A. Household Transactions and Demographics Information

This study uses two data sets to analyze the trading behaviors of households. The first data set contains transactions for a subset of individual investors in the United States, while the second contains transactions of all investors in Finland. The individual trade data for the United States come from a major U.S. discount brokerage house that recorded the daily trades of 78,000 households from January 1991 to December 1996. This is the same data set as used in Barber and Odean (Reference Barber and Odean2000).Footnote ⁷ We focus only on the common stock transactions of households in this study, which account for nearly two thirds of the total value of household investments. We exclude from the current analysis investments in mutual funds, American Depositary Receipts (ADRs), warrants, and options.

Our final sample includes over 66,000 households with close to 800,000 transactions. The data set includes for each transaction, the number of shares traded, the transaction price, and value of the position at market close. The data set also includes demographic information for a smaller subsample of households (37,664 households), such as income, age, gender, occupation, and marital status.

To address concerns that our findings may be specific to the data and sample period we study, we repeat our analyses using an alternative transaction data set from Finland. This data set comes from the central register in the Finnish Central Securities Depository (FCSD). The register officially records all the trades of all Finnish investors (both individual and institutional) daily from January 1995 to December 2003. Compared to the U.S. data set, the Finnish data set has better coverage as it includes the complete trading records of all market participants rather than a subset of market participants. For the purposes of this study, we ignore institutional trades and utilize only the trades of individual investors in Finland. Like the U.S. data set, the Finnish data set reports each transaction, the number of shares traded, the trading price, and the daily closing price. We can also observe the initial holdings for each account at the beginning of the sample period, which allows us to keep track of the daily holdings of households. While the data set reports demographic information such as age and gender for a subset of investors, it does not include information about income, occupation, and marital status. A more detailed description of the Finnish data set can be found in Grinblatt and Keloharju (Reference Grinblatt and Keloharju2000), (Reference Grinblatt and Keloharju2001). To calculate stock and firm characteristics for the Finnish stocks, we obtain data from Datastream.

We use the following investor characteristics in our analyses: investor age (Age), log of annual income in dollars (Log (Income)); a dummy variable that is equal to 1 if the trader is married (Married Dummy); a gender dummy that is equal to 1 if the trader is male (Male Dummy), a dummy to capture if the trader holds a technical or managerial position (Professional Dummy); a dummy that takes on the value of 1 if the trading account is a retirement account (Retirement Acct Dummy); and a dummy that equals 1 if the trader is retired (Retired Dummy).

We also identify certain trader characteristics from each household’s trading history and define the following control variables: Foreign Securities Dummy, equals 1 if the household has ever traded foreign securities; Option User Dummy, equals 1 if the household has ever traded options; and Short User Dummy, equals 1 if the household has ever held a short position. We also estimate the log of the average total dollar value of each household’s equity investments, Log (Equity Portfolio Value). Finally, we estimate the concentration of each household’s portfolio (Portfolio Concentration) computed as the sum of the squared value weights of each stock in a household’s portfolio following Ivkovic et al. (Reference Ivkovic, Sialm and Weisbenner2008). We calculate the value of equity investments monthly and calculate all the other trade and investor characteristics over the full sample period.

B. Measures of Transaction Costs and Firm Controls

Transaction costs are multifaceted and are usually defined in terms of the costs and risks associated with trading financial securities. These costs incorporate price impact, asymmetric information, and inventory risk. A number of different measures of transaction costs have been proposed and used in the literature. Instead of relying on a single measure, we use several different measures that are commonly used in previous papers and can be estimated for both the U.S. and Finland data sets.

The first measure is the Amihud illiquidity ratio ( $ Illiq $ ) from Amihud (Reference Amihud2002), calculated as follows:

(1) $$ {Illiq}_{i,t}=\frac{1}{D_{i,t}}\sum \limits_{d=1}^{D_{i,t}}\frac{\left|{r}_{i,d}\right|}{dvol_{i,d}} $$

where $ {r}_{i,d} $ is the daily return for stock i in day d. $ {dvol}_{i,d} $ is the dollar volume for stock i in day d. $ {D}_{i,t} $ is the number of trading days in month t. The Amihud measure is similar to Kyle’s lambda and captures the price impact of trades over a specific time period. Following Acharya and Pedersen (Reference Acharya and Pedersen2005), we adjust the Amihud measure as in the following to remove outliers and to make it stationary: $ AdjIlli{q}_{i,t}=\min \left[0.25+30\times {Illiq}_{i,t}\times {M}_{t-1},30\right] $ , where $ {M}_{t-1} $ is the ratio of the capitalizations of the market portfolio at the end of the month t-1 to that of the market portfolio in July 1962. The higher the adjusted Amihud ratio, the more illiquid the stock is.

The second measure uses the proportion of trading days with zero returns (Zerofreq) to capture transaction costs. Following Lesmond, Ogden, and Trzcinka (Reference Lesmond, Ogden and Trzcinka1999), we compute the proportion of days with zero returns for each stock each year as Zerofreq. The higher the Zerofreq, the more illiquid the stock is.

