Salience-driven mispricing in retail-dominated markets: evidence from Korea

Traditional asset pricing models assume that investors are rational and utilize all available information. However, a substantial body of research documents that investors' attention and processing capacity are limited (Kahneman, 1973; Daniel et al., 1998). Bordalo et al. (2012) propose the salience theory, arguing that decision-makers' attention gravitates toward the most unusual attributes of available options. In asset pricing (Bordalo et al., 2013), this implies that investors over-weight salient payoffs, such as a stock's potential for extreme returns, leading them to overprice assets with salient upsides and underprice those with salient downsides. This mechanism generates predictable return patterns, a hypothesis strongly supported by US evidence (Cosemans and Frehen, 2021) ^[1].

This study investigates the salience effect in the Korean stock market (Yun et al., 2009; Han et al., 2020), a setting that provides a particularly powerful laboratory to test the underlying channels of salience theory for two key reasons. First, the Korean market is characterized by an exceptionally high and persistent level of retail investor participation, with individuals consistently accounting for over 70% of total trading volume (Kang et al., 2013; Kim et al., 2025) ^[2], coupled with highly volatile sentiment among retail investors (Seok et al., 2019; Kim and Park, 2015; Byun et al., 2023). As retail investors are generally considered less sophisticated and more prone to behavioral biases (Barber and Odean, 2000; Kumar, 2009), this retail-dominated structure allows for a clean test of whether salience-driven mispricing is indeed amplified by the presence of behaviorally biased investors ^[3]. The importance of investor attention and sentiment in the Korean market is well-documented (e.g. Fan et al., 2022) ^[4]. Importantly, the availability of daily trading data disaggregated by investor type from FnGuide allows us to directly examine whether individual investors are indeed net buyers of high-ST stocks, as the theory suggests ^[5].

Second, the Korean market has a history of implementing market-wide short-sale bans in response to external shocks (Eom et al., 2021; Kim and Seo, 2015; Chun and Cho, 2019; Lee, 2024) ^[6]. These regulatory interventions serve as powerful quasi-natural experiments that enable us to provide causal evidence on the role of limits to arbitrage, a channel that prior studies have often proxied using cross-sectional firm characteristics. The combination of high retail participation and stringent, time-varying arbitrage constraints creates a unique environment where behavioral mispricing can become particularly pronounced and persistent ^[7].

Using comprehensive data on Korean stocks from March 2005 to December 2024, our empirical results provide strong support for salience theory in the Korean context. The equal-weighted high-minus-low ST quintile portfolio generates average monthly returns of −1.31% (t-stat = −4.85), with a four-factor alpha of −1.50% (t-stat = −5.91). The value-weighted high-low ST portfolio earns −1.07% per month (t-stat = −3.16) with a four-factor alpha of −1.19% (t-stat = −3.59). The fact that equal-weighted portfolios show stronger effects than value-weighted portfolios is consistent with small stocks having greater limits to arbitrage and being more susceptible to behavioral mispricing ^[8].

We demonstrate that the salience effect is distinct from related well-known anomalies. Given that ST is conceptually linked to investor overreaction and the demand for lottery-like stocks, we explicitly test its incremental explanatory power. In bivariate-sort analysis and Fama and MacBeth (1973) regressions, controlling for short-term reversal (REV), lottery-demand proxies (MAX), and idiosyncratic volatility (IVOL), ST remains a highly significant predictor of future returns. To further isolate the unique content of ST, we construct orthogonalized ST variables that are purged of the influence of REV, MAX, and IVOL. The results show that these orthogonalized ST variables continue to have significant predictive power. For example, a portfolio sorted on ST orthogonalized with respect to MAX still yields a significant four-factor alpha of −0.40% (t-stat = −2.55). This confirms that the salience measure captures a dimension of investor behavior not fully subsumed by other anomalies.

Moreover, and most importantly, we provide direct evidence that the salience effect is driven by retail investors. We conduct Fama and MacBeth (1973) regressions, including an interaction term between ST and the net buying intensity of individual investors. The results show a significantly negative coefficient on this interaction term (−0.269, t-stat = −3.96), confirming that the negative ST-return relation is significantly amplified by retail investor trading. This finding is corroborated by non-parametric double-sorting analyses (Table 6), which show the ST return spread is largest and most significant in the quintile of stocks with the highest retail investor participation: the spread is −1.03% (t-stat = −3.22) in the high individual-net-buying quintile, compared to an economically small and insignificant −0.16% (t-stat = −0.55) in the low quintile. In contrast, the ST effect weakens or disappears when institutional or foreign net buying is high. These results directly link the mispricing to the trading behavior of the investor class most likely to be affected by salience bias.

Finally, we show that the salience effect is not uniform and varies predictably with the market's features. First, we find that the salience effect is significantly stronger among stocks with greater limits to arbitrage, such as small-cap stocks, low-priced stocks, and stocks with high idiosyncratic volatility ^[9]. This is consistent with the hypothesis that arbitrage constraints allow salience-induced mispricing to persist. More directly, we analyze periods of market-wide short-sale bans and find that the ST effect is significantly stronger when these bans are in place. As detailed in our time-series regressions, the coefficient on the short-sale ban indicator is −0.038 (t-stat = −3.27) in the full model. Furthermore, the salience effect also exhibits time-variation linked to investor sentiment or market uncertainty. Using Korean-specific sentiment proxies, we find that the salience effect is substantially stronger during low-sentiment or high uncertainty periods ^[10]. Rather than intensifying in optimistic markets, salience-driven mispricing appears to be more pronounced when sentiment is subdued or uncertainty is high. This pattern is consistent with evidence in the Korean market showing that certain behavioral anomalies strengthen in low-sentiment states, potentially because limits to arbitrage become more binding and capital constraints more severe during such periods.

