Investor Memory: Godker, Jiao & Smeets (2025)

What this is. The paper’s core results, the experimental designs, and the estimating specifications with exact magnitudes: enough to know what was found and how, without reading all 46 pages. To replicate or extend, read the full source at the original or download the replication data from Harvard Dataverse.

TL;DR

Across three incentivized experiments (N = 229, 498, 487), the paper documents a positive memory bias for investment outcomes. After 1 week, subjects overremember the positive returns and underremember the negative returns of stocks they chose to invest in. This positive memory bias: (a) distorts beliefs about stock quality by 8.16 percentage points toward optimism relative to the objective Bayesian posterior, (b) doubles the rate of suboptimal reinvestment (41.9% vs 20.5%), and (c) raises overconfident betting on one’s own stock picks by 47%. The mechanism is motivated memory suppression (not genuine forgetting), consistent with the framework of Benabou and Tirole (2002): raising the financial incentive for accurate recall eliminates the bias for subjects prone to self-deception, and the bias disappears when subjects did not actively choose their investment (passive endowment condition).

Core results

Magnitudes and significance are as reported; * p < .1, ** p < .05, *** p < .01. Locators point into the source PDF.

#	Result	Locator	Magnitude
R1	Subjects overremember positive investment outcomes after 1 week	Table 2, p. 1614; Figure 1, p. 1613	Delay: +0.89 outcomes overremembered (p = .000); Immediate: +0.27 (p = .058); difference p = .018
R2	Subjects underremember negative investment outcomes after 1 week	Table 2, p. 1614; Figure 1, p. 1613	Delay: -0.73 outcomes underremembered (p = .000); Immediate: -0.28 (p = .037); difference p = .038
R3	Beliefs are 8.16 percentage points too optimistic in the Delay condition relative to the objective Bayesian posterior	Figure 2, p. 1616; Internet App. D	t-test p = .004; effect is absent in Immediate condition
R4	Memory bias directly predicts belief distortion: each recalled positive outcome raises subjective belief by 6.20pp; each underremembered loss reduces it by 7.96pp	Table 3 cols 1-2, p. 1617	6.20*** (0.92) and -7.958*** (1.02); R2 = 0.52-0.55
R5	Suboptimal reinvestment doubles in the Delay vs Immediate condition (41.9% vs 20.5%)	Figure 3, p. 1619; Table 4 col. 2, p. 1620	Odds ratio 3.542*** (1.44) for Delay treatment dummy; significant at 1%
R6	Memory bias magnitude predicts suboptimal reinvestment: each overremembered positive outcome raises probability of suboptimal reinvestment by 59.9%	Table 4 cols 3-4, p. 1620	Odds ratio 1.599*** (0.22) for positive-outcome memory bias; 0.617*** (0.10) for negative-outcome bias
R7	Overconfident betting is 47% higher in Delay vs Reminder condition; Delay subjects bet 46.3 vs 31.5 points suboptimally	Figure 4, p. 1621; Table 5 col. 2, p. 1622	14.788*** (4.18) additional suboptimal points in Delay; p < .01
R8	Motivated memory suppression drives the bias: high recall incentives eliminate the memory bias for high-SDE subjects; only actively choosing investors exhibit the bias	Table 7, p. 1625; Table 6, p. 1623	HighStake coeff for high-SDE: -0.469** (0.19); NoChoice treatment effect 0.18 (p = .398, insignificant)

Overall (paper’s conclusion). Investors systematically overremember their gains and underremember their losses, and this positive memory bias operates as a microfoundation for three stylized facts in behavioral finance: gains weigh more than losses when investors learn from experience (Kaustia and Knupfer (2008)); investors are more likely to repurchase stocks previously sold at a gain (Strahilevitz, Odean, and Barber (2011)); and overconfident investors trade excessively (Barber and Odean (2001)). The paper also extends evidence from Kuhnen (2015) on asymmetric belief updating from financial information by adding a memory channel. The bias is motivated (driven by self-image concerns) rather than a passive cognitive limitation.

Theory / model

The paper has no formal structural model in the body. The hypothesis tested is:

H (Positive memory bias): When subjects have time to form memories (Delay condition, 1 week between observation and elicitation), they recall more positive investment outcomes and fewer negative outcomes than actually occurred, relative to a control group that elicits memory immediately (Immediate condition).

The identification logic is a between-subject random assignment to Delay vs Immediate conditions, with the 1-week gap as the treatment that activates memory processes. The Immediate condition holds constant all non-memory factors (information acquisition, salience, attention) that could influence recall.

