Acknowledgments: We thank the Editors and referees of this journal for some very comprehensive and helpful comments on an earlier draft of this paper.
Corresponding author: Ronny Manos, The College of Management, School of Business Administration, 7 Yizhak Rabin Boulevard, Rishon LeZion 75190, Israel. Tel: +972 36438753, Fax: +972 36438753, email: email@example.com.
We investigate the impact of the March 1995 move to screen-based trading on the Mumbai Stock Exchange, using separate samples of more liquid (A) and less liquid (B) shares. Following the move, the average cumulative abnormal return for A shares was 4.5%, whereas that for B shares was over 12%; market liquidity and efficiency increased but the effect on volatility was more ambiguous. We identify a significant cross-sectional relationship between the size of cumulative abnormal returns and firm-specific improvements in liquidity, efficiency, and volatility, with differences in the effects of reform on A and B shares.
During the past two decades, a range of market microstructure reforms has been implemented in stock markets around the world. The reforms have been aimed at improving market efficiency, including modernization of trading and information systems, expanding membership and revamping the regulatory framework. Stoll (2006) emphasizes the benefits to be expected from the introduction of modern trading technology. There have been numerous studies on the effectiveness of microstructure reforms in the more established stock markets, focusing in particular on trading and information systems; and most of the studies suggest that the reforms have had a positive impact, creating efficiency gains in the price discovery process, increased liquidity and lower volatility. Examples include studies of Milan (Amihud et al., 1990; Majnoni & Massa, 2001), Tokyo (Amihud & Mendelson, 1991), Tel Aviv (Amihud et al., 1997; Kalay et al., 2002), London (Chelley-Steeley, 2008), Dublin (Chelley-Steeley & Lucey, 2008), Paris (Pagano & Schwartz, 2003), and New Zealand (Blennerhassett & Bowman, 1998).
Many emerging stock markets have participated in the reform process, but there have been relatively few studies of emerging markets. This is an important omission for several reasons. First, in comparison with industrial countries, emerging stock markets invariably start from a lower base, with less frequent trading, lower levels of efficiency evidenced by higher serial correlation in prices, slower adjustment of prices to news, and indirect evidence of more insider trading (Bekaert & Harvey, 2002). Therefore, reforms that are effective in developed markets might be less useful in emerging markets and vice versa. Indeed, the relatively few studies of microstructure reforms in emerging markets show mixed results. Chang et al. (1999) find no change in liquidity or in price-efficiency but an increase in volatility, following the move to a continuous auction system in Taipei. Kim & Singal (2000) find that opening the stock exchange to foreign investors was substantially more effective in increasing efficiency in 20 emerging stock markets than was reforming the trading system itself.
Second, there is mounting evidence that trading systems (and, hence, reforms) that are appropriate for more liquid securities might be inappropriate for less liquid securities, and vice versa (Muscarella & Piwowar, 2001). Even though most exchanges have replaced floor trading with electronic trading,1 evidence from developed stock markets suggests that automated trading reduces execution costs for liquid stocks, but floor trading might be better at reducing information asymmetries and trading costs for less liquid stocks (Benveniste et al., 1992; Madhavan & Sofianos, 1998; Venkataraman, 2001). Chung & Kim (2009) find that the specialist system on the New York Stock Exchange provides better liquidity services than the NASDAQ dealer market, particularly in times of high return volatility. On a parallel theme, it is often argued that for trading less liquid stocks, a periodic auction is a more efficient and lower cost method than a dealer market (Barry & Brown, 1984; Madhavan, 1992). However, in a study on the London Stock Exchange, where trade at the opening and closing may take place either through a call auction or the dealer system, Ellul et al. (2005) find that less liquid stocks traded more frequently in the dealer market, even though the auction offered cost savings and greater pricing efficiency. As dealers guarantee the availability of a counterparty, whereas call auctions depend on coexisting public orders, Ellul et al. (2005) argue that there is a “thick-market” externality: traders will only enter an auction if they believe sufficient counterparty traders will be present so as to establish an informationally efficient price. Camilleri & Green (2009) find that opening and closing call auctions on the National Stock Exchange of India (NSE) have markedly different effects on shares of differing liquidity.
Therefore, as lower liquidity is a central feature of most emerging markets, we would argue that, for a full understanding of microstructure reforms in emerging markets, it is essential to identify any differences in the effects of the reforms on high and low liquidity shares. Furthermore, as variations in liquidity affect share values (Lee, 2009), we would expect the liquidity effect of reforms to be impounded in share prices in a measurable way.
In the present paper, we study the Mumbai Stock Exchange (BSE),2 and investigate the impact on share values and market characteristics of the March 1995 switch from floor- to screen-based trading. The new system was dubbed “Bombay On-Line Trading” (BOLT). We extend the work of Shah & Thomas (1996) and follow Amihud et al. (1997) and others in using an event study as our basic investigative tool. We control for thin trading in our estimation procedures; we test the duration of the event’s impact; and, most importantly, we compare the impact of the event on two distinct samples of shares: more liquid (A) and less liquid (B). Our analysis enables us to pin down more precisely the effects of BOLT on market performance, and how much these effects differ between more and less liquid shares.