Our study contributes to the literature in three significant ways. First, we provide robust out-of-sample evidence for salience-based asset pricing that moves beyond confirmation to an in-depth analysis of its underlying mechanisms. While many US-based anomalies weaken or disappear in international samples (Hou et al., 2011; McLean and Pontiff, 2016), we demonstrate that the salience effect documented by Cosemans and Frehen (2021) and Goh et al. (2026) is not only strong but economically more pronounced in the Korean market. By exploiting Korea's history of market-wide short-sale bans as a quasi-natural experiment, we provide causal evidence that limits to arbitrage are a critical channel that allows salience-driven mispricing to persist and grow. This finding substantiates the theoretical link between arbitrage constraints and behavioral biases in a clean, event-based setting.

Second, we contribute to the literature on individual investor behavior (e.g. Barber and Odean, 2000; Kumar and Lee, 2006; Han and Kumar, 2013; Goh and Kim, 2024) by providing direct evidence that the salience effect is driven by retail investors. While prior research suggests that stocks with high retail trading tend to be overpriced, our study identifies a specific, salience-driven attention mechanism through which this mispricing occurs. The availability of disaggregated trading data in Korea allows us to directly test this channel in a way that is difficult with US data. Our findings from Fama and MacBeth (1973) regressions with interaction terms and double-sorting analyses clearly show that the salience effect is concentrated in stocks heavily purchased by individual investors and is attenuated when more sophisticated institutional investors are active. This intuition is also consistent with recent Korean evidence showing that investor attention and market sentiment play an important role in shaping stock-price responses in the Korean market (Fan et al., 2022).

Finally, we clarify the unique explanatory power of the salience measure by carefully disentangling it from related, well-known anomalies. A key challenge in this literature is that salience (ST) is conceptually and empirically correlated with other behavioral phenomena, such as short-term reversal (REV) and the preference for lottery-like stocks (MAX) (Kang and Sim, 2014). While some studies, such as Cakici and Zaremba (2022), have examined these effects in international markets, a rigorous test of ST's incremental predictive power is crucial. Through orthogonalization procedures and double-sorting analyses, we demonstrate that the salience measure retains significant, incremental predictive power for future returns even after controlling for these other effects. This finding confirms that salience theory offers a distinct lens for understanding the cross-section of stock returns, capturing a dimension of investor psychology not fully subsumed by other behavioral models.

The remainder of this paper is organized as follows. Section 2 describes the data and variable construction. Section 3 presents our main empirical results on portfolio sorts and cross-sectional regressions. Section 4 provides additional analyses examining cross-sectional variation, investor trading behavior, market state dependence, and robustness checks. Section 5 concludes.

We use stock data from the Korean market, listed on the Korea Composite Stock Price Index (KOSPI) and Korean Securities Dealers Automated Quotations (KOSDAQ) markets, from March 2005 to December 2024 ^[11]. The sample period spans 20 years and includes multiple market cycles, the 2008 financial crisis, and the COVID-19 pandemic. This provides a robust test of salience theory across different market conditions. We obtain daily and monthly stock returns, outstanding shares, trading volume, and share prices from DataGuide, provided by FnGuide. DataGuide is the most comprehensive database for Korean stocks and is widely used in academic research on Korean markets.

We also obtain book equity from annual financial statement data. Accounting information is measured at the fiscal year-end and is assumed to become available to investors at the end of June of the following year. Accordingly, book equity for fiscal year t is matched to monthly stock returns from July of year t+1 to June of year t+2. This convention ensures that our analysis uses only information available to investors at the time of portfolio formation. This avoids look-ahead bias that could contaminate our results.

A distinctive feature of our data is the availability of daily trading volumes disaggregated by investor type. DataGuide provides daily trading volumes for each market participant, categorized as individual, institutional, and foreign investors. Individual investors include retail traders and high-net-worth individuals. Institutional investors include domestic mutual funds, pension funds, insurance companies, and securities firms trading for their own accounts. Foreign investors include both institutional and individual investors based outside Korea. This disaggregation enables a more precise analysis of the direct trading behavior of different market participants and allows us to test whether individual investors are indeed attracted to high-ST stocks, as salience theory predicts.

After constructing the variables, we apply several filters to ensure data quality. We exclude financial firms from the sample due to their distinct financial structure. We also exclude observations with a stock price of less than 1,000 Korean won (approximately 1 US dollar) ^[12] to alleviate concerns that our results are driven by micro-cap stocks or influenced by market microstructure issues that affect very low-priced stocks. Finally, we require stocks to have more than 15 trading days in a month to construct meaningful salience measures.

Our final sample includes 378,582 stock-month observations from 2,911 unique stocks. KOSDAQ stocks tend to be smaller, more volatile, and have lower institutional ownership than KOSPI stocks, which allows us to examine whether salience effects vary across different types of stocks. The average firm-month observation has a market capitalization of approximately 700 billion Korean won, though there is substantial variation ranging from very small firms to large chaebols.