We also compute a number of measures using intra-day trades for the U.S. sample. We use a 5-second delay to match trades with quotes and apply the same filters discussed in Hvidkjaer (Reference Hvidkjaer2006). Effective Spread/Price is the difference between the transaction price and the quoted bid–ask midpoint multiplied by 2 and scaled by the transaction price. Relative Spread/Price is defined as the quoted bid–ask spread divided by transaction price, and Relative Spread/Mid is defined as quoted spread scaled by the bid–ask midpoint. Depth is defined as the midpoint of bid size and offer size (both in number of round lots). As depth tends to be skewed, we use log(1 + depth) in our analyses. To reduce potential endogeneity arising from contemporaneous measurement and to smooth out idiosyncratic changes, we use the 12-month moving average of each liquidity measure in our analyses. The liquidity measures are calculated at the end of the month prior to transactions. For instance, if an individual investor has a sell transaction on May 15, the liquidity measures would be calculated as of Apr. 30.

Finally, we use the actual trades of investors to measure realized transaction costs following Barber and Odean (Reference Barber and Odean2000). We estimate the closing price spread (Closing price Spread %) for purchases as the negative of the closing price from CRSP divided by the transaction price minus 1. Closing price spread for sales is equal to the closing price from CRSP divided by the transaction price minus 1. We also calculate commissions (Commission %) as the amount charged by the brokerage for the trade scaled by the dollar value of the trade. In the analyses, we use the sum of commissions and purchase spread (Closing price Spread % + Commission %).

We control for a number of firm characteristics in the analyses. These are firm size measured by the log of market capitalization (Size), B/M, momentum calculated using returns over the past 12 months excluding the previous month (Momentum), idiosyncratic volatility (Ivol), maximum daily return over the past 1 month (MaxPrc), and the CAPM Beta (Beta). We also control for Unrealized Returns as (selling price ‒ purchase price) / purchase price to capture potential disposition bias. In calculating Unrealized Returns, if a sale is never observed and the sale price is unavailable, we use the stock price on the last day of our sample period. With the exception of unrealized returns, which are calculated on the transaction date, all other stock characteristics are calculated as of the end of the month prior to the transaction. For instance, if an individual investor has a sell transaction on May 15, the stock characteristics would be calculated as of Apr. 30.

Table 1 reports the summary statistics for stock and investor characteristics for the United States. Panel A reports descriptive statistics for stocks that are traded by households in the data set. For comparison, Panel B provides descriptive statistics for the CRSP stock universe during the same sample period. Summary statistics are calculated by pooling annual stock-level observations from 1991 to 1996. Panels A and B show that the price, size, B/M, and past returns for stocks in our sample are similar in magnitude to those in the entire CRSP universe. For example, the median, 25th percentile, and 75th percentile prices are the same for our sample of stocks and for those in the CRSP universe. The average B/M for our sample of firms is 0.78, which is slightly higher than the average B/M of 0.72 for the CRSP universe, while the median is 0.57 for our sample and 0.56 for the CRSP universe. The average and median sizes of our sample firms are also slightly larger than those of the CRSP universe. The differences between the 2 samples and their statistical significance are reported in the last column in Panel B. The transaction costs measures are marginally lower in our sample compared to the larger CRSP universe. Overall, the differences are economically small, indicating that our sample of stocks is representative of the entire stock market during the sample period.

Table 1 Summary Statistics of Stock and Investor Characteristics in the United States

Panel C reports the summary statistics for the U.S. individual investor characteristics. The majority of the investors are in their 40s and 50s, with an average (median) age of 49.58 (48), and 15% of the investors are retired. Only 10% of the primary U.S. account holders for the transactions analyzed in this study are female, 76% are married, and 66% hold technical or managerial positions. The mean (median) portfolio value is $80,342 ($22,952) for the households analyzed in this study, and the mean (median) annual income is $76,840 ($87,500) for these investors over the sample period. In addition, 14% of the households have traded options, 22% have traded foreign securities, and 38% have held a short position at some time over the sample period analyzed. The mean (median) U.S. individual investor’s portfolio concentration is 0.52 (0.48), which roughly corresponds to holding two stocks with equal weights.

A. Holding Periods and Transaction Costs

In this section, we provide empirical evidence in support of our Hypothesis 1. We begin by computing a holding period for each transaction in the data set. The holding period for a transaction is defined as the number of trading days from the first purchase to the first sale of that stock, following the approach of Seru, Shumway, and Stoffman (Reference Seru, Shumway and Stoffman2010). This generates 799,469 holding period observations, with a median (mean) of 207 (550) trading days for retail investors in the United States.

