive a score based on forecast error, $S_t = 100 \times \max(0, 1 - |FE_t|/\sigma)$, converted to a dollar bonus (p. 4).

The lab task follows Afrouzi et al. (2023), adapting their overreaction elicitation to forecast hypothetical stock returns. The paper uses two measures of lab overreaction. Equation (1) regresses each participant’s forecast on the most recent return; Equation (7) fits an exponentially weighted average of 40 past returns and separates overall sensitivity from recency (pp. 4, 11). For the mechanism evidence, the authors use GPT-4o-mini to code chat exchanges for uncontested first proposals and to generate a separate self-selection score over 100 queries (Fig. 2, p. 9). The Self-Selection treatment panel follows votes and forecast errors over repeated rounds (Table 4, p. 10).

Mutual fund field design. The authors compare 308 teams with the equal-weighted average behavior of the same managers when they worked alone, rather than comparing unrelated solo- and team-managed funds. The underlying panel covers 1980-2018, with 467 managers and 847 funds. The trading response to past returns is estimated at the fund level using stock holdings, returns, controls for stock characteristics and flow-induced trading, and fund-by-quarter fixed effects (Eqs. 8-9, p. 12). The team-level transmission regressions include team controls and, where reported, style fixed effects. To address measurement error in the generated solo counterfactual, the paper instruments it using estimates from disjoint subsamples, following Jegadeesh et al. (2019) (Eqs. 10-11 and Table 8, pp. 13-15).

Empirical specifications

The lab measure regresses participant i’s reported expectation on the latest realized return (Eq. 1, p. 4). Participants see the same 40-return information history in each round, repeated over 20 rounds; this is a participant-level time-series regression, not a treatment regression:

$$\widehat{E}_{t} x_{i,t+1} = \alpha_i + \beta_i x_t + \varepsilon_{i,t} \tag{1}$$

The participant-level treatment regression relates the estimated coefficient to Group assignment and survey controls (Eq. 2, p. 5):

$$\widetilde{\beta}_i = \alpha_2 + \gamma G_i + \delta' X_i + \eta_i \tag{2}$$

Here G_i equals one for Group participants. Table 2 reports 704 observations in columns 1-3 and 703 in column 4. Columns 2-4 add demographic controls, column 3 adds financial-sophistication controls, and column 4 adds US region and income-category fixed effects. Standard errors are clustered at the team level. The estimated Group coefficient ranges from -0.092 to -0.101 (Table 2, p. 6).

The four-treatment specification estimates the overreaction coefficient on treatment indicators and controls (Eq. 4, p. 8):

$$\widetilde{\beta}_i = \alpha_3 + \gamma_{IR} IR_i + \gamma_{SS} SS_i + \gamma_G G_i + \delta' X_i + \epsilon_i \tag{4}$$

It uses all 1,512 participants, with the Individual treatment as the omitted group. The controls are added successively across columns; the final column includes demographics, financial-sophistication controls, and US region and income-category fixed effects. Standard errors are clustered at the team level. Equation (5) reports the decomposition from the most controlled column (Table 3, p. 8):

$$\underbrace{-0.098}_{\Delta\widehat{\beta}_G\;[\text{s.e.}=0.035]} = \underbrace{-0.001}_{\Delta\widehat{\beta}_{IR}\;[\text{s.e.}=0.034]} + \underbrace{-0.067}_{\Delta\widehat{\beta}_{SS}\;[\text{s.e.}=0.034]} + \underbrace{-0.031}_{\Delta\widehat{\beta}_{ES}\;[\text{s.e.}=0.035]} \tag{5}$$

Equation (6) tests whether the previous round’s forecast error changes the next-round leadership/vote outcome conditional on who led in the previous round (p. 9):

$$Y_{i,t} = \alpha + \beta \text{MostVotes}_{i,t-1} + \gamma |\text{Error}_{i,t-1}| + \delta \text{MostVotes}_{i,t-1} \times |\text{Error}_{i,t-1}| + \varepsilon_t \tag{6}$$

The outcome is either votes cast or an indicator for being the decision maker. The absolute forecast error is demeaned and standardized. Table 4 uses 7,657 round-participant observations, team-clustered standard errors, and demographic and financial-sophistication controls plus region and income fixed effects in columns 2 and 4; the controls also interact with MostVotes. The interaction estimates are -4.440 [0.645] for vote counts and -0.137 [0.018] for decision-maker status (Table 4, p. 10).

The alternative belief-formation model weights past returns exponentially (Eq. 7, p. 11):

$$\widehat{E}_{t} x_{i,t+1} = \lambda_0 + \lambda_1 \frac{\sum_{j=0}^{N} \lambda_2^j x_{t-j}}{\sum_{i=0}^{N} \lambda_2^i} + \varepsilon_{i,t} \tag{7}$$

Here N is set to 39, corresponding to 40 past returns. The model is estimated separately by treatment; the Appendix reports nonlinear least squares. Figure 3 reports $\lambda_2 = 0.886$ for Individual and 0.953 for Group. The main text does not give standard-error or fixed-effect details for these treatment-specific estimates (Fig. 3, p. 11).