The Bayesian benchmark for subjective beliefs is the objective posterior (equation 1, p. 1604):

$$\mu_t^G(h_t) = \frac{1}{1 + \frac{1-\mu_0^G}{\mu_0^G} \times \left(\frac{\theta}{1-\theta}\right)^{t - 2n_t^+}} \tag{1}$$

where $\mu_0^G = 50\%$ is the prior probability that the stock is good, $\theta = 60\%$ is the probability that a good stock generates a positive outcome each period, $t$ is the total number of observed outcomes, $h_t$ is the history of outcomes, and $n_t^+$ is the number of positive outcomes observed. The objective Bayesian posterior $\mu_t^G(h_t)$ is the benchmark against which subjective beliefs are compared to measure belief distortion.

The motivated-memory mechanism is tested via two conditions in experiment 2. The HighStake condition increases the financial incentive for accurate memory reporting from 8 to 50 GBP/USD per correct memory answer (following Zimmermann (2020)), so that if the true information is still in memory (suppressed, not deleted), high incentives should induce subjects to report it more accurately. The NoChoice condition randomly endows subjects with a stock rather than letting them choose, so that ego-relevance of outcomes is reduced (following Mather, Shafir, and Johnson (2000, 2003)).

The Internet Appendix formalizes a model in which memory of investment outcomes is systematically biased in a motivated way: memory bias stems from quasi-Bayesian belief updating with a probability of underremembering specific previously observed signals, where this probability depends on whether the signals are consistent with the decision-maker’s motivation (Section 3, p. 1617 reference to the Internet Appendix model).

The parimutuel betting payoff in experiment 3 follows Enke, Graeber, and Oprea (2023), equation 2 (p. 1611):

$$\text{payoff}_i = \frac{b_i}{\frac{\sum_{l=1}^{10} x_l b_l}{\sum_{l=1}^{10} b_l}} + (100 - b_i) \tag{2}$$

where $b_i$ denotes points bet by subject $i$ and $x_i$ is an indicator equal to 1 if the subject’s investment choice was optimal. This design provides an objective measure of overconfidence: subjects should bet 0 if they believe their choice was suboptimal and bet all 100 if they believe it was optimal, so suboptimal betting (betting > 0 when the choice was suboptimal) directly reveals overconfidence.

Method

The paper uses a randomized between-subject design with exogenous variation in the time span between observing investment outcomes and eliciting memory (Delay vs Immediate conditions). Three experiments share a common framework but vary in outcomes, conditions, and elicited beliefs/choices.

Common structure. Subjects observe 12 sequential investment outcomes from a risky stock over 12 periods (each shown for 2 seconds on screen). Outcomes are drawn from known distributions (positive/negative with known probabilities depending on stock quality). After the observation phase, memory, beliefs, and investment choices are elicited. The Delay condition places the elicitation in week t+1 (1 week later); the Immediate condition places it in week t (immediately after).

Memory elicitation. Subjects report how many positive and negative outcomes they observed (experiment 1: how often each specific outcome value occurred; experiments 2-3: total counts summing to 12). The memory bias at the individual level is the difference between recalled and actually observed counts of positive (or negative) outcomes.

Identification via timing variation. The comparison of Delay vs Immediate conditions isolates the effect of memory from attention, salience, and information processing, which are held constant because both groups observe the same outcomes with the same level of engagement. A NoRecall condition in experiment 1 (no memory task) confirms that the memory elicitation task itself does not affect subsequent beliefs or investment decisions.

The randomized-survey-experiment primitive underlies all three experiments: random assignment to conditions, incentivized elicitation of beliefs and choices, and exogenous variation in a single treatment variable (delay vs immediate, or HighStake vs Baseline, or NoChoice vs Baseline).

Empirical specifications

Memory bias (R1, R2). The memory bias is regressed on a constant and tested with a t-test against zero for each condition (Table 2, p. 1614):

$$\text{MemoryBias}_{i,s} = \alpha + \beta \cdot \mathbf{1}[\text{Delay}_i] + \gamma \cdot \text{Session}_{is} + \varepsilon_{is}$$

where $\text{MemoryBias}_{is}$ is the subject’s recalled minus actually observed count of positive (or negative) outcomes, and $\text{Session}_{is}$ is a session fixed effect. Column 1 of Table 2 reports the t-test of the Delay group mean against zero; column 3 tests the Delay-Immediate difference.