Following Bordalo et al. (2012, 2013) and Cosemans and Frehen (2021), we construct the salience theory variable (ST) in five steps. The ST variable is designed to capture how salience-driven attention distorts investors' expectations about future returns.

First, for each stock i in each month t, we collect all daily returns (⁠$ris,t$⁠) during the month. We require more than 15 trading days to construct a reliable salience measure. For each trading day s in month t, we calculate the equal-weighted market return (⁠$r̅s,t$⁠) as the average return across all eligible stocks trading on that day. We use equal-weighting rather than value-weighting because it preserves the properties of the salience function, as demonstrated by Bordalo et al. (2012).

Second, for each stock-day observation, we calculate the salience of stock i's return on day s relative to the market return using the salience function. The salience function measures how much a stock's return stands out relative to the market by comparing the absolute difference between the stock return and market return to the average magnitude of these returns ^[13]. Specifically, salience is defined as the absolute difference divided by the sum of absolute values plus a small constant θ = 0.1, as follows:

Third, within each stock-month, we rank the daily returns as $kis,t$ in descending order of salience. The most salient return receives rank 1 (i.e. $kis,t=1$⁠), and the least salient return receives rank $St$⁠, (i.e. $kis,t=St$⁠) where $St$ is the number of trading days in month t ^[14]. We then compute the raw salience weight for each stock i on day s relative as $δkis,t$⁠, by applying an exponential decay function with parameter δ = 0.7 to the rank $kis,t$⁠. This weighting scheme assigns greater weight to more salient days and captures the idea that investors pay more attention to days with unusual returns ^[15].

Fourth, the salience weights are normalized so that they sum to one within each stock-month as follows:

This ensures that the weights represent a probability distribution over days and allows us to interpret ST as the difference between two expected returns: one computed using salience weights and one using equal weights.

Finally, we calculate ST as the covariance between the normalized salience weights and the daily returns. Equivalently, ST equals the salience-weighted average return minus the equal-weighted average return as follows:

If a stock's most salient days had high returns, ST is positive, indicating that salience leads investors to overestimate future returns. If the most salient days had low returns, ST is negative, indicating that salience leads investors to underestimate future returns. Salience theory predicts that high-ST stocks will be overpriced and earn lower subsequent returns, while low-ST stocks will be underpriced and earn higher subsequent returns ^[16].

We construct a comprehensive set of control variables following the asset pricing literature. These variables are designed to capture known return predictors and alternative explanations for our findings. By controlling for these variables, we can isolate the unique predictive power of ST and distinguish salience effects from other phenomena.

Standard firm characteristics include: (1) ME (size), the market capitalization at the end of month t; (2) BM (book-to-market), calculated as book equity divided by market equity using book values from the fiscal year-end; (3) MOM (momentum), defined as the cumulative return from month t − 12 to t − 2, skipping the most recent month to avoid overlap with short-term reversal; (4) REV (short-term reversal), measured as the stock return in month t − 1; (5) ILLIQ (Amihud illiquidity), calculated as the average of daily absolute return divided by trading volume in million Korean won over month t − 1, following Amihud (2002), (6) MAX, the maximum daily return within month t − 1, following Bali et al. (2011), (7) BETA (market beta), estimated by regressing daily excess returns on daily market excess returns over month t − 1; (8) IVOL (idiosyncratic volatility), calculated as the standard deviation of residuals from the beta regression, following Ang et al. (2006), These variables capture well-known return predictors documented in the cross-sectional asset pricing literature.

Table 1 reports summary statistics and correlations for the main variables. ST is strongly positively correlated with the past one-month return (REV), reflecting that extreme daily returns both increase salience and mechanically affect short-term cumulative returns. In addition, ST is positively associated with maximum daily return (MAX) or idiosyncratic volatility, consistent with salient experiences being more likely for stocks with asymmetric return distributions.

Table 2 reports the time-series averages of firm characteristics for portfolios sorted into deciles based on the salience theory variable (ST). The cross-sectional patterns reveal that stocks with high salience exhibit distinct features, although not always in a monotonic fashion.

Stocks in the highest ST decile (Decile 10) are generally related to lottery-like features and recent strong performance. They have the highest idiosyncratic volatility (IVOL at 0.06), the most extreme positive daily returns (MAX at 0.17), and the highest short-term past returns (REV at 0.25). Other key characteristics display U-shaped or hump-shaped patterns. Market Capitalization (ME) and Price (PRC) both exhibit an inverted U-shape. They increase from the lowest ST decile to the 3rd decile (ME peaks at 9.41, PRC peaks at 29.06) and then decline sharply towards the highest ST decile (ME at 2.59, PRC at 14.21). This indicates that the most salient stocks are not the micro-cap stocks found in the lowest ST decile, but rather smaller-cap stocks relative to the broad market average. Market beta (BETA) displays a U-shaped pattern, being highest in the extreme deciles (1.05 in Decile 1 and 0.96 in Decile 10) and lower in the middle deciles. Momentum (MOM) shows a similar, albeit weaker, U-shaped tendency.

Notably, the book-to-market ratio (BM) shows a distinct hump-shaped pattern. It rises from 1.02 in the lowest ST decile to a peak of 1.33 in Decile 3, and then declines steadily to 0.77 in the highest ST decile. This finding is particularly interesting as it differs significantly from the results for the US market reported by Cosemans and Frehen (2021), where BM showed a more consistent decline. Our result suggests that in the Korean market, the most salient stocks are strongly growth-oriented, while stocks with moderately negative salience exhibit value characteristics. This divergence could be attributable to the strong preference for high-growth, speculative stocks among Korean retail investors, who are the primary traders of high-ST stocks.

