Intraday characteristics of stock price crashes

Abstract

This article presents the first detailed analysis of the intraday characteristics of idiosyncratic stock price crashes. The analysis focuses on the impact of large crashes in single stocks on their intraday returns and liquidity in the US market. Furthermore, optimal intradaily behaviour during crashes is studied. Crashes are found to happen rather quickly, usually during a time interval of a few hours. In general, a strong increase in trading activity is observed during a crash, indicating that investors are able to sell their stocks even in distressed markets. The level of liquidity change is linked to the size of the crash. However, there is little evidence that the large sales volume during a crash drives down stock prices. After a stock price crash a significant momentum effect is found for several hours. Stock price crashes appear to reduce information asymmetries.

Notes

¹ A model by Morris and Shin (2003) illustrates how a stock market crash can trigger the so-called liquidity black holes characterized by a small capacity to absorb selling pressure from active traders.

² This hypothesis is analogue to the paper by Atiase (1985), stating that the response of the stock price to second-quarter earnings is inversely related to the market capitalization. This finding is attributed to the fact that the available information on a company increases with its capitalization.

³ Unfortunately, we do not have access to a longer time frame of data.

⁴ Longer spacings are also used, but do not return significantly different results in general.

⁵ The analysis is repeated using a sample consisting of 30 S&P 500 companies to test if there are size-associated effects. However, no size-associated effect is found and the results are essentially the same. Hence, the results of the additional analysis are not discussed separately in this article.

⁶ However, there has to be at least a time frame of 1 week between crashes of one company.

⁷ For penny stocks the minimum tick size of 0.01 USD can be a considerable proportion of the total price, causing erratic returns which are not a consequency of information processing, but of the tick size.

⁸ As an example, a frequent case in the dataset is a crash on a day with a total price decline of 15.5%. The main crash often causes a price drop of about 13%. The remaining 2.5% price decline occurs during an extended time period of many hours after the crash. Choosing the crash identification threshold to be 15% falsy identifies the post-crash period as the crash period. Thus, facing the empirical evidence, it is useful to set the crash identification period to the point where two-thirds of the minimum total crash size occurred.

⁹ This ensures that the CIP is during the time of the main price drop and not accidentally set to be during the post-crash momentum period.

¹⁰ The 9.30 a.m. time point in the constructed grid captures only the first trade executed on the market. Since there are large trades reported right after the close of the market, we assume that those trades are entered in the NYSE trading system before the close but reported later. Although NYSE stops accepting orders at 4.00 p.m., specialists might still manage orders entered just prior to the close. Thus, the NYSE closing trade is sometimes recorded several seconds after the closing bell. For further information on opening and closing trades refer to Bacidore and Lipson (2001). Those trades are summarized in the last time interval finishing at 4.05 p.m.

¹¹ x = 5 results in crash periods whose beginnings and ends are characterized by a period of 25 min during which the prices do not fall. To control for the effects of the chosen parameter x, the analysis is repeated using x = 20 (i.e. a BIP and an EIP of 100 min). In general, the results are comparable to the x = 5 case and not reported for brevity.

¹² The end of the crash is determined by average prices being equal to or higher than the last price measured during the crash. Thus, average returns in the x 5-min periods after a crash have to be nonnegative by the research design.

¹³ The cumulated return in period j is , where i indicates the number of the crash and r_i,j is a 5-min return. In this graph the different crash durations are not taken into account.

¹⁴ Results not reported in Table 1 for brevity. Before the crash, mean autocorrelations are −0.0537, −0.0177 and −0.0176 for lags 1, 2 and 3, respectively. The Ljung–Box test and the Breusch–Godfrey test (refer to Greene, 2003, for details) show significant autocorrelations at the 10% level for 47 and 50 of the 162 crashes, respectively. After the crash mean autocorrelations of −0.0576, −0.0253 and −0.0131 are found for lags 1, 2 and 3, respectively. The significance of the Ljung–Box test and the Breusch–Godfrey test increases, returning a significant autocorrelation for 63 and 59 of the 162 crashes, respectively. Serial correlation of returns appears more pronounced after the crash than before. However, the level of this autocorrelation is limited.

¹⁵ The observed contemporaneous movement of return volatility and turnover can also be found when analysing intraday return seasonality (compare, e.g. Gwilym et al. (1999) for an analysis of the LIFFE futures market).

¹⁶ Christie et al. (2002) find a median spreads of 0.625 USD after a mid-day trading halt and a median of 0.30 USD on a normal day.

