key: cord-1046529-uu4xmwxg
authors: Aktas, Osman Ulas; Kryzanowski, Lawrence; Zhang, Jie
title: VOLATILITY SPILLOVER AROUND PRICE LIMITS IN AN EMERGING MARKET
date: 2020-05-30
journal: Finance research letters
DOI: 10.1016/j.frl.2020.101610
sha: 3965f6b13b3678c9aefc3dcca641ccf945720ff3
doc_id: 1046529
cord_uid: uu4xmwxg

Abstract The intraday volatility effects of price-limit hits for stocks in the BIST-50 index during a volatile period are examined. Our evidence supports the volatility no-effect, dampening and spillover hypotheses depending on whether the lower or upper price limit is hit and on when the hit begins and ends. Post-hit volatilities tend to be lower for limit hits near the beginning of the first trading session, unchanged for those that transcend a trading session and for upper price-limit hits near the end of either trading session, and higher for lower price-limit hits near the end of either trading session. These results are robust using samples differentiated by cross-listed status, same-day news, equi-distant and trade-by-trade returns and volatility measures accounting for return-series autocorrelations. Our findings have implications for emerging markets planning to impose price-limit bands or to increase their efficacy.

We gratefully acknowledge financial support from the Senior Concordia University Research Chair in Finance, IFM2 and Social Sciences and Humanities Research Council of Canada (SSHRC). The usual disclaimer applies. INTRODUCTION Many stock markets, including those in North America, use different mechanisms with different configurations to address stock price volatility. Little agreement exists on the merits of one such mechanism, price limits, for dealing with price volatility. Advocates (e.g., Anderson, 1984; Chou et al., 2000) argue that price limits moderate stock price volatility, and help to correct short-term overreaction by providing a cooling off period for investors to digest new information, and that price reversals, lower price volatilities and thinner trading volumes are more likely after price limit hits (henceforth PLHs).

Critics (e.g., Kim and Rhee, 1997; Lee and Kim, 1995) argue that price limits only delay price discovery by inefficiently stopping order flow and causing volatility spillover. The mixed empirical evidence includes some studies finding no stock volatility effects (e.g., Phylaktis et al., 1999; Diacogiannis et al., 2005) .

We investigate the effects of the price-limit configuration on stock market volatility for the Borsa Istanbul (BIST) Stock Exchange. This provides an ideal laboratory for such a test since the BIST is a fully computerized order-driven market using price and time priority for order matching. This market which is open to foreign investment is situated in a major developing economy with a Western-style economy with great importance to regional peace, stability and security among Balkan countries due to its unique geographic location, and its role in oil distribution from the Middle East to Europe and in NATO (second largest standing military force). The price limit rules in place during our sample period allowed stock prices to fluctuate to a maximum of 10% plus any rounding up to the nearest tick in each session from the "base price" calculated based on the previous trading session. A "limit lock" occurred when trading was either interrupted or continued at the limit prices given an excessive number of buy (sell) orders at a limit price with no offsetting orders (Bildik and Gulay, 2006) . The interruption continued until a seller (buyer) was willing to trade at a price within the set limits for that session. For limit locks remaining at session close, trading resumed in the following session after a resetting of the price limits.

To ensure a sufficient number of observations to compute measures of volatility, our sample is confined to limit hits surrounded by 30-minute windows with no transacted price at the limit hit. We find evidence to support the volatility no-effect, dampening and spillover hypotheses since the impact on volatility depends on whether the lower or upper limit is hit, on when during the day the PLH begins and ends, and whether the stock is cross-listed. Our base-line findings are robust using ARIMA models with either trade-by-trade or equi-distant returns.

Our paper contributes to the literature on price-limit effects by examining the performance of a specific price-limit configuration during a period of extreme global volatility that is likely to repeat given the recent COVID-19 pandemic. It also contributes to the extensive and growing literature on the effects of a global shock, the 2008-9 financial crises, on such diverse topics as: cost of corporate debt financing (e.g. Pianeselli and Zaghini, 2014), financial constraints or frictions on various firm decisions (e.g.