All independent variables in cross-sectional regressions are winsorized at the 1st and 99th percentiles. They are standardized each month to have a zero mean and unit variance. This standardization serves two purposes. First, it facilitates comparison of economic magnitudes across variables with different scales. Second, it ensures that our regression coefficients represent the effect of a one-standard-deviation change in each variable, making the results more interpretable. 1n portfolio sorts, we use raw (unstandardized) values to assign stocks to portfolios, as this better preserves the economic interpretation of the sorting procedure.

We begin by examining whether stocks with high ST values earn lower subsequent returns than stocks with low ST values, as predicted by salience theory. At the end of each month, we sort all stocks into decile portfolios based on ST and calculate equal-weighted and value-weighted returns over the following month. Portfolios are rebalanced monthly, and we report time-series averages of portfolio returns over our sample period from March 2005 to December 2024. We calculate both raw excess returns (in excess of the risk-free rate) and risk-adjusted alphas from several factor models.

Table 3 presents the results. The first column reports raw monthly excess returns in percent. Subsequent columns report alphas from Fama and French (1993) three-factor and Carhart (1997) four-factor models. We calculate t-statistics using Newey and West (1987) standard errors with 12 lags to account for potential autocorrelation in portfolio returns.

The results provide strong overall support for salience theory in the Korean market, but they also reveal a distinct return pattern that differs from prior US evidence. Panel A shows the results for equal-weighted portfolios. A non-monotonic pattern is observed across the deciles. Returns are positive and fluctuate without a clear trend for the first seven deciles, with the low-ST portfolio (Decile 1) earning 1.04% per month (t-stat = 2.35). However, a sharp and economically significant decline begins at Decile 9. The portfolio with the highest salience (Decile 10) earns a substantial negative raw return of −0.27% per month (t-stat = −0.46). The long-short strategy (High-minus-Low ST) generates a raw return of −1.31% per month, which is highly statistically significant (t-stat = −4.85). This spread represents an annualized return of approximately −15.72%, indicating the economic significance of the salience effect. After adjusting for standard risk factors using the Fama-French four-factor model, the alpha of this spread is even more pronounced at −1.50% per month, with a t-statistic of −5.91 ^[17]. These results confirm that a strategy of shorting high-ST stocks and buying low-ST stocks would have generated substantial abnormal returns ^[18].

It is important to note that this significant spread is almost entirely driven by the severe underperformance of the stocks in the top two deciles (Deciles 9 and 10). This pattern contrasts with the more monotonic decline observed in the US market by Cosemans and Frehen (2021). This distinction suggests that in the Korean market, which is dominated by retail investors, the mispricing effect of salience is particularly concentrated at the extreme positive end. Investors appear to be disproportionately attracted to stocks with the most salient upside potential (the highest ST), leading to severe overpricing and a subsequent, sharp price correction. The behavior for stocks with average or moderately negative salience is less pronounced.

Panel B, which reports the results for value-weighted portfolios, tells a similar story. The high-minus-low spread remains economically large and statistically significant, with a raw return of −1.07% per month (t-stat = −3.16) and a four-factor alpha of −1.19% per month (t-stat = −3.59). The fact that the equal-weighted portfolios show slightly stronger effects than the value-weighted portfolios is consistent with the interpretation that salience-driven mispricing is more severe among smaller, less liquid stocks where limits to arbitrage are greater. The non-monotonic pattern also persists in the value-weighted results, reinforcing the conclusion that the effect is predominantly driven by the sharp underperformance of the highest-ST stocks.

A key challenge in establishing the validity of the salience effect is to demonstrate that the ST variable provides incremental explanatory power beyond other well-known anomalies that are also linked to extreme returns and investor overreaction (e.g. preference for lottery-like stocks (MAX) and short-term reversal (REV)). Given the conceptual and empirical overlap between these variables, as evidenced by the characteristics in Table 2, a rigorous analysis is required to disentangle their effects.

To address this concern, we first perform a sequential two-way portfolio sort analysis. Each month, we first sort stocks into quintiles based on a control variable, then within each quintile, we further sort stocks into quintiles based on ST. This procedure creates 25 portfolios, and we calculate the return spread between high-ST and low-ST stocks for a given level of each firm-characteristic variable over the following month. Portfolios are rebalanced monthly, and we report the average high-low ST spread across each control variable quintile, with three-factor alphas and four-factor alphas. If ST simply proxies for the control variable, the spread should be zero or insignificant within quintiles of that variable.

According to Panel A of Table 4, the salience effect persists across virtually all controls. When controlling for size (ME), the average spread is −1.20% (t-stat = −4.80) and remains significant. Similarly, when controlling for book-to-market (BM), the average spread remains significant (−1.18%, t-stat = −5.24). Also, the salience effect remains significant after controlling for short-term reversal (REV). The average spread is −0.78% (t-stat = −3.93), which is smaller than the unconditional spread of −1.31% but still economically and statistically significant.