We begin our analyses by sorting stocks into two broad transaction cost groups each year based on our main transaction cost measure, namely the adjusted Amihud illiquidity measure. One group consists of stocks in the highest transaction cost decile, while the other group comprises the rest of the stocks in the other 9 deciles. We plot Kaplan–Meier survival probabilities for these two broadly defined groups of stocks in Figure 1. The x-axis shows the number of days that have passed since the purchase of a representative stock in each group, while the y-axis represents the probability that the investor will continue to hold this representative stock conditional upon no sale up to that point in time. The solid line plots the survival probability of a representative stock in the highest transaction cost decile, while the dashed line graphs the survival probability of a representative stock for the other 9 deciles. In Figure 1, Graph A, we plot survival probabilities for stocks in the United States, while in Graph B, we plot survival probabilities for stocks in Finland. Investors are more likely to sell stocks with lower transaction costs as the survival probabilities are lower for these stocks. The figures provide preliminary evidence that holding periods are strongly related to measures of transaction costs.Footnote ⁸

Figure 1 Survival Probabilities for Stocks in the United States and Finland

Figure 1, Graph A plots the Kaplan–Meier survival probabilities for two groups of stocks held by households in the United States over the 1991–1996 time period. Illiquid stocks in the figure are stocks that belong to the top decile based on their adjusted Amihud illiquidity measure. The solid line represents the probability of holding onto these illiquid stocks, and the dashed line represents the probability of holding all of the other stocks. Figure 1, Graph B plots the Kaplan–Meier survival probabilities for two groups of stocks held by households in Finland over the 1995–2003 time period. Illiquid stocks in the figure are stocks that belong to the top decile based on their adjusted Amihud illiquidity measure. The solid line represents the probability of holding onto these illiquid stocks, and the dashed line represents the probability of holding all of the other stocks.

We use a hazard model to analyze the relationship between holding periods and transaction costs, controlling for the confounding effects of stock and investor characteristics.Footnote ⁹ Specifically, we model investors’ sell versus hold decision using a Cox proportional hazard model with time-varying as well as static explanatory variables. The hazard model takes the following form:

(2) $$ h(t)={h}_0(t)\mathit{\exp}\left({\beta}^{\prime }X+{\theta}^{\prime }{Z}_t\right) $$

The left-hand side variable, $ h(t) $ , is the hazard rate, the probability of selling a stock on day t conditional upon holding that stock until that point (t) in time. X is a vector of explanatory variables which are static and do not change over time. Z_t represents a vector of time-varying covariates which can take on different values at different points in time.Footnote ¹⁰ $ {h}_0(t) $ is the baseline hazard rate and describes the hazard rate when the independent covariates are all equal to 0. Using the Cox (Reference Cox1972) estimator, we can estimate coefficients on X and Z_t without specifying a baseline $ {h}_0(t) $ hazard rate. Positions that are not closed by the end of the sample period are treated as censored observations.

We control for investor characteristics that are directly observable such as age, income, gender, marital status, employment status, and occupation, as well as another set of less readily observable variables that are extracted from investors’ positions and trades, such as the total wealth invested in their portfolios and whether the individual investors ever short stocks, trade options, or trade foreign securities. We also control for size, B/M, and momentum to account for investors’ preferences for stocks with certain characteristics that are known to be associated with expected returns.

As there is also likely to be seasonality in purchases and sales, we further include calendar year and month dummies in the hazard regressions. Open stock positions, for instance, may be closed out in December for tax reasons. Finally, we use unrealized gains/losses as a control variable. Although momentum does capture the effect of past returns on trading decisions, unrealized gains and losses for each individual investor could be different based on the original purchase price.

Table 2 reports the results from the hazard regressions. Following standard reporting conventions, we report the hazard ratios instead of the estimated coefficients. The hazard ratio is similar to the odds ratio estimated from a binary choice model and is defined as the ratio of two hazard rates when one explanatory variable is changed by 1 unit from 0 holding all other variables constant. A hazard ratio of less than 1 would suggest that the explanatory variable reduces the probability of selling the stock. In contrast, a hazard ratio larger than 1 would suggest that a higher exposure to the explanatory variable would increase the likelihood of selling the stock, thus reducing the likelihood that the investor would continue holding on to the stock.

Table 2 Impact of Transaction Costs on Households’ Holding Periods in the United States, Hazard Analysis

In Table 2, we report the results using only the adjusted Amihud illiquidity ratio for all specifications. We provide results for alternative measures of transaction costs in Table 3. Column 1 of Table 2 shows that the estimated hazard ratio for the adjusted Amihud illiquidity ratio is 0.981 when we do not control for stock or investor characteristics. It is less than 1 and statistically significant, suggesting that the sale probability of a stock declines with higher transaction costs. Specifically, the average investor would be 9.3% less likely to sell a stock in the 75th percentile in terms of illiquidity compared to a stock with a median level of illiquidity using the adjusted Amihud illiquidity ratios.Footnote ¹¹

Table 3 Alternative Transaction Costs on Households’ Holding Periods in the United States, Hazard Analysis

As households could have different preferences and potentially have different holding periods, we control for heterogeneity across households within the hazard framework. We assume different baseline hazard rates for each household and estimate a model with partial likelihood stratification. The household-level stratification allows for the possibility of each household having a different baseline holding period, which is analogous to using household fixed effects in OLS regressions. Similarly, we use firm stratification to allow for the possibility that each stock has a different average holding period. In column 2 of Table 2, we calculate hazard ratios using firm and household stratifications to account for household and firm fixed effects. The estimated hazard ratio for the adjusted Amihud measure (AdjIlliq) is 0.973 and statistically significant, consistent with earlier results. Controlling for household and firm-level fixed effects suggests that a 1-standard-deviation increase in the adjusted Amihud illiquidity ratio would reduce the sale likelihood by 18.5%.