Campello, Graham and Harvey, 2010; Benmelech, Frydman and Papanikolaou, 2019) , social capital's association with firm performance (Lins, Servaes and Tamayo, 2017), stock market contagion (Bekaert et al., 2014) and media tone and readability for IPO firms (Bhardwaj and Imam, 2019) .

Our sample consists of BIST-50 index members, an index which represents 70% of the BIST's total stock trading volume during the studied period. We used all cleaned transactions and incoming orders, and our reconstructed full limit-order book for all stocks in our sample. Following Bildik and Gulay (2006), we use and for upper and lower price limits, respectively, where is the volume-weighted average price from the previous session s-1 and 0.10 is the permitted price divergence from .

As in Kim and Yang (2008) , limit hits are classified as single, consecutive and closing. Consecutive limit hit are as a series of limit hits without a 30-minute window available between two consecutive limit close. Illiquidity is not a concern since at least 91% and 94% of the pre-and post-PLH windows, respectively, have at least 30 seconds with trades.

The volatility spillover hypothesis argues that price volatility increases after a limit hit or lock (e.g., Kim and Rhee, 1997) if price limits or locks cause greater uncertainty due to informational asymmetry (Spiegel and Subrahmanyam, 2000) and delay price discovery by adversely affecting trading (Fama, 1989; Lehmann, 1989) . However, if this type of market intervention provides traders with time to obtain information to reduce informational asymmetry, reassess the market price, and avoid or correct overreaction, then it follows from the volatility dampening hypothesis that price volatility is expected to be lower after a limit hit or lock (Kim and Yang, 2008) . Another possible outcome is that these two opposing effects neutralize each other so that price volatility remains unchanged. The resulting testable null hypothesis is: No pre-to-post PLH change in volatility ( : volatility no-effect hypothesis).

Alternative hypotheses are lower volatility post-hit ( : volatility dampening hypothesis) and higher volatility post-hit ( volatility spillover hypothesis).

Among the three types of benchmarks used in the literature to test the volatility spillover hypothesis, we choose the one that compares volatilities in the 30 minutes before and after a limit hit or lock (see Section A3.1, Online Supplementary Appendix (OSA) for greater details). While most studies choose a discrete sampling scheme to control for microstructure noise, the severity of any bias introduced with this choice is unclear. Since the value of the metric being examined seldom coincides with the end of each equi-distant interval due to trade randomness, the calculation of evenly spaced high-frequency returns necessarily relies on some form of interpolation among prices recorded around the endpoints of the given sampling intervals. This results in a nonsynchronous trading or quotation effect that may induce negative autocorrelation and heteroscedasticity in the interpolated return series. These biases may be exacerbated in a multivariate context since varying degrees of interpolation are employed in the calculation of the returns for different securities. Since the impact on the examined metric (e.g., price) of intra-interval events with informational content is imperfectly captured by the choice of equi-distant intervals, the equidistant intervals may discard valuable information. Thus, the choice of measurement interval involves a tradeoff between statistical measurement error that decreases and untreated microstructure-induced bias (e.g., significant autocorrelation from bid-ask bounce (Roll, 1984) ) that may increase as the sampling frequency increases.

Our base-case results use ten three-minute (equi-distant) intervals 1 based on an examination of estimated ARIMA (1,0,0) models using trade-by-trade returns. Dickey-Fuller tests find evidence of a unit root in both the paired pre-and post-windows for only seven lower PLHs and only two upper PLHs. We also find that price reversals or bid-ask bounce is likely to be a problem using trade-by-trade returns for all samples. Thus, Section 4 provides results using unequally spaced intervals to assess the impact (if any) of this high level of negative autocorrelation in our trade-by-trade returns on our base-case inferences.

