2.1 Data

Consistent with a large body of prior studies (O’Hara, 2014; Chen, 2018; Chen, 2019; Chen, 2020), our intraday tick data are taken from a transaction dataset compiled by Bloomberg from Jan to Dec 2016. Bloomberg renders real-time, historical and descriptive data, analytics, and news on a number of markets and securities in a single integrated platform around the world. Each record comprises the trading time, the transaction price, the number of shares traded, and the quotation code. To avoid massive computations if employing the full data set, a random sample of 100 traded stocks is constructed in each market.

Before enforcing a random selection, we drop the observation satisfying with the following conditions for each stock: (i) outside the normal trading period; (ii) negative prices or sizes; (iii) ask prices less than bid prices; (iv) trades beyond the spread; (v) trades without the normal quotation code; and (vi) prices beyond the range between 50 % and 150 % of prior traded prices.

To be eligible for inclusion in the final sample, we further impose additional screening criterion: (i) stocks must have more than one trade per minute; (ii) stocks prices must be greater than one; and (iii) stocks must be actively traded for more than one hundred days. Our final sample consists of 2,200 stocks across 22 countries.

2.2 Measure of retail trading

The Bloomberg dataset does not include any information on whether the trade is a buy or sell for the vast majority of countries, so the inference of trade direction is made in light of the tick rule. Compared to Lee and Ready’s (1991) procedure, the tick rule provides better estimates of signed volume (Finucane, 2000) and of the aggregate side of trading (Easley et al., 2016). Concretely, this rule determines the trade direction with current prices relative to prior ones. The trade is classified as a buy if the current price is higher than the last price (an uptick) or as a sell if lower (a downtick). If there is no price change between trades (a zero tick), the preceding price that differs from the current one is utilized to decide the trade direction.

Owing to the unavailability of the identity code, it is unlikely to identify retail traders in our dataset. To solve this problem, existing work proposes various approaches where transactions are assigned to retail trades if less than a certain threshold. While one strand of literature uses a preset cutoff point of 500 shares (Battalio and Mendenhall, 2005) or of $5,000 (Barber et al., 2009), the other develops a dynamic methods to calculate different firm-size cutoff points (Hvidkjaer, 2008). To improve the classification accuracy, our paper capitalize on the stock-specific dollar-based methodology to extract individual trades out of the dataset.

To make it work, we first determine the small-trade cutoff point as follows. Each day d, all trades of stock s are sorted based on their trading volumes. The 20^th percentile relating to trading volume (TV20sd) can be found, which serves the small-trade cutoff point. Subsequently, for stock s on day d, transactions with the trading volume less than TV20sd are used to proxy for individual investor trades in our study. After implementing such a procedure, roughly 20 percent of the trades in our sample belong to retail investors.

Descriptive statistics of small-trade cutoff points are presented for each country in Table 1. After adjusting with inflation rates and exchange rates (Barber et al., 2009), the mean cutoff points in every markets are generally comparable to a trade size of $5,000, which Lee and Radhakrishna (2000) suggest as an optimal benchmark to identify retail trades.

In line with previous researches (Barber et al., 2009; Kelley and Tetlock, 2013; Barrot et al., 2016), retail trading in this paper is measured by retail order imbalances, which equal the difference between the retail trading volume bought and the retail trading volume sold over the total retail trading volume.

2.3 Validation of retail trading measure

In this subsection, three validation tests are harnessed to verify whether our small-trade cutoff point is a good proxy for separating retail trades from the market. First, we examine the reliability of the assumption that individual market participation is positively related to retail trading. It serves as a necessary condition for the accuracy of the chosen measure. If there is no positive relationship between these two, our proxy is very unlikely to capture the overall activity of retail investors. We obtain retail participation ratios (populations) directly from Grout et al. (2009). Cross-country correlations are estimated in Table 2 Panel A, where country-based retail trading has a positive association with the participation ratio (0.270) and population (0.717). These preliminary findings are consistent with the assumption used to develop our measure of retail trading.

Second, we form a retail attention measure by using the Google Search Volume (GSV) for each stock with its Bloomberg ticker (Da et al., 2011). Next, we evaluate whether it is positively linked with our retail trading measure. No correlation suggests that our proxy is questionable. Table 2 Panel B provides direct evidence that the 20^th percentile of trading volume is a good proxy for retail trades. Specifically, contemporaneous correlations between retail trading and retail attention are shown to be positive in 17 out of 22 countries from the statistical perspective. Moreover, the cross-country average correlation is 0.205, with the significance at the 1 % level.