When controlling for MAX, the lottery-demand proxy of Bali et al. (2011), the results provide compelling evidence for the incremental explanatory power of ST. Within each of the five MAX quintiles, the High-minus-Low ST portfolio spread is consistently negative in terms of Fama-French three-factor alphas. The average spread across the five MAX quintiles is −0.35% per month (t-stat = −2.41). Moreover, the average spread generates a Fama-French four-factor alpha of −0.41% (t-stat = −2.76). While both ST and MAX relate to extreme positive returns, they capture different aspects of investor psychology. MAX measures the single highest return regardless of its context, whereas ST measures how much a stock's returns stand out relative to the market's concurrent performance. The results from double-sort analysis in Appendix Table A1 confirm that even among stocks with similar lottery-like features (e.g. short-term reversal or idiosyncratic volatility), the salience of their return distributions, as measured by ST, has additional, independent predictive power for future returns.

To further strengthen this conclusion and address the high correlation between ST, REV, and IVOL, we conduct a more stringent test using an orthogonalization procedure. We create new, orthogonalized ST variables by taking the residuals from monthly cross-sectional regressions of ST on REV, IVOL, and MAX, respectively. At the end of each month, we sort all stocks into decile portfolios based on orthogonalized ST and calculate equal-weighted and value-weighted returns over the following month. Portfolios are rebalanced monthly, and we report time-series averages of portfolio returns, with three-factor alphas and four-factor alphas. This process purges the ST measure of its correlation with these other anomalies.

As shown in Panel B, portfolios sorted on these orthogonalized ST variables continue to yield significant negative alphas. For instance, the High-minus-Low portfolio formed on ST orthogonalized with respect to REV (ST_REV) produces a four-factor alpha of −0.68% (t-stat = −5.53), and the portfolio based on ST orthogonalized with respect to MAX (ST_MAX) produces a significant alpha of −0.40% (t-stat = −2.55). This provides robust evidence that the salience measure captures a distinct dimension of investor behavior not subsumed by other prominent return predictors (e.g. size, value, momentum, reversal, illiquidity, beta, idiosyncratic volatility, or lottery demand). This conclusion is further strengthened by the cross-sectional regression analysis that follows.

To test whether the negative relation between ST and future returns holds at the individual stock level after simultaneously controlling for multiple firm characteristics, we perform Fama and MacBeth (1973) cross-sectional regressions. We run Fama and MacBeth (1973) regressions each month, regressing month t+1 returns on month t firm characteristics, including ST and various control variables. We then calculate the time-series average of the monthly coefficient estimates and compute t-statistics using Newey and West (1987) standard errors with 12 lags. This approach allows us to control for many variables simultaneously and assess the marginal predictive power of ST.

Table 5 reports the results. Model 1 includes only ST as the explanatory variable. The coefficient is −0.460% per month (t-stat = −6.67), indicating that a one-standard-deviation increase in ST is associated with a 0.46% decrease in next month's return. This corresponds to approximately −5.52% per year and is highly statistically significant. Model 2 adds standard firm characteristics (ME, BETA, BM, MOM, ILLIQ, and PRC). The coefficient on ST remains negative and highly significant at −0.412% (t-stat = −6.84), indicating that the salience effect is not subsumed by conventional size, value, momentum, market beta, liquidity, or price effects ^[19].

Models 3 and 4 examine whether the ST effect survives after controlling for short-term reversal (REV). ^[20]. In Model 3, which adds REV, the coefficient on ST remains strongly negative and significant (−0.405%, t-stat = −5.25), while REV itself is negative but statistically insignificant (−0.131%, t-stat = −1.21). In Model 4, which includes a set of control variables, ST remains negative and significant (−0.337%, t-stat = −5.85), although its magnitude slightly declines relative to earlier specifications. Taken together, these models show that ST captures a distinct dimension of return predictability beyond mechanical reversal.

Models 5, 6, and 7 control MAX, IVOL, and progressively incorporate the full set of controls. In Model 5 (including MAX), the coefficient on ST drops sharply to −0.073% (t-stat = −1.04) ^[21]. In Model 6, which includes IVOL, ST remains economically small and only marginally significant (−0.110%, t-stat = −1.93) ^[22]. Finally, in Model 7, the full specification including REV, MAX, IVOL, and all firm characteristics, the coefficient on ST is −0.144% (t-stat = −2.73).

While the coefficient on ST is indeed attenuated in Model 5, particularly after controlling for MAX, it would be incorrect to conclude that its predictive power is entirely subsumed. As demonstrated in our non-parametric double-sort analysis in Section 3.2, the ST effect remains statistically and economically significant even after first controlling for MAX. Furthermore, an even more stringent test using orthogonalized ST variables (Panel B of Table 4), which purges the ST measure of its linear association with REV, IVOL, and MAX, confirms that ST retains significant, independent predictive power. Taken together, these results show that while there is a meaningful overlap between these variables, the salience effect does not vanish.

This observation motivates a deeper investigation into the primary drivers of this phenomenon, which is a central theme of our paper. As we will demonstrate in detail in Section 4.1, the predictive power of ST is critically dependent on the type of investor trading in the stock. Specifically, we will show through interaction-term regressions that the salience effect is almost entirely concentrated in stocks that are being actively purchased by individual investors.

A direct implication of salience theory is that investors attracted to salient stocks should be those most susceptible to behavioral biases, namely, individual investors. Given that individual investors account for 70–80% of trading volume, as shown in Figure 1, this represents considerable buying pressure that can move prices away from fundamental values. If we observe this pattern in direct trading data, rather than inferring it indirectly, this might provide unusually clear evidence of the mechanism through which salience affects prices.