We use the following model-free but definition dependent metrics to measure volatility or variation in each 30-minute window:

where and are the not demeaned and demeaned 3-minute returns, respectively, for limit hit i in window j (pre-or post-limit hit) for a window of length (= 30 minutes herein). The measure given by (4) 

We report summary statistics for each of the four volatility estimates for lower and upper PLHs for the full sample and three differentiated subsamples in Tables 1 and 2 Table 2 ). Thus, the average effect of both a lower and upper PLH depends upon when during the day the PLH is first triggered and also on whether it remains at trading session close.

[Please place Tables 1 and 2 here] 2 We get no negative Vol estimates using the demeaned three-minute returns but do using the non-demeaned threeminute returns. 3 Not unexpectedly, the mean differences either become insignificant or change sign for the full sample for both lower and upper PLHs and become insignificant for the sample of upper PLHs triggered during the first session's first 30 minutes when trade-by-trade returns are used in the calculation of the two volatility measures that capture the effect of bid-ask bounce; namely, and .

For robustness, we examine PLH samples of cross-listed firms with(out) U.S. ADRs and for firms with/without no same-day firm-specific news. The OSA results (Section A3.3) provide support for all three specifications of our first hypotheses since the post-window impact of a PLH depends on its beginning and ending time-of-day. While PLH volatilities tend to be lower post-limit hit near first trading session beginnings, higher (unchanged) post-hit for lower-price (upper-price) limit hits near trading session ends, and unchanged post-hit for those still in place at session end. The changes post PLHs tend to differ depending upon cross-listed status but not on same-day news.

Our previous variance results based on 3-minute returns are now shown to be robust using ARIMA models with trade-by-trade returns in this section and three-minute interval returns in OSA section A4.

This buttresses our previous inference that our test results support our second null and our two second alternative hypotheses since the post-window impact on return variances of a PLH depends on time-ofday when the PLH begins (e.g. beginning or end of a trade session) and when it ends (e.g., same or subsequent session).

We use ARIMA models with trade-by-trade returns since several studies suggest ARIMA(0,0,1) or ARIMA(1,0,0) or models (known as MA(1) and AR(1) models, respectively) help to mitigate any autocorrelation biases when using trade-by-trade returns. 4 The MA(1) and AR(1) models estimated herein are:

MA(1):

AR(1):

where ⁄ is the first-order for the MA(1); is the correlation between successive return observations in the AR(1); and and are the error terms assumed to be mean zero and IID normal distributed. As widely documented, the theoretical variance is given by for the MA(1) model and by ⁄ for the AR(1) model. Pre-and post-windows for each PLH are examined only for the series found to be stationary (i.e., ) based on the Dickey-Fuller test.

We only tabulate summary results of interest for the MA(1) model for the lower PLHs (Table 3) and

for the upper PLHs (Table 4) [Please place Tables 3 and 4 here.]

This study analyzes the effect of price limits on stock return volatilities using intraday data from an We find supportive evidence for the volatility no-effect, volatility dampening and volatility spillover hypotheses for members of the BIST-50 index depending on when during the day the PLH 5 The mean (median) post-minus-pre PLH windows for both ARIMA models generally carry the same sign and lead to the same statistical inferences based on conventional levels of significance. Exceptions are positive values for lower price PLHs during both sessions' last 30 minutes that are significant and nearly significant for the MA(1) and AR(1) models, respectively. begins and/or ends, and whether it is a lower or upper PLH. Our findings are robust using trade-by-trade returns as well as ARIMA models. These results have implications for studies examining PLH effects across markets with different PLH configurations, and for emerging markets planning to impose pricelimit bands or to adjust to more efficient price limits rules. Strict price limit applications without considering time-of-day effects when configuring price limits may not lower stock price volatility without damaging the price discovery process.

All authors made an equal contribution.

Wei, S. X. and C. Zhang, 2005, Idiosyncratic risk does not matter: A re-examination of the relationship between average returns and average volatilities, Journal of Banking and Finance 29, 603-621. 

The industrial organization of futures markets: A survey, 1-33

The global crisis and equity market contagion

Power and bipower variation with stochastic volatility and jumps

Financial frictions and employment during the Great Depression

The tone and readability of the media during the financial crisis: Evidence from pre-IPO media coverage