Binary Investment Bias

How to cite

Hanspal, Tobin; Fattinger, Felix; Steshkova, Alina; Koval, Borys. 2025. "Binary Investment Bias." AEA RCT Registry. AEARCTR-0015973. https://doi.org/10.1257/rct.15973-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Tobin Hanspal* , Felix Fattinger , Alina Steshkova , Borys Koval

* Primary investigator

Registry ID
AEARCTR-0015973
DOI
https://doi.org/10.1257/rct.15973-1.0
First registered
Last updated
Status
In development
Trial dates
2025-05-30 – 2025-07-31
Countries
Austria ()
JEL codes
G02, G11, G12, G40
Registry record
http://www.socialscienceregistry.org/trials/15973

Abstract

A longstanding literature in behavioral economics and psychology suggests that people use simple heuristics when making complex (financial) decisions. At the same time, recent advances in financial economics depart from models where investors hold rational expectations and suggest that people tend to extrapolate from previous realized returns when forming expectations about the future. In this project, we propose a novel belief formation mechanism which affects investment choices in both the field and laboratory and has important asset pricing implications. We hypothesize that the sign of past returns is a salient characteristic investors use to inform their trading decisions. Specifically investors employ a simple heuristic of counting the number of positive vs. negative daily returns when forming expectations about future returns. In this belief formation process, more recent returns weigh more heavily than more distant returns. This “binary investment bias” leads (retail) investors to overtrade which results in non-fundamental price pressure, i.e., mispricing, that eventually reverts. While the binary bias has received notable attention in psychology, it has largely been ignored by financial economists. Our research project combines various research designs: (1) an analysis of historical stock market and proprietary trading records to test whether the price dynamics and trading behavior implied by our hypothesized belief-formation mechanism are borne out in real-world data, and (2) laboratory experiments to establish causality and to control for potential confounding effects ubiquitous in the field.

Intervention

With a simple, standard investment experiment we examine if investors extrapolate from recent signed returns, and how attributes of recent returns affect portfolio choice. Participants partake in an investment experiment consisting of 27 independent investment decisions. In each decision, participants allocate an endowment of 10,000 experimental currency units between a risky and a risk-free asset. To inform their investment decisions, past returns are shown. While observable return distributions differ in the number of negative vs. positive returns, they have stationary means, standard deviations, and skewness, and thus provide the same information about the expected value and variability of future returns. We vary the proportion of positive to negative returns that participants observe (mostly positive vs. mostly negative) as well as the way past returns are depicted to understand how format and salience of signed returns affect investment choices. We also elicit participants' beliefs about the next period return.

Experimental design

We address our research question (i.e., if participants extrapolate from recent signed returns, and how attributes of recent returns affect portfolio choice) with a simple, standard investment experiment. Subjects participate in an investment experiment consisting of 27 independent decisions. In each decision task, participants allocate an endowment of 10,000 experimental currency units between a risky and a risk-free asset. To inform their investment decisions, past returns are shown. While past return distributions differ in the number of negative vs. positive returns, they have stationary means, standard deviations, and skewness, and thus provide the same information about the expected value and variability of future returns. We interact the main treatment (mostly positive vs. mostly negative returns) with three different visualization treatments: tables, bar plots, and colored (pos vs. negative) bar plots. The different visualizations potentially create different levels of treatment salience. For every third investment decision, we also elicit participants' beliefs about the next period return. In addition, we also implement several between-subject treatment variations: 1) standard experiment described above, 2) alteration with different ordering of visualization formats, 3) alteration controlling for 'streaks', i.e., successively positive (negative) recent returns, and 4) alteration controlling for 'recency bias', i.e., more recent vs. more distant prevalence of positive (negative) returns. Thus, in total, we will have 2 sources of within-subject variation (sign distribution of observed returns, visualization format), and 4 sources of between-subject variation (standard, visualization ordering, streaks vs. no streaks, recent vs. distant sign prevalence).

The experiment concludes with a short survey eliciting investment experience and basic sociodemographic information (age, gender, education, self-assessed risk aversion). At the end of the experiment, one of the 27 investment decisions is randomly chosen and implemented for payment. The investment outcome is then divided by 1000 and rounded to the nearest Euro amount. In case the chosen round contained belief elicitation, probability estimates are incentivized via a ranked probability score. We expect an average payoff between 10 and 13 EUR per participant.

The experiment is implemented in Qualtrics. We are going to run the experiment at WU Labs and we expect that participants will complete it in approximately 30 to 45 minutes. Note, we abstain from any form of deception.

Design details

Randomization method
The main treatment is fully randomized and determined by the randomization algorithm in the software used to code the experiment (Qualtrics, JavaScript). Specifically, from a distribution matching the historical moments of daily stock returns, 9 triplets with 3 different 20-return observations are sampled. Within each triplet, the first three moments are (almost) identical across all 3 samples. Every triplet consists of (i) a sample with at least three times as many positive than negative returns, (ii) a sample with an equal number of positive and negative returns, and (iii) a sample with at least three times as many negative than positive returns. In total, this gives 27 return samples, 9 for each visualization treatment. For each participant, the embedded code then randomly assigns 3 triplets to each visualization treatment. Within each visualization treatment, the order of the 9 samples is again fully randomized.
Randomization unit
The main source of within-subject randomization consists of the randomly selected return triplets (see above) for the 9 investment decisions within each visualization format. In the standard case, we aim for increasing treatment salience: For the first 9 investment decisions, sampled returns are shown in tables. For the second 9 investment decisions, sampled returns are illustrated with bar plots. For the final 9 investment decisions, sampled returns are illustrated with color-coded (positive vs. negative) bar plots. The between-session randomization occurs through different session structures: 1) standard treatment, 2) alternated ordering of return visualizations (bar plots, colored bar plots, tables), 3) alternated ordering of returns within samples ensuring the presence and absence of streaks, and 4) alternated ordering of returns within samples ensuring variation in the recency of mostly positive (negative) returns.
Clusters
Outcomes are clustered at the individual subject level and controlled for the session in which the subject participated in. Empirical analyses will control for subject and session effects. Depending on budgetary constraints, we are planning to collect data from approximately 250 subjects.
Observations
Approximately 250 subjects each participating in 27 (9) investment decisions (belief elicitations) yielding a total of approximately 6750 (2250) observations.
Arms
All participants are subject to all 3 visualization treatment arms. We plan for approximately 25 subjects per session across 10 sessions, amounting to 250 subjects in total. • From these 10 sessions, 4 will receive the standard treatment ~ 4x25x27 investment decisions = 2700 observations clustered by 100 subjects • From these 10 sessions, 4 will receive the recency and streaks variation treatments ~ 4x25x27 investment decisions = 2700 observations clustered by 100 subjects • From these 10 sessions, 2 will receive the order variation treatment ~ 2x25x27 investment decisions = 1350 observations clustered by 50 subjects
Primary outcomes
Experimental currency units invested in risky asset, investor beliefs about future returns and return distributions

Cited by

No citation data available for this registration yet.

Keywords