For fund j and stock s, the field measure is estimated from changes in holdings on past stock returns, controls, and fund-time fixed effects (Eqs. 8-9, p. 12):

$$\text{trades}_{s,j,t+1} = \alpha_j + \beta_j^X r_{s,t-4\to t} + \gamma_j' C_{s,t} + \theta_{j,t} + e_{s,j,t+1}, \qquad j = 1, \ldots, J \tag{8}$$ $$\text{trades}_{s,j,t+1} \equiv \frac{(\text{shares}_{s,j,t+1} - \text{shares}^{\text{split-adj}}_{s,j,t}) P_{s,t+1}}{TNA_{j,t+1}} \tag{9}$$

The past-return regressor is the weighted sum of four quarterly returns, with weights proportional to $\lambda^j$ and $\lambda = 0.56$. The controls include stock characteristics and flow-induced trading controls; $\theta_{j,t}$ is a fund-quarter fixed effect. The panel uses the stock holdings history from the 1980-2018 field sample. The paper refers to Internet Appendix IA5 for the full control list and investment universe; the main text does not specify a standard-error treatment for this fund-level estimation (Eq. 8, p. 12).

The team-level regression compares each observed team’s sensitivity with its solo-manager counterfactual (Eq. 10, p. 13):

$$\widehat{\beta}_j^{TM} = \alpha + \widehat{\beta}_j^{CF} (\delta_0 + \delta_1 D_j^E) + \delta_2 D_j^E + \delta_3 C_j + \epsilon_j \tag{10}$$

Here D_j^E marks extrapolative teams and C_j contains team controls. Table 8 estimates this cross-section for 308 teams (307 in controlled columns), with team controls and style fixed effects included as indicated and standard errors reported in brackets; the main text does not state a clustering level. The IV columns use disjoint-subsample estimates to instrument the generated counterfactual. The simpler model reported in Table 8 columns 1, 2, 5, and 6 is (Eq. 11, p. 14):

$$\widehat{\beta}_j^{TM} = \alpha + \delta_0 \widehat{\beta}_j^{CF} + \delta_1 C_j + \epsilon_j \tag{11}$$

In the full IV model, the estimate of $\delta_0 + \delta_1$ is 0.4536 in column 7 and 0.4997 in column 8; the null of full extrapolative transmission is rejected at p = 0.015 and p = 0.018, respectively. The null that contrarian behavior is fully transmitted, $\delta_0 = 1$, is not rejected in IV column 5 (p = 0.568; Table 8, p. 15).

Datasets used

Dataset	Role in paper	Wiki page
Lab experiment data (Labvanced / Prolific, 2024)	Primary data for Sections 2-2.5: 1,512 participants, four treatments, 20 prediction rounds each; chat transcripts analyzed by LLM	No page yet (original data; replication package on Mendeley Data)
CRSP monthly stock returns (via WRDS)	Quarterly stock-level returns for the fund trading regression (Eq. 8); past four-quarter return predictor	WRDS (licensed)
Thomson Reuters Mutual Fund Holdings (via WRDS)	Quarterly holdings of US stocks per fund; used to construct the split-adjusted trade measure (Eq. 9)	WRDS (licensed)
Morningstar	Fund investment objectives, fund family, expense ratios, fund age; used as controls in the field analysis	Morningstar (licensed)
SEC mandatory fund filings	Fund managerial structure (team vs individual management identification); fund-level panel 1980-2018	EDGAR

Sample (field): 467 unique managers, 847 unique funds, 308 unique team observations, quarterly 1980-2018. Sample (lab): 1,512 participants across four treatments, run June-November 2024.

When to read the full paper

Read the original at https://doi.org/10.1016/j.jfineco.2025.104219 if you are studying team effects on belief formation and behavioral biases (Section 2 for the experimental design and Tables 2-5 for the core results); implementing the three-channel decomposition of team effects (Section 2.3 and Eq. 3-5 for the framework); analyzing the dynamic feedback between forecast errors and team leadership roles (Section 2.4 and Table 4); studying how organizational structure (solo vs team management) affects fund manager trading behavior and fund performance in the field (Section 3 and Tables 6-8); or looking for evidence linking overreaction to investment underperformance (Section 3.6.1 and Fig. 4). The Internet Appendix contains preregistration documents, the LLM prompts (Appendix IA3), and robustness checks.

Attribution and rights

Source: peer-reviewed, Journal of Financial Economics 176 (2026), article 104219. This distillation was extracted by an LLM on 2026-06-24 and is not human-verified or independently reproduced. All rights reserved; this page is extract-only.

Barahona, Ricardo, Stefano Cassella, Kristy A.E. Jansen, and Vincenzo Pezone. “Do teams alleviate or exacerbate overreaction in beliefs?” Journal of Financial Economics 176 (2026): 104219. DOI: 10.1016/j.jfineco.2025.104219. © 2025 Elsevier B.V. All rights reserved.