We test our main conjecture by examining whether the ST effect in next-month returns varies across stocks with different levels of investor-type net buying. Net buying (Netbuy) is defined as (buy volume − sell volume) divided by shares outstanding and expressed as a percentage. Each month, stocks are independently sorted into quintiles based on Netbuy and ST. This procedure creates 25 portfolios, and we report the return spread between the highest and lowest ST quintiles (Q5 − Q1) within the bottom and top Netbuy quintiles of each market participant over the following month. Portfolios are rebalanced monthly, and we report the high-low ST return, three-factor alphas, and four-factor alphas.

Table 6 reports the ST return spreads (Q5 − Q1) conditional on investor-type Netbuy. For individual investors (Panel A), the ST effect is economically small and statistically insignificant among stocks with low individual net buying (Diff = −0.16, t-stat = −0.55). In contrast, among stocks with high individual net buying, the ST spread is large and strongly significant (Diff = −1.03, t-stat = −3.22). The effect remains highly significant after adjusting for common risk factors (FF3 alpha = −1.22, t-stat = −3.89; FF4 alpha = −1.35, t-stat = −4.71). This pattern indicates that salience-based mispricing is substantially stronger when individual investors are actively buying.

Panels B and C present analogous results for institutional and foreign investors. For institutional investors, the ST spread is negative but only marginally significant in the low-Netbuy group (Diff = −0.37, t-stat = −1.26) and becomes economically negligible in the high-Netbuy group (Diff = 0.08, t-stat = 0.27). A similar pattern emerges for foreign investors: the ST effect is strong and significant when foreign net buying is low (Diff = −1.07, t-stat = −3.89), but weaker when foreign net buying is high (Diff = −0.53, t-stat = −1.92).

This stark difference in trading behavior provides direct evidence supporting the behavioral interpretation. Individual investors appear to be attracted to stocks with salient upsides, consistent with salience theory's prediction. In contrast, sophisticated investors recognize this bias and trade in the opposite direction, attempting to profit from the mispricing. However, the fact that institutional and foreign selling only partially offset individual buying explains why the mispricing persists in equilibrium. Short-sale constraints and capital constraints limit how aggressively sophisticated investors can trade against the bias. The results are consistent with the previous literature (Kim et al., 2025; Jeong et al., 2025, 2026), stating the different behaviors among individual vs. institutional or foreign investors.

Furthermore, we perform the Fama and MacBeth (1973) regressions that include an interaction term between ST and the net buying intensity of individual investors (ST $×$ Netbuy_IND), with all the control variables the same as Table 5. We additionally include abnormal turnover (ABTURN), defined as monthly share turnover (volume divided by shares outstanding) normalized by its average over the prior 12 months. This variable captures unusually high trading activity relative to its typical level.

The results in Table 7 are striking. In the full specification corresponding to Model 7 of Table 5, we now include this interaction term. We find that while the standalone ST coefficient becomes insignificant, the coefficient on the interaction term ST $×$ Netbuy_IND is strongly negative and highly significant (−0.269, t-stat = −3.96).

This finding provides a clarification for the results in Table 5. The attenuation of the ST coefficient when controlling for MAX and IVOL does not mean the salience effect disappears; rather, it indicates that the unconditional ST variable is an imperfect proxy. The true channel through which salience operates is via the trading of retail investors. The mispricing is not a general feature of all high-ST stocks, but is specifically concentrated in those high-ST stocks that are being actively purchased by individuals. When sophisticated investors (like institutions) are net buyers, the effect is muted or even reversed.

Therefore, the cross-sectional regression results, when interpreted together, provide compelling evidence for our central thesis: salience-driven mispricing is a real and robust phenomenon, but its manifestation is critically dependent on the presence of behaviorally biased retail investors and is most pronounced in stocks where their trading impact is largest.

If the salience effect reflects behavioral mispricing, we expect it to be stronger among stocks with greater limits to arbitrage. When sophisticated investors face high costs or constraints in trading against mispricing, behavioral biases can have a more persistent impact on prices. We test this hypothesis using two complementary approaches: first, by examining cross-sectional variation in the ST effect based on firm-level arbitrage-cost proxies, and second, by directly testing the impact of market-wide short-sale bans over time.

We begin by examining whether the salience effect varies across stocks sorted by well-known proxies for limits to arbitrage: size (ME), price (PRC), and idiosyncratic volatility (IVOL). High-IVOL stocks, in particular, are widely interpreted as being costlier to arbitrage due to greater firm-specific risk and higher uncertainty (Cho and Yoo, 2022). Each month, we sort stocks into terciles based on one of these variables and then, within each tercile, sort stocks into quintiles based on ST. We then calculate the return spread between high-ST and low-ST stocks for a given level of each variable over the following month, accompanied by three-factor alphas and four-factor alphas.

Table 8 reports the results, which strongly support the limits-to-arbitrage interpretation. When sorting by size, the equal-weighted high-low ST spread is −0.80% (t-stat = −3.17) for small stocks compared to −0.41% (t-stat = −2.03) for large stocks ^[23]. Similarly, when sorting by stock price, the spread is a substantial −1.39% (t-stat = −5.17) for low-price stocks versus an insignificant −0.19% (t-stat = −1.08) for high-price stocks. The pattern is most striking for idiosyncratic volatility. The salience spread is statistically insignificant of 0.17% (t-stat = 1.31) for low-IVOL stocks but is a large and significant −1.14% (t-stat = −3.20) for high-IVOL stocks ^[24]. These cross-sectional patterns are consistent: the salience effect is most pronounced where arbitrage is expected to be most constrained.