The BSE has a particular advantage for the present study in that it has certain features that make it an interesting laboratory for studying reform in emerging markets. First, several major reforms took place in the 1990s. Here, we concentrate on the switch from open outcry to electronic trading, which took place in 1995. Second, data availability in India is generally good; the BSE has a large number of quoted companies,3 with data available for the pre-reform era. Third, as the market is of long standing origin, we can obtain a relatively large balanced panel of companies to evaluate the impact of reforms in a consistent way. Fourth, the market consists of a wide range of shares. Some are relatively heavily traded, but many others are very thinly traded. This enables us to investigate how reform has impacted on securities of varying liquidity, and to identify channels through which reform might have been most effective. Of course, debating the desirability per se of electronic trading is somewhat moot because virtually all exchanges utilize automated trading, and we would expect automation to add value overall on the BSE, notwithstanding that floor trading does retain certain useful properties. The key issue in the present paper is the balance of gains and losses between the more and less liquid groups of shares. Differences between these groups might suggest that effective trading reforms in emerging markets need to be sufficiently flexible to take into consideration the different characteristics and trading needs of more and less liquid securities.
Our results suggest that the introduction of BOLT was followed by a significant improvement in market performance, both in terms of share valuations and in direct measures of liquidity, efficiency, and volatility (LEV). However, we also find important differences between the more and less liquid shares in their response to the reforms, broadly confirming our conjecture that stock liquidity is an important consideration in the reform process. Specifically, we provide new evidence to show that the effects of reform in India were conditional on share liquidity: less liquid B shares appeared to benefit significantly more from the reforms, both in increased liquidity and in increased value, than did the more liquid A shares.
The rest of the paper is structured as follows. In Section 2 we summarize key changes that took place in the BSE during the reform era. In Section 3, we describe the data set used in the analysis. The price reactions to the introduction of BOLT are estimated in Section 4. We concentrate on implementation and do not control for any effects of the pre-announcement of BOLT.4 Possible reasons for the price reactions are investigated in Section 5, where we study changes in LEV. In Section 6, we analyze the relative contributions of improved LEV to the overall price improvements observed in the market. Section 7 contains some concluding remarks.
2. Mumbai Stock Exchange and the Reform Process
Established in 1875, the BSE was India’s only NSE until 1994, when the NSE commenced trading.5 Until the 1990s, India’s securities markets were governed by legislation that was highly restrictive of share trading, new share issues and foreign participation in securities markets. Companies had little incentive to finance their activities through the private capital market (Agarwal, 1996; Singh, 1998). During the same epoch, the brokerage industry operated largely as a cartel and trading on the BSE was conducted by open outcry (Shah & Thomas, 1997). The market was characterized by high transactions costs, lack of liquidity, a fragile clearing system, and slow and unreliable physical settlement. A unique feature of the BSE was its system of leveraged trading called “carry forward” or “badla,” which permitted the postponement of cash settlement to the next account period (Patel, 1995). Two problems with badla were that settlement could in theory be postponed indefinitely, and that there were no formal margin requirements. It therefore fostered complex highly leveraged positions and exacerbated settlement risks. Following a share-dealing fraud and subsequent investigation, badla was suspended in 1993 (Shah, 1997), reintroduced in 1996 on the recommendation of the Patel Committee (Patel, 1995),6 and abolished entirely in 2001 when T + 5 rolling settlement was introduced. According to Berkman & Eleswarapu (1998), the abolition of badla decreased market liquidity and increased noise, and the market reacted positively to its later reintroduction, suggesting that, overall, badla had been valued positively by the market, notwithstanding its regulatory problems.
Capital market reform began with the repeal of old legislation and the establishment in 1992 of the Securities and Exchange Board of India, which introduced, inter alia capital adequacy rules for brokers, a share depository system involving (from 1997) the progressive dematerialization of securities, and an investor guarantee fund (Agarwal, 1996). Capital markets were opened up to foreign institutional investors, and the Reserve Bank of India liberalized interest rates and credit controls. In 1994, the NSE opened a new nationwide screen-based trading system, which quickly attracted a large part of India’s bond trading, but the BSE retained much of the share-trading business, in part because of its own reforms.7
The centerpiece of the BSE reforms was the introduction of screen-based trading: BOLT.8 Initially, this was a hybrid order-and-quote-driven system that accepted and disseminated two-way price quotations from authorized dealers, and market and limit orders from authorized brokers. The order book functioned as an auxiliary dealer, matching orders only where there were no quotes for a particular security, and improving the price competitive character of the market in cases where individual investors were willing to deal at prices better than the best dealer’s quote.9 BOLT was introduced on 14 March 1995 when 818 shares were transferred to electronic trading (all “A” shares and the most liquid “B” shares).10 The remaining shares (over 5000) were transferred in stages, and all shares were on the new system by the 50th trading day. As BOLT was introduced shortly after the NSE was established, there is a risk that the analysis of either event may be confounded by spillover effects from the other. We tackle this problem in two ways: first by choice of estimation period in the calculation of the abnormal returns, and second by a cross-sectional analysis in which we consider whether the effects of BOLT differed between firms that were quoted on both exchanges and those that were only quoted on the BSE.