Controlling for heterogeneity among households and stocks leads to stronger results as the hazard ratio is reduced from 0.981 to 0.973. To better understand the source of this variation, we run a regression of holding periods on household and stock fixed effects. We find that household fixed effects explain about 35% of the cross-sectional variation in holding periods, while stock fixed effects explain about 18% of the variation. These results suggest that both household and stock fixed effects influence holding periods and that households also differ in their baseline holding periods.

We examine in detail how specific stock and investor characteristics affect households’ trading decisions. We add stock characteristics first in column 3 of Table 2, and then further control for investor characteristics and unrealized returns in column 4. Since demographic information is only available for a subset of investors in the data set, the number of observations reported is lower in column 4. Our initial finding on transaction costs is unchanged with these additional controls. The loading on the AdjIlliq in column 3 is still less than 1 at 0.981 and statistically significant. The estimated hazard ratio for momentum is statistically significant and larger than 1 (1.135), which indicates that investors are more likely to sell recent winners. More specifically, a 1-standard-deviation increase in the past 10-month momentum returns (from month-12 to month-2) would increase the probability of sale by 30.6%. The estimated hazard ratio for size is 0.649 and that for the B/M is 0.681, both of which are less than 1 and are economically and statistically significant, suggesting that U.S. individual investors tend to hold large and value stocks for longer periods.

In column 4, we control for Unrealized Returns, to account for the impact of disposition effect—the tendency of individual investors to hold on to losing stocks for too long and to sell winners too quickly on our results. Our basic inferences regarding the impact of transaction costs on retail investors’ holding periods are unaffected when we control for unrealized returns along with trader demographics and trade characteristics. The estimated hazard ratio on AdjIlliq is 0.975, comparable to our findings in the initial 3 columns. The coefficient on Unrealized Returns is statistically significant and greater than 1 (1.134), suggesting that retail investors are more likely to sell shares that have higher unrealized returns. This finding is consistent with the disposition bias documented in the literature.

For robustness, we repeat our main analyses by excluding holding periods less than 2 days to remove day traders. Specifically, in column 5 we repeat our analyses in column 4 by excluding observations with 1- and 2-day holding periods. Our main findings are unaffected by this restriction.Footnote ¹²

For robustness, we also control for additional variables that prior studies have shown to affect individual investor trading decisions. Prior studies have shown that individual traders tend to buy attention-grabbing stocks.Footnote ¹³ To control for investor attention, we add stock characteristics that are positively correlated with investor attention to our baseline hazard regression. The stock characteristics we use are idiosyncratic volatility (Ivol), maximum daily return over the past 1 month (MaxPrc), and CAPM Beta (Beta). The hazard regression results with these controls are reported in column 6 of Table 2. Beta and Ivol have statistically significant hazard ratios of greater than 1, 1.111 and 2.807, respectively, while the estimated hazard ratio for MaxPrc is almost 1 and statistically insignificant. These results are consistent with investors trading attention-grabbing stocks more frequently. The estimated hazard ratio in column 6 for AdjIlliq remains significant at 0.977 and is similar in magnitude to the hazard ratio reported in column 4.

We conduct an alternative analysis in Table A1 in the Supplementary Material Section A.1. Specifically, we rank all stocks by the Amihud illiquidity ratio and create a dummy variable (AdjIlliq Dum) that takes on a value of 1 if stock belongs to the highest illiquidity quintile. The use of the dummy variable makes it easier to interpret our results in Table 2. We find in Table A1 in the Supplementary Material that in the United States a stock in the highest transaction cost quintile is 20% less likely to be sold than a stock that has lower transaction costs but with similar firm and investor characteristics, consistent with other results in the paper as well as with the predictions of theoretical models of liquidity.

We repeat our analyses using six alternative measures of illiquidity described earlier, namely, Zerofreq, Closing price Spread + Commission (%), Effective Spread/Price (%), Relative Spread/Price (%), Relative Spread/Mid (%), and Log (1 + depth). In Table 3, we repeat our analysis conducted in column 1 of Table 2 using these alternative measures. All estimated hazard ratios in columns 1 through 6 are less than 1 and statistically significant. Our results are qualitatively similar regardless of the illiquidity measure we use. The economic significance levels of these variables are also similar to those using the adjusted Amihud measure. For example, the estimated hazard ratio for Closing price Spread + Commissions (%) reported in column 2 is 0.945. This suggests that a 1-standard-deviation increase in Closing price Spread + Commissions (%) would lead to a 15.37% reduction in the average household’s likelihood to sell.

It is possible that individual investors may care more about the trading costs incurred at the time of purchase rather than at the time of sale. To better understand whether there is an asymmetry in how purchase and sale transaction costs incurred are incorporated in holding period decisions, we investigate the impact of buy and sell transactions separately. Following Odean (1999), we calculate the closing price spread for purchases and sales separately for each trade. These results are reported in Table A3 in the Supplementary Material Section A.3. We repeat the analyses for Finland and report the relevant results in Table A4 in the Supplementary Material Section A.4. The coefficients on the purchase and sale transaction costs are similar in magnitude. Overall, these results are consistent with investors incorporating transaction costs incurred both at the time of purchase and at the time of sale.