Inspired by Campbell et al. (2009), we finally check whether individual transaction under the proposed cutoff rule is effective in explaining the weekly change in retail ownership. Because data on retail shareholding are not available, we gauge the change in retail ownership by the change in institutional ownership on the ground that no others would be involved in the trading. Table 2 Panel C shows the average adjusted R² after running a stock-level regression of retail ownership changes onto its retail trading lagged level and the lagged change in retail ownership. Overall, these statistics range from 4.0 % (Taiwan) to 65.2 % (US) with a global mean of 13.8 %, which are greater than the 10.9 % reported by Campbell et al. (2009). Besides, there are 11 countries with an adjusted R² above 10 %, suggesting that our measure track retail trading. To some degree, all three tests justify the utilization of the 20^th trading volume percentile to identify retail trading.

3.1 State space model: Informed or noise traders?

The first analysis of price discovery depends on a trading environment where true prices (permanent components) are unobserved and observed traded prices are linked by measurement errors (transitory components). Therefore, the state space model is a natural choice (Menkveld et al., 2007) for exploring the interdependence among retail trading, permanent price changes, and temporary price changes. In concrete terms, the transaction price of a stock is represented by a nonstationary efficient element, which is represented as an unobservable state variable plus a stationary temporary element capturing the price error.

where ps,t is the transaction price after taking the log at time t for stock s, ms,t is the permanent portion of observed prices at time t for stock s, and ss,t is the transitory portion of observed prices at time t for stock s. In its simplest form, the efficient price is assumed to be a martingale process consisting of two components:

where ws,t is an innovation that affects the price permanently by reflecting the arrival of new information at time t for stock s. In line with Brogaard et al. (2014), the permanent price increment is further modeled to introduce the overall impact of retail trading.

where RTˆs,t is the residual of an autoregressive model with 10 lags on retail trading at time t for stock s. Its coefficient (αs) can be used to examine whether retail investors impart information that impacts permanent price changes. Positive (negative) coefficients indicate retail trading in the same (opposite) direction of permanent price changes. Moreover, εs,t is the firm-specific information disturbance irrelevant with retail trading, which follows a white noise process.

To quantify temporary price elements, we specify them as an equation encompassing an autoregressive component and retail trading as follows:

where RTs,t is retail trading (the difference between buyer- and seller-initiated trading volume standardized by the total volume). Hence, the coefficient (γs) can be used to evaluate the noise trader hypothesis.

The state space model conditional on 1-minute retail trading is estimated for each stock day. The estimated coefficients are averaged cross-sectionally for each country and then presented in Table 3. In terms of global impact, retail trading is positively associated with permanent price changes, indicating a positive contribution made by individuals to price discovery. Coupled with the presence of similar findings in individual countries, our results are consistent with the informed trader hypothesis.

However, we also find a positive and significant coefficient on RT in the equation for the temporary price element, demonstrating that retail investors on average move in the same direction of pricing errors. In addition, this transitory price change continues with an AR(1) coefficient of approximately 0.512. Therefore, the persistent pricing errors caused by retail trading seem to lend support to the noise trader hypothesis.

To sum up, while we discover that retail trading is so informative as to improve price efficiency, it has the side effect of introducing noises into the market. Our findings reconcile these conflicting hypotheses developed in the prior literature because empirical evidence in favor of either is sought in our study.

3.2 Weighted price contribution: Incorporation of new information

In the last subsection, we show that retail trading delivers information about future price movements. However, the level of retail trading activity as a whole remains below that of other market participants, such as institutional investors. To study the amount of price discovery throughout the trading session, we follow Barclay and Warner (1993) to gauge how much new information belonging to retail trades is incorporated into prices via the weighted price contribution.

Given that stock s has N trades on day d, either retail or other, we use the trade index (n) to denote the transaction sequence. Consequently, the weighted price contribution (WPC) attributable to retail trading on day d for each country with S stocks is defined as follows:

where φn is an indicator taking the value of one if trade n falls in the retail category, and rs,d,n is the price difference between trade n and trade n−1 for stock s on day d. The first term represents the ratio of absolute returns for each stock over the sum of absolute returns for all stocks. This weighting factor helps alleviate the heteroskedasticity in calculating WPC (Barclay and Warner, 1993). The second term reflects the relative contribution of retail trading to the total return on day d.