While the cross-sectional tests are supportive, a more direct and powerful test of the limits-to-arbitrage channel comes from exploiting the unique regulatory environment of the Korean market. Korea has a history of imposing market-wide short-sale bans during periods of market stress. These events provide a quasi-experimental setting to test how the complete removal of a key arbitrage tool affects salience-driven mispricing.

To test this directly, we follow the methodology of Cakici and Zaremba (2022) and run time-series regressions of the ST long-short portfolio returns on a dummy variable (SHORT) that equals one during months with a market-wide short-sale ban and zero otherwise. The SHORT dummy captures the incremental effect on the ST spread during periods when this critical arbitrage mechanism is unavailable. Short selling in the Korean market was temporarily banned during several periods in our sample, including October 2008 to May 2009 (global financial crisis), August 2011 to November 2011 (European sovereign debt crisis), and February 2020 to March 2025 (COVID-19 pandemic and related market interventions).

Table 9 reports the results. We control for investor sentiment (SENT), market uncertainty (VKOSPI), and market-wide limits-to-arbitrage measures, including idiosyncratic risk (IRISK), firm size (ASIZE), trading volume (DVOL), and illiquidity (AMIH). In Model (4), the full specification including all variables, the coefficient on the SHORT dummy is −0.038 with a t-statistic of −3.27 ^[25].

This result indicates that the profitability of the ST long-short strategy is significantly higher (by 3.8% points per month) during periods when short-selling is banned. This provides strong, direct evidence supporting our hypothesis. When the primary tool for correcting overpricing is removed, the mispricing caused by investors' salience bias becomes more severe and persistent, leading to larger subsequent price corrections. Taken together, the consistent findings from both our cross-sectional and time-series analyses provide compelling support for the interpretation that the salience effect represents behavioral mispricing that survives and thrives in environments characterized by greater limits to arbitrage.

We investigate whether the salience effect varies with market conditions. For example, Cheon and Lee (2018) document that the MAX effect is more pronounced during high-uncertainty periods, suggesting that investors' preference for lottery-like stocks intensifies when market conditions are volatile.

The link between salience and lottery demand is particularly relevant in this context. Salience theory provides a more general framework that encompasses but extends beyond pure lottery preferences. While MAX captures only the single highest return, ST captures the entire distribution of salient days weighted by how much they stand out. This salience-driven mispricing should be particularly severe because investors not only chase lottery-like payoffs but also overweight any returns that capture their attention.

We test these predictions by constructing Korean-specific proxies for market sentiment and uncertainty. For sentiment, we use the Economic Sentiment Index (ESI) ^[26], a composite indicator that combines the BSI and the CSI. The ESI captures overall economic sentiment among private economic agents and is published by the Bank of Korea. We classify months as high-sentiment when the sentiment is above their median values, and low-sentiment otherwise. This composite approach ensures that our results are not driven by any single sentiment proxy. For uncertainty, we use the implied volatility derived from KOSPI 200 options (VKOSPI index). We classify months as high-uncertainty when either measure exceeds its median, and as low-uncertainty when both measures are below their medians. This classification focuses on periods of clearly elevated or subdued uncertainty.

For each monthly classification of sentiment or uncertainty, we sort all stocks into quintiles based on ST and calculate equal-weighted and value-weighted returns over the following month. Portfolios are rebalanced monthly, and we report time-series averages of portfolio returns over our sample period from March 2005 to December 2024. We calculate both raw excess returns (in excess of the risk-free rate) and risk-adjusted alphas from several factor models.

Panel A of Table 10 shows that the salience effect is, in fact, stronger during low-sentiment periods. The equal-weighted Q5 − Q1 spread is −0.91% (t-stat = −5.17) in low-sentiment months, compared to −0.68% (t-stat = −2.29) in high-sentiment months. The difference in risk-adjusted performance is similar: the FF3 alpha is −0.96% (t-stat = −5.37) under low sentiment versus −0.60% (t-stat = −1.76) under high sentiment, and the FF4 alpha is −0.97% (t-stat = −5.52) versus −0.88% (t-stat = −2.74). Thus, although the effect remains negative in both states, it is economically larger and statistically stronger when sentiment is subdued.

This pattern is consistent with recent evidence in the Korean market (Byun et al., 2023), which documents that certain behavioral anomalies intensify during low-sentiment periods rather than high-sentiment booms. One possible interpretation is that when overall sentiment is low, limits to arbitrage become more binding and capital constraints more severe, allowing salience-driven mispricing to persist more strongly. In contrast, during high-sentiment periods, broad-based optimism may compress cross-sectional differences, reducing the relative impact of salience on return predictability ^[27].

Similarly, the effect is stronger during high-uncertainty periods. Using VKOSPI as the uncertainty proxy and splitting months at the median, the equal-weighted spread is −1.01% (t-stat = −4.30) during high uncertainty, compared to −0.41% (t-stat = −1.31) during low uncertainty. The FF3 alpha is −0.97% (t-stat = −3.79) in high-uncertainty months versus −0.46% (t-stat = −1.73) in low-uncertainty months, and the FF4 alpha is −1.06% (t-stat = −3.91) versus −0.72% (t-stat = −2.59).