B. Investor Sophistication

In this section, we investigate the impact of heterogeneity across households on the relationship between transaction costs and holding periods of investors. We provide empirical evidence in support of our Hypothesis 2.

Following Goetzmann and Kumar (Reference Goetzmann and Kumar2008), we assume that financial sophistication is correlated with education and resources available to each investor. We create a sophistication measure based on household and trade characteristics. Specifically, we use seven criteria to construct our sophistication measure (Sophistication), which increases by 1 with each of the seven criteria being met. The criteria include: if the investor has income greater than $75 K; if the investor works in a technical or managerial position (Professional Dummy = 1); if the investor is ranked among the top 25% of all investors in terms of total equity holdings; if the investor has ever traded an option (Option User Dummy = 1); if the investor has ever traded in foreign securities (Foreign Securities Dummy = 1); if the investor has ever shorted any equity (Short User Dummy = 1); and if the investor’s portfolio concentration is greater than 0.48, the median investor’s level of portfolio concentration. The last criterion is based on findings in Ivkovic et al. (Reference Ivkovic, Sialm and Weisbenner2008), who propose and empirically document that investors who hold more concentrated portfolios are financially more sophisticated as they possess informational advantages that allow them to outperform investors with diversified portfolios. The value of Sophistication ranges from a minimum of 0 for the least sophisticated investors to a maximum of 7 for the most sophisticated investors.

We sort investors into three groups based on their sophistication scores. Group 1 includes those with sophistication scores between 0 and 2, Group 2 is for investors whose sophistication scores are between 3 and 5, and Group 3 contains the most sophisticated investors with scores of 6 or 7. We then run a separate hazard regression for each of these three sophistication groups and examine how the relationships between transaction costs and holding periods change among investors with different levels of sophistication. Since many of the demographic variables and trade characteristics are used to calculate the sophistication score, these variables are not included as independent variables in our analyses in Table 4. Columns 1 to 3 of Table 4 report the estimated results for the least sophisticated group of households, the medium sophistication group, and the most sophisticated group, respectively.

Table 4 Transaction Costs and Holding Periods for Investors of Various Sophistication

We find that the coefficient on the adjusted Amihud illiquidity measure, AdjIlliq, is significantly positive and less than 1 for all sophistication groups. The estimated hazard ratios decrease monotonically from 0.984 for Group 1 to 0.975 for Group 2, and to 0.948 for Group 3 (the most sophisticated households). The estimated hazard ratio for AdjIlliq is 0.984 in column 1, indicating that an investor in the least sophisticated group would be 0.921 as likely to sell the stock in the 75th percentile of AdjIlliq as a stock with a median AdjIlliq. This would make them 7.9% less likely to sell. Similarly, the estimated hazard ratios for AdjIlliq in columns 2 and 3 would suggest that retail investors in Group 2 would be 0.879 as likely and those in Group 3 would be 0.762 as likely to sell the stock in the 75th percentile of AdjIlliq as a stock with median AdjIlliq. In other words, retail investors in the medium sophistication group would be 12.1% less likely to sell, while those in the most sophisticated group would be 23.8% less likely to sell when transaction costs increase. Overall, these results are consistent with Hypothesis 2 that financially more sophisticated investors pay closer attention to the impact of transaction costs when they trade.

A. Stock Splits

An extensive literature documents a significant reduction in transaction costs and improved liquidity subsequent to stock splits (Conroy, Harris, and Benet (Reference Conroy, Harris and Benet1990), Desai, Nimalendran, and Venkataraman (Reference Desai, Nimalendran and Venkataraman1998), Kryzanowski and Zhang (Reference Kryzanowski and Zhang1996), and Schultz (Reference Schultz2000)). There is also evidence of a positive abnormal return reaction on the split announcement day for splitting firms and findings that post-split performances of splitting firms are statistically indistinguishable from those of similar non-splitting firms in the long run (see, for instance, Byun and Rozeff (Reference Byun and Rozeff2003)).

We first verify empirically that stock splits indeed increase liquidity and reduce transaction costs. We identify a total of 3586 stock splits that took place in the United States between 1991 and 1996 for our sample. We remove reverse splits and splits that have a split factor of less than 0.25 (717 in total). Our final sample includes 2869 forward split events.Footnote ¹⁴ Consistent with the findings in the literature, in Table A6 in the Supplementary Material, we show that there is a significant decline in our main measure of transaction costs (AdjIlliq) after a split event for stocks in our sample.

If investors hold illiquid securities for longer periods, then the reduction in transaction costs after stock splits should lead to shorter holding periods. We examine individuals’ trading behavior over the same 6-, 9-, and 12-month periods after a split event using a dynamic hazard regression framework. To construct the appropriate data set for the dynamic hazard regression, we split the duration of a position into multiple periods.Footnote ¹⁵ The first period covers the time period before the split event. In this first period (pre-event), we assign a value of 0 to the After-Split Dummy. The second part is the time period from the split until the end of the event window of interest (i.e., three windows with a length of 6, 9, and 12 months). For the second period, After-Split Dummy takes on a value of 1. The third period corresponds to the time period after the split window (post-event), during which the After-Split Dummy takes on a value of 0.Footnote ¹⁶ For non-splitting stocks, the After-Split Dummy is always 0o. In the analyses of forward splits, we exclude reverse splits.