Table 4 presents the results on the WPC of retail trading. Several findings emerge from this table. First, on average, 24.8 % of the price discovery is contributed by retail trading around the world. In contrast with its aggregate share of volume (5.6 %), retail trades release more information content than their incidence from the statistical viewpoint. Second, a positive and significant WPC is observed in all countries, ranging from 1.3 % in Australia to 62.4 % in Singapore, indicating that informed retail trading takes place globally. Third, our WPCs are comparable to those reported in O’Hara et al. (2014), who focus on the odd-lot trades in the US. Combined, retail investors reveal information through their trading to move the price in international markets, in tandem with the informed trader hypothesis stated before.

3.3 Price discovery surrounding earnings announcements

According to the above analysis, retail trades carry information relevant with future price changes. However, it remains ambiguous as to what types of information retail investors can explore. Therefore, we aim to shed light on this question by examining an earnings announcement in which retail investors are capable of accessing information benefiting price discovery.

The purpose of an earnings announcement is to release firm-specific information that draws significant attention from various types of investors in the market. If individuals are well informed in advance, they would trade around financial disclosures aggressively using small quantities of shares. Meanwhile, retail trading faces less constraints in terms of diversification requirements or short sales compared to institutional counterparts. Hence, this reasoning points to a notion that retail trading around earnings announcements may matter to price discovery.

To proceed, we collect the announcement dates of our sample stocks from Bloomberg in 2016. Consistent with earlier studies, we uncover post-earnings-announcement drifts, suggesting traders’ failure to fully absorb the information surrounding events. Consequently, such a setting creates an opportunity to investigate how retail investors improve the price discovery process. If informed individuals do contribute to price discovery, their trading around financial announcements should positively predict the ensuing drifts. Otherwise, if individuals move prices beyond their efficient level following earnings disclosures and exacerbate inefficient price changes (i. e., overshooting), we expect a negative relation between retail trading and post-event returns. In the spirit of Kaniel et al. (2012), the subsequent model is specified to examine our conjecture.

where CARs,[2,11] is the cumulative abnormal return from 2 to 11 days post announcement for stock s, and CARTs,[−1,1] is the cumulative abnormal daily retail trading from 1 day before to 1 day after an announcement for stock i. Since many announcements happen either before market open or after market close (Johnson and So, 2018), we set the event window as one day before and after the announcement date.

Next, the pooled regression is performed for all stocks with announcements in each country with the results tabulated in Table 5. Overall, the cross-country coefficient on retail trading (0.112) is positive and significant at the 1 % level, suggesting that individuals move in the same direction as subsequent price changes. It is equivalent to implying that retail investors are able to acquire information released from the announcement and continue to push prices toward efficient levels by their trading.

When looking at individual markets, retail trading is documented to exert a positive impact on price discovery across the board. In particular, this positive relation is witnessed in the US, parallel to Kaniel et al. (2012) who show that pre-announcement retail buying predicts post-announcement returns. To summarize, our results regarding earnings announcements seek favorable evidence in line with the view that retail trading facilitates the incorporation of new information from the scheduled event rather than triggers the transitory price movement.

3.4 Unscheduled news release

Besides the scheduled earnings announcement, retail investors are likely to gather information from an unscheduled event. In order to look for an unscheduled news release, we take advantage of the weekend effect as an alternative scenario. In other words, stock prices may alter dramatically after a weekend close due to the disclosure of unscheduled news. Compared to weekday overnight periods, the weekend poses increased uncertainties and heightened inventory risks (French, 1980). Thus, the weekend effect offers a unique platform to ascertain the role of retail trading in price discovery. If informed individuals continue buying (selling) on Friday in anticipation of good (bad) news over the weekend, it leads to a positive relation between their trading and weekend returns, which is identical to the prevalence of price discovery. Conversely, a negative or insignificant association illustrates the absence of price discovery. To figure out this problem, we perform the following regression for each stock:

where rs,wkd (mrs,wkd) is stock s’s weekend (market) return, calculated by the difference between the closing price (benchmark index) on Friday and the opening price (benchmark index) on the coming Monday, and ARTs,Fri is the abnormal last-hour retail trading on Friday for stock s.

The regression results of Eq. (7) are given in Table 6. As expected, the cross-country mean coefficient pertinent to retail trading is 0.044 and statistically different from zero, substantiating the aforementioned hypothesis that retail investors are conducive to reflect the impact of unscheduled news over the weekend into prices before its arrival. That is to say, individual trading plays a crucial role in accelerating the price discovery process. Furthermore, this coefficient of interest is positive and significant in all markets. A relatively high adjusted R2 (0.199) indicates that our model has an explanatory power for the weekend return. In a nutshell, we demonstrate evidence that retail investors are beneficial for the price discovery process when confronting scheduled news (earnings announcements) or unscheduled information (weekend effects).