¹⁷ The asymmetry between bid and ask distances is particularly strong for the sample of the S&P 500 stocks, which is analysed, but not discussed separately in this article. The sample is created by choosing the 30 stocks out of the S&P 500 with the largest daily price drop from August to October 2002. The bid price of the S&P 500 companies during the crash is at a mean discount of 4.86% (1 − 0.95136), whereas the ask price is only at a mean premium of 2.84% to the next transaction price. This asymmetry is supported by the fact that only the mean bid distance is significantly different from zero at the 1% level, in contrast to the ask distance, which is not even significant at the 10% level.

¹⁸ Results on the bid and ask sizes as measured in number of shares are not reported in tables for brevity.

¹⁹ Because of a crash duration below three periods or data errors, we eliminate nine of the 162 crashes from the dataset for this analysis.

²⁰ The following analysis is also performed using levels of the liquidity variables during the crash and a strong link to the crash size is found. However, the explanatory content is lower than for the changes of the variables as discussed in this section. Although not in the focus of this article, our results can deliver insights to improve risk management systems which model the interaction of liquidity and stock price crashes. Compare, among others, work on liquidity adjusted value-at-risk by Jarrow and Subramanian (1997), Bangia et al. (1999), or Angelidis and Benos (2006).

²¹ The analysis is repeated with a longer time frame for the pre-crash means, using data from the beginning of the dataset until the last 5-min period before the crash. The results are robust and are not reported here for brevity.

²² The following analysis is also performed using levels of the liquidity variables during the crash as explanatory variables for crash duration. However, a very weak link between the levels of liquidity variables and the crash duration is found.

²³ The analysis is repeated with a pre-crash period lasting from the beginning of our dataset to the last 5-min interval before the crash. The results stay essentially unchanged and are not reported here.

²⁴ Other settings for x give comparable results.

²⁵ Unlike before, liquidity is split in buy side and sell side liquidity, making it necessary to use longer time intervals for stable liquidity statistics. Furthermore, a considerable number of transactions is not classified at all because of a lack of price changes. Robustness tests not reported here show that the results do not depend on the length of the interval.

²⁶ The parameter x is set to values ranging from three to 12 periods, which is very conservative, but results do not change substantially.

²⁷ For comparability reasons we also analyse the order flow characteristics of 90 nondistressed companies. This sample of 90 companies contains the 30 Dow Jones companies, 30 mid-sized companies from the S&P 500, and the 30 smallest companies in the S&P 500. While the OIN measure indicates a balanced order flow for these stocks in our event period, the OIT measure indicates a sales surplus (average OIT between 1.541 for large and 1.764 for small companies). However, with the equity market indices being stable in this period, the sales surplus does not drive down the markets. When running tests for difference for the size-based company samples and our crash sample the following t-statistics are obtained for the pre-crash period: t-statistics of −4.496, −3.906 and −2.986 for the OIT ratio as well as −8.546, −7.861 and −6.241 for the OIN ratio for large, medium and small companies, respectively. Thus, order flow imbalance is higher for stocks that are about to experience a crash.

²⁸ For brevity, results are not reported here.

²⁹ The parameter of 0.0125 for OIN is significant at the 5% level (t-statistic of 2.0854). Constant selling pressure appears to have little impact on the intraday price drift. The intercept is −0.0422 and is significant at the 1% level.

³⁰ However, the application of a cross-sectional regression analysis assumes that the relationship between the order flow imbalance measures and crash size are more or less equal across the companies in the sample. This might perhaps not be the case in reality.

³¹ We use a Newey–West procedure to obtain a robust covariance matrix, which is also used as weighting matrix in the estimation. A two-step GMM procedure is applied, using the identity matrix as weighting matrix in a first step. The inverse of the resulting estimated covariance matrix is used as weighting matrix in the second GMM estimation. Our orthogonality condition is that all independent variables are orthogonal to the error term.

³² Compare De Winne and Platten (2003) for more details.

³³ Since the spreads of distressed companies are also higher than the ones of nondistressed companies, the order processing cost components in the spreads of distressed companies in USD are higher as well.

³⁴ As expected, the bunching procedure results in an increased trade reversal probability of 0.427 and 0.425 before and after the crash, respectively. These numbers are lower than the ones obtained by Huang and Stoll (1997) (their trade reversal probability is above 0.676). However, our numbers appear more realistic since they indicate a slight tendency towards trade continuation.

³⁵ The results in this paragraph are given using the parameters estimated for the bunched trades sample.