Since it is possible for transactions to be open 6, 9, or 12 months after a split, this setup ensures that After-Split Dummy will only equal 1 when a sale event falls within the event window, and as time elapses to the post-event window period, After-Split Dummy will switch back to 0. After-Split Dummy captures the marginal impact of stock splits on sale decisions over a distinct event horizon. Since the baseline hazard rate in the Cox regression model captures the increasing probability of a sale as time passes, After-Split Dummy captures the marginal impact of being in the split window period on the probability of a sale and does not simply capture a mechanical relationship due to the fact that the probability of a sale increases as time passes on.

Table 5 reports the estimated results of dynamic hazard regressions. All the regression models control for stock characteristics: size, book-to-market, and momentum, as well as calendar year and month effects. It is possible that split-event returns may lead to second-order effects that may influence investors’ trading decisions. To control for the impact of post-split returns, we calculate split-event returns for each period and control for these returns in models 2, 3, 5, 6, 8, and 9 in the table. Finally, we account for the possibility that stock splits may lead to clientele effects: forward (reverse) splits may attract clienteles that prefer lower-priced (higher-priced) equities. Columns 3, 6, and 9 address the clientele issue by controlling for stock prices at the time of sale. If no sale takes place until the end of the data set, then we use the last observed stock price.

Table 5 Impact of U.S. Stock Splits on Holding Period Decisions

Given the reduction in transaction costs after stock splits, we expect households to be more likely to reduce their holding periods, and thus we expect the estimated hazard ratio of After-Split Dummy to be greater than 1 in all specifications. We find that the estimated hazard ratio for After-Split Dummy is indeed greater than 1 and statistically significant at the 1% level for all specifications (see Panel A of Table 5). The estimated hazard ratio for After-Split Dummy in model 1 is 1.161, indicating that investors are 16.1% more likely to sell a stock in the 6 months after its split, controlling for other stock characteristics. Our results are robust across different event windows: the hazard ratio for After-Split Dummy takes on a statistically significant value of 1.171 for the 9-month window analysis in column 4, and 1.175 for the 12-month window analysis in column 7. These results suggest that investors are 16.1% to 17.5% more likely to sell their stock holdings within the first year after the split.

Next, we repeat our analyses conducted in Panel A of Table 5 for reverse splits. Reverse splits are much rarer compared to forward splits. In these analyses, we exclude forward splits. Panel B of Table 5 reports the estimated results of the dynamic hazard regression using reverse split events instead of forward split events. All regression models in the table control for calendar year and month-specific effects as well as for stock characteristics (i.e., size, book-to-market, and momentum). Models 2, 3, 5, 6, 8, and 9 further control for post-reverse split returns, while models 3, 6, and 9 control for stock price values at the time of sale to account for potential clientele effects.

Given the increase in transaction costs after reverse stock splits, we would expect households to increase their holding periods, and thus we would expect the estimated hazard ratio on After-R-Split dummy to be less than 1. In all specifications, we find that the estimated hazard ratio for After-R-Split dummy is less than 1 and economically and statistically highly significant. For example, the estimated hazard ratio for After-R-Split dummy in model 1 is 0.491, indicating that investors are 50.9% less likely to sell a stock in the 6 months after its reverse split. In Panel B, we exclude forward splits from our analyses, as including them would artificially strengthen our findings.

B. AMEX Tick Size Change

On Sept. 3, 1992, the American Stock Exchange (AMEX) reduced its minimum price increment from 1/8^th of a dollar to 1/16^th of a dollar for stocks priced between $1 and $5. One of the motivations for this change was to reduce bid–ask spreads. A number of papers document that both quoted and effective spreads declined subsequent to this change (Ahn et al. (Reference Ahn, Cao and Choe1996), Crack (Reference Crack1996)). This quasi-exogenous shock to transaction costs presents us with another opportunity to assess our main hypothesis and address potential endogeneity issues.

To investigate the impact of this event, we use a similar approach as we used above to examine the impact of stock splits on investors’ holding periods. We examine event windows of 6 months, 9 months, and 12 months after the implementation of the new tick size rule. We create a dummy variable (After-AMEX tick change dummy) that takes on a value of 1 if a month falls within the 6-, 9-, or 12-month period.

Unlike splits, which are staggered over time, the AMEX tick size change event occurred at a single point in time. To control for potential confounding market-wide factors, we compare the change in the affected stocks that were priced $1 to $5 listed on AMEX to three groups of control stocks. That is, in addition to calculating the change in the holding period for stocks affected by the tick size change (i.e., treated) before and after the event, we also calculate the change in holding period for non-affected stocks (i.e., control) during the same time period. We then compare the change in the holding period for treated firms to the change in the holding period for non-treated (control) firms. The first control group contains firms that were priced between $1 and $5 but listed on the NYSE and NASDAQ exchanges and as such were not affected by the tick size change. The second group contains all firms on AMEX that were priced $5 or more. The third group contains all non-affected stocks on the 3 major exchanges.