Memory-based beliefs (R3, R4). OLS regressions with session fixed effects (Table 3, p. 1617):

$$\text{SubjProb}_{is} = \alpha + \beta_1 \cdot \text{MemBias}_{\text{pos},is} + \beta_2 \cdot \text{MemBias}_{\text{neg},is} + \gamma \cdot \text{ObjProb}_{is} + \delta \cdot \text{Session}_{is} + \varepsilon_{is}$$

Dependent variable: subjective probability that the stock is good (1-100). Columns 3 and 4 use belief distortion (difference between posterior log-likelihood ratios of subjective and objective probabilities) as the dependent variable. N = 188, R2 = 0.31-0.55. Standard errors in parentheses.

Reinvestment behavior (R5, R6). Logit regressions with session fixed effects (Table 4, p. 1620):

$$\Pr(\text{Invest}_{is}) = F\left(\alpha + \beta \cdot \text{Delay}_i + \gamma_1 \cdot \text{MemBias}_{\text{pos},is} + \gamma_2 \cdot \text{MemBias}_{\text{neg},is} + \delta \cdot \text{Session}_{is}\right)$$

Dependent variable: a dummy equal to 1 if the subject reinvested in the stock (Inv.) or a dummy equal to 1 if the subject reinvested suboptimally from a Bayesian perspective (Inv. (Subopt.)). Odds ratios reported. N = 152, pseudo-R2 = 0.05-0.13. Sample restricted to subjects who invested in the stock.

Overconfidence / suboptimal betting (R7, R8). OLS regressions (Table 5, p. 1622):

$$\text{PointsBet}_{is} = \alpha + \beta \cdot \text{Delay}_i + \gamma_1 \cdot \text{MemBias}_{\text{pos},is} + \varepsilon_{is}$$

Dependent variable: number of points bet suboptimally (Points bet (subopt.) = points bet when the stock chosen had fewer than 6 positive outcomes, i.e., was suboptimally chosen from a Bayesian perspective). N = 191 (col. 2), N = 83 (col. 3 for Delay subjects only). R2 = 0.02-0.06. No session fixed effects in experiment 3 (collected in one online session).

Mechanism tests (R8). OLS regressions split by median self-deceptive enhancement (SDE) score (Table 7, p. 1625); Treatment (HighStake) is the dummy for the high-incentive condition. The memory-suppression prediction is that HighStake reduces bias for high-SDE subjects (column 1: coeff -0.469, p < .05) but not for low-SDE subjects (column 2: 0.115, p = .60).

Datasets used

Dataset	Role in paper	Wiki page
Experiment 1 lab data (Hamburg University, N = 229)	Primary identification of memory bias, beliefs, and reinvestment (Results 1-6)	No page yet
Experiment 2 online lab data (CESS Oxford, Xlab UC Berkeley, N = 498)	Memory suppression and active-choice mechanism tests (Result 8, Table 6-7)	No page yet
Experiment 3 online data (Prolific, UK, N = 487)	Overconfidence measure via parimutuel betting (Results 6-8)	No page yet

All three experiments use primary hand-collected data: incentivized laboratory or online experiments with randomized treatment assignment. No external financial databases are used; the investment outcomes are computer-generated draws from known distributions. Experiment 1 programmed in z-Tree (Fischbacher 2007); experiments 2 and 3 in oTree (Chen, Schonger, and Wickens 2016). Experiment 3 preregistered at AsPredicted under ID 153791.

When to read the full paper

Read the original if you are: studying the microfoundations of investor overconfidence or asymmetric learning from experience; replicating the experimental paradigm (exact outcome distributions in Appendix D, instructions in Appendices A-C); extending the motivated-memory model to field settings or longer horizons; or connecting this to reinforcement learning in finance (Kaustia and Knupfer (2008)) and stock repurchase anomalies (Strahilevitz, Odean, and Barber (2011)). The exact Table/Figure locators above point to the key results. Replication code and data are publicly available at Harvard Dataverse (https://doi.org/10.7910/DVN/7K6ZPK).

Attribution and rights

Source: peer-reviewed, The Review of Financial Studies 38(6), 2025. Published by Oxford University Press on behalf of the Society for Financial Studies. All rights reserved. Commercial re-use requires reprints permission from OUP. This distillation was extracted by an LLM on 2026-06-06 and is not human-verified or independently reproduced. Extract-only: the OUP standard publication reuse rights do not permit mirroring the verbatim PDF.

Gödker, Katrin, Peiran Jiao, and Paul Smeets. “Investor Memory.” The Review of Financial Studies 38, no. 6 (2025): 1595-1640. DOI: 10.1093/rfs/hhaf006. © 2025 The Author(s). Published by Oxford University Press on behalf of the Society for Financial Studies.