This finding aligns closely with Cheon and Lee (2018). When market conditions are volatile and future outcomes are difficult to predict, investors appear to rely more heavily on simple heuristics like salience to guide their decisions. Uncertainty makes it harder to assess fundamental values, causing investors to anchor more strongly on recent salient experiences. Moreover, during uncertain times, the emotional impact of extreme returns (whether positive or negative) becomes more pronounced, intensifying the salience-driven distortion in expectations.

In sum, the results are consistent with the behavioral interpretation and provide additional evidence that salience effects reflect investor psychology rather than rational risk premia. If the return spread represented compensation for risk, we would not expect it to vary systematically with sentiment and uncertainty in this manner. Instead, the patterns suggest that salience-driven mispricing intensifies when behavioral biases are likely to be strongest.

We conduct extensive robustness tests to verify that our main findings are not driven by specific methodological choices, sample selection criteria, or measurement issues. Table 11 reports the results of quintile univariate-sort analysis using alternative constructions of the ST variable.

Panel A examines sensitivity to the salience function parameters. Following Bordalo et al. (2012), our baseline specification uses θ = 0.1 and δ = 0.7. We verify robustness by varying these parameters within reasonable ranges suggested by experimental evidence. When we increase θ to 0.15 (decreasing sensitivity), the spread is −0.99% (t-stat = −4.51), virtually identical to the baseline. This demonstrates that our results are not driven by the specific choice of θ. Similarly, when we vary the attention decay parameter δ, the results remain robust. With δ = 0.6 (faster decay, more weight on the most salient day), the spread is −0.94% (t-stat = −4.44). The fact that the effect is strongest with lower δ values is consistent with investors focusing most intensely on the single most salient experience rather than averaging across multiple salient days. This pattern aligns with psychological evidence on recency and peak-end effects in memory.

Panel B examines the effect of different time horizons for constructing ST. Our baseline uses daily returns over the past month (approximately 20 trading days). When we extend the horizon to three months (quarterly), the spread is −1.09% per month (t-stat = −3.95). The smaller magnitude is expected because including more distant returns dilutes the impact of recent salient experiences. However, the effect remains highly significant, indicating that salience influences expectations even over longer horizons. When we extend to six months, the spread is −0.82% (t-stat = −3.10), still significant but substantially weaker. This pattern is consistent with investors focusing primarily on recent experiences, as predicted by salience theory and documented in the psychology literature.

Panel C shows the effect of different price filters. Our baseline sample excludes stocks priced below 1,000 Korean won. When we lower this threshold to 500 won, the high-low spread is −0.98% (t-stat = −4.43), slightly larger than baseline. When we set this threshold to 1,500 won, the high-low spread is −0.95% (t-stat = −4.46), slightly smaller than baseline. The robustness across different price thresholds indicates that our results are not driven by penny stocks or microstructure issues affecting very low-priced securities.

Finally, Panel D shows the results of the subperiod analysis. To ensure our results are not driven by specific time periods, we split our sample into two ten-year subperiods: pre–2015 and post–2015. The high-low spreads are −0.93% (t-stat = −4.48) and −0.66% (t-stat = −2.21), respectively. All subperiods show negative and statistically significant effects, demonstrating that our findings are not specific to any particular market regime or time period. The consistency across subperiods is particularly notable given that these periods encompass very different market conditions: the 2008 financial crisis, the post-crisis recovery, and the recent COVID-19 period.

This paper investigates the salience theory of Bordalo et al. (2012, 2013) in the unique institutional setting of the Korean stock market. Using data from 2005 to 2024, our comprehensive analysis confirms that salience is a powerful and robust predictor of the cross-section of stock returns. A portfolio that goes long on low-salience (ST) stocks and short on high-salience stocks generates a significant four-factor alpha of −1.50% per month. This effect is not only statistically and economically significant but is also distinct from well-known anomalies such as short-term reversal and the preference for lottery-like stocks.

More importantly, our study provides direct evidence on several channels through which this salience-driven mispricing is generated and sustained. First, we demonstrate that the salience effect is fundamentally driven by the trading of individual investors. The effect is muted or non-existent for stocks favored by more sophisticated institutional and foreign investors. Second, we provide causal evidence that limits to arbitrage allow this mispricing to persist. By exploiting Korea's history of market-wide short-sale bans as a quasi-natural experiment, we find that the ST effect is significantly stronger during periods when this critical arbitrage tool is unavailable. This result, combined with cross-sectional evidence showing the effect is more pronounced in small, illiquid, and high-volatility stocks, confirms that constraints on arbitrage are a key reason why salience-driven mispricing is not immediately corrected by rational investors. Furthermore, our analysis of market states reveals that the salience effect is stronger during periods of low sentiment and high uncertainty, times when arbitrage is generally considered riskier and costlier.

Several limitations and directions for future research merit mention. First, our sample is limited to one emerging market; testing salience effects in other Asian markets would enhance external validity. Second, micro-level analysis using individual investor account data could directly test whether retail investors with limited cognitive resources exhibit stronger salience-driven trading than sophisticated investors.

Overall, our results extend the salience-based asset pricing literature beyond developed markets and demonstrate that behavioral biases may be amplified in emerging market contexts. The persistence of these effects throughout our sample period indicates that behavioral biases remain economically important even as markets develop. These findings have implications for understanding return patterns in markets where limits to arbitrage prevent full efficiency and suggest that incorporating salience theory into asset pricing models can improve our understanding of cross-sectional return predictability.

The supplementary material for this article can be found online.