The dynamic hazard regression results controlling for stock characteristics are reported in Table 6. Panels A, B, and C report the results for 6-, 9-, and 12-month event windows, respectively. The coefficients reported under the row heading “Treated” are the hazard ratios estimated from the hazard regression on the interaction of the Treated group indicator with the After-AMEX tick change dummy, Treated × After-AMEX tick change dummy, while the coefficients reported under the row heading “Control” refer to the coefficient from the hazard regression on the interaction of the Control group indicator with the After-AMEX tick change dummy, Control × After-AMEX tick change dummy. The row with the header “Treated – Control” reports the difference in hazard rates between these two sets of interaction variables. In column 1 of each of Panels A, B, and C, we focus only on firms listed on AMEX that were priced between $1 and $5 and thus directly impacted by the tick-size rule change. Specifically, we investigate the holding period decisions of investors for these impacted stocks and find that investors’ likelihood of selling their impacted shares significantly increased after the tick-size reduction. For example, in Panel A, we observe that the estimated hazard ratio for the After-AMEX tick change dummy in model 1 is 1.167, indicating that investors were 16.7% more likely to sell an impacted stock (AMEX stock priced $1 to $5) in the 6-month period after the tick-size change. The results are qualitatively similar in Panels B and C, using 9-month (1.164) and 12-month (1.125) event windows, respectively.

Table 6 Holding Period Changes around Sept. 3, 1992, AMEX Tick Size Changes

In columns 2 through 4 for Panels A, B, and C, we investigate the differential impact of the tick-size rule change on investors’ holding periods of impacted stocks (i.e., AMEX stocks with prices between $1 and $5) as well as those of different sets of control firms. The regression includes controls for size, book-to-market, momentum, and unrealized returns. In column 2, we use stocks that are similarly priced (with prices between $1 and $5) but listed on the NYSE and NASDAQ exchanges that were not affected by the tick size change as the control group. We find that sale probabilities of investors in treated stocks increase significantly by 18.7%, 20.5%, and 17.4%, respectively, in Panels A, B, and C after the tick size change. However, the sales probabilities of stocks priced similarly but listed on NYSE and NASDAQ increase by much smaller magnitudes, specifically, 6.7%, 12.6%, and 14.1%, respectively, in the 6, 9, and 12 months after the event. The differences between the coefficients in columns 2 of Panels A, B, and C for treated and control groups are always statistically significant at the 1% level, with meaningful economic differences.

In column 3, we compare the differential impact on holding periods for treated firms versus all other AMEX unaffected stocks. Similarly, we find that sale probabilities of retail investors in treated stocks increase by 12.5% throughout Panels A, B, and C, while sale probabilities of the rest of the AMEX stocks only increase by 2.6%. 3.7%, and 3.4% during the 6-, 9-, and 12-month windows, respectively. The differences in increases of sales probabilities are again significant both economically as well as statistically across the three event windows.

Finally, in column 4, we investigate the differential impact of the tick-size change on retail investors’ holding periods for treated firms versus all other non-treated stocks (including all stocks on NYSE and NASDAQ, as well as AMEX stocks that are priced more than $5). We find that sale probabilities of retail investors in treated stocks increase by 11.8%, 12.9%, and 9.4%, respectively, while those for control stocks only increase by 1%, 7.3%, and 4.3%, respectively, for the 6-, 9-, and 12-month windows. The differences in increases of sales probabilities between treated and control stocks again are significant both economically and statistically across all three windows. Overall, we find that the tick-size reduction leads retail investors to reduce their holding periods of treated firms in reaction to reduced transaction costs, consistent with our main hypothesis.

C. Finland Transactions

There may be sample selection concerns as the U.S. sample covers only a subset of individual investors. To address this concern, we replicate our analyses using a transaction-level data set from Finland that covers complete trading records of all individual investors between 1995 and 2003.

Table 7 reports the summary statistics for the Finnish stock and investor characteristics. Summary statistics are calculated by pooling annual observations over the 1995–2003 time period. All liquidity measures are calculated as described in Section III.B. The results show that our main transaction cost measure—adjusted Amihud ratio (AdjIlliq)—is positively skewed, with a mean of 10.61 and a smaller median of 6.21. Other transaction cost measures show a similar pattern. For example, Zerofreq has a mean of 21.90% and a median of 20.64%. Finally, we estimate Closing price Spread (%) following Barber and Odean (Reference Barber and Odean2000). The mean Closing price Spread (%) is 0.083, while the median is close to 0.Footnote ¹⁷ Size is also positively skewed, with the average market capitalization approximately 10 times as large as the median one.

Table 7 Summary Statistics of Stock and Investor Characteristics in Finland

The mean (median) investor age is 39.5 (40). About 33% of the primary account holders are female. The mean (median) household stock portfolio value is 10,823 (2079) Euros in Finland. The mean (median) portfolio concentration is 0.20 (0.17), roughly corresponding to holding five stocks with equal value weights of 20%. Furthermore, 4% of households have traded options at least once, and less than 1% have ever held a short position during the 1995–2003 time period. This is not surprising since Bris, Goetzmann, and Zhu (Reference Bris, Goetzmann and Zhu2007) suggest that short selling became legal in Finland in 1998 but that tax laws inhibit would-be short sellers.

We use a similar framework to the one we utilize for the United States to test the validity of Hypotheses 1 and 2 for Finnish investors. We run the same hazard regression, modeling the conditional probability that a stock is sold, controlling for stocks’ transaction costs, firm characteristics, available investors’ demographic information, and trade-related characteristics. Consistent with standard reporting convention, we report estimated hazard ratios from the hazard regressions instead of estimated hazard coefficients in Table 8.

Table 8 Impact of Liquidity on Households’ Holding Periods in Finland, Hazard Analysis

The results estimated from the transaction-level Finnish data set in Panel A of Table 8 are remarkably similar to our findings for the United States. In the baseline model in column 1, the hazard ratio of the adjusted Amihud illiquidity measure (AdjIlliq) is 0.984, less than 1 and statistically significant. This indicates that if transaction costs (AdjIlliq) increase by 1 standard deviation (10.25), the investor is 15.2% less likely to sell that stock. We obtain comparable results using Zerofreq in column 2 and the Closing price Spread (%) in column 3. After we control for household- and firm-specific effects using stratification in column 4, the estimated hazard ratio of the adjusted Amihud illiquidity measure (AdjIlliq) is still less than 1 (0.976) and statistically significant. This coefficient indicates that with a 1-standard-deviation increase in AdjIlliq, the representative investor is 22% less likely to sell.

To explore how stock, investor, and trade characteristics affect holding periods, we include additional controls in the regressions reported in columns 5 and 6 of Table 8. Controlling for stock characteristics (i.e., Size, B/M, Momentum, and Unrealized Returns) in addition to household-specific effects in column 5 yields a statistically significant hazard ratio of less than 1 (0.979) for the adjusted Amihud illiquidity measure (AdjIlliq). We further control for both investor and stock characteristics available in the data set in column 6. The estimated hazard ratio for the adjusted Amihud illiquidity measure (AdjIlliq) is statistically significant at 0.988, suggesting that the average investor is 11.6% less likely to sell a stock when the stock’s transaction cost increases by 1 standard deviation.

The hazard ratios for investor characteristics are also quite similar to those for the U.S. sample. Specifically, the hazard ratio for age is less than 1, implying that older investors have lower turnover. In contrast, the hazard ratio for the male dummy is larger than 1, suggesting that male investors tend to have shorter holding periods and trade more frequently. The hazard ratios for all trade-related variables are larger than 1, suggesting that investors who trade options, who invest more capital in the stock market, and who concentrate their investments in fewer securities have shorter holding periods, consistent with our findings in the U.S. data.

The loadings on stock characteristics are also similar to those in the United States except for size. Similar to U.S. investors, Finnish investors are also more likely to sell past winners while holding value stocks for longer periods. Unlike in the United States, investors in Finland do not prefer to hold larger firms for longer periods. Altogether, the results in Panel A of Table 8 are similar to our U.S. findings reported in Table 2. Individual investors in Finland are also cognizant of and pay attention to transaction costs when they make trading decisions.

In Panel B of Table 8, we further investigate heterogeneity in the relationship between transaction costs and holding periods. In particular, we examine if financially more sophisticated investors pay more attention to transaction costs. As in the U.S. analysis, we assume that financial sophistication is correlated with education and resources available to each investor. We construct a similar sophistication measure. Sophistication score increases by 1 for each of the following three criteria met: if the investor is ranked among the top 25% based on the amount of capital invested in the stock market; if the investor has experience trading options (i.e., Option User Dummy = 1), or if the investor’s portfolio concentration is above that of the median investor.Footnote ¹⁸ Since the Finland transaction data do not provide information regarding investors’ income, their professions, or whether the investor has ever traded any foreign securities, we exclude these criteria in the construction of the Finnish sophistication measure. Our sophistication measure for Finland ranges from a minimum of 0 for the least sophisticated investors to a maximum of 3 for the most sophisticated investors.

We then divide Finnish investors into two subgroups based on their financial sophistication. Group 1 is comprised of the less sophisticated Finnish investors with Sophistication Score values of 0 or 1, while Group 2 includes the more sophisticated investors in Finland with Sophistication Score values of 2 or 3. Column 1 of Panel B in Table 8 reports that the hazard ratio of the Amihud illiquidity measure (AdjIlliq) for Group 1 is 0.992, and column 2 reports that the hazard ratio of AdjIlliq for Group 2 investors is 0.987. Both hazard ratios are statistically significant and less than 1. The hazard ratio for the more sophisticated investors is smaller in magnitude compared to the hazard ratio for the less sophisticated investors.

These results suggest that the Finnish investors who are more sophisticated hold stocks with higher transaction costs for a longer period of time than less sophisticated investors, consistent with more financially sophisticated investors paying more attention to transaction costs compared to their less sophisticated peers. In both columns we control for Size, B/M, Momentum, Unrealized Returns, Age, and Male Dummy. Using alternative transaction cost measures, including the proportion of zero-return days, (Zerofreq), as well as actual transaction costs (Closing Price Spread (%)), generates comparable results. Overall, these findings suggest that our findings in the United States are unlikely to be driven by the specific sample of investors and the time period we study.