When BOLT was introduced, securities traded on the BSE were classified into specified (or “A”) and non-specified (or “B”) shares. “A” shares are the more liquid and, to attain the “A” classification a company had to be sufficiently large with a widespread shareholding, steady dividend, good growth record and, more particularly, a large volume of secondary market trading. The remaining less liquid securities were classified as B shares.11
Our data consist of the daily closing prices of firms that were quoted continuously on the BSE from 1 September 1993 to 29 November 1996.12 The sample amounts to 69 A shares and 83 B shares, and is smaller than the total number of quoted firms at any point in the sample period because it excludes financial firms, quoted subsidiaries of foreign companies, and firms whose data were not continuously available over the full period. We restricted our sample in this way for the following reasons. First, we wanted to focus on privately owned Indian firms. Public sector or foreign-owned firms might not respond in the same way to market reforms as would private Indian firms.13 Second, we ruled out financial companies. Many were wholly or partly owned by the Indian Government or state governments and, therefore, belonged to the public sector. Moreover, insofar as market reforms alter the financial structure of non-financial companies, they will have indirect effects on financial firms because of their impact on their normal lending and borrowing business. These indirect effects could easily obscure the direct trading effect of market reforms. Third, trading and data availability for many firms prior to the introduction of BOLT were very fragmented. When a share is traded very infrequently there are too few observations with which to construct measures of the impact of BOLT, and the power of any tests is correspondingly reduced. Fourth, we excluded all firms incorporated after 1990 so as to avoid a “new firm” effect following the introduction of market reforms beginning in 1989. Fifth, there was, in general, a substantial turnover in quoted firms during the sample period. Overall, the true number of firms that were continuously quoted, priced, and traded during the 3 years of the study was substantially less than the number observed at any point in time.
The samples of A and B shares (Table 1) broadly reflect the population differences between the two groups. The average firm in the A sample is approximately three times larger and 9 years older than that in the B sample. Manufacturing firms and business groups are dominant, a feature that also reflects the population of non-financial companies in India.
Table 1. Data description The percentages are with respect to the total number of firms in the respective sample. Crore = 107 (i.e. 10 million). Membership in leading indices shows the number and proportion of the sample shares included in the index. For example, 67% or 97% of our sample A shares are in the BSE500, and 83% or 100% of the B shares are in the BSE500. Ownership group is based on the PROWESS classification. See Centre for Monitoring the Indian Economy (1997).
Number of firms
Membership in leading indices
Year of incorporation
Sales 1995 (crore Indian rupees)
Indian business groups
4. Introduction of Bombay On-line Trading
4.1. Cumulative Abnormal Returns Methodology
We begin by estimating the impact of the introduction of BOLT using an event study (MacKinlay, 1997). French (1980) distinguishes between a trading time model in which returns are generated only in trading time, and a calendar time model in which returns are generated continuously, implying that returns distributions will differ among Mondays, post-holidays, and other working days. Available evidence does not strongly favor either model (Lakonishok & Smidt, 1988). In the present paper, we adopt a calendar time approach and use the market model as the equilibrium model of security returns:
We define Rk,t as the return on stock k between two consecutive days when the exchange was open, from day t – n to day t, which include nt calendar-day returns; and Mt as the corresponding return on the market index. Therefore, and are the returns per calendar day during period nt, on stock k and the market index respectively; εk,t is the usual error term.
The standard event study method involves estimating Equation (1) outside the event window, and calculating the abnormal return for each trading day in the event window (ARk,t) as:
The cumulative abnormal returns (CARk,t) are:
where t = (T − u,…, T,…, T + v) is the event window; and t = T is the event day. These can then be averaged across shares to obtain the mean CAR for each day in the window.
An equivalent procedure for calculating the AR involves estimating:
across the non-window period and the event window itself using dh,t as daily dummies within the event window (dh,t = 1 for h = t and zero otherwise). Estimates of and from Equation (4) are identical to those from Equation (1), and the estimated are identical to the AR obtained using Equation (2) (Karafiath, 1988). This “event dummy” approach has the useful property that it permits direct computation of F-tests on the estimated dummies that can be used to check the length of the event window. This is important in studying the impact of new technology where market participants may incur costs of learning about the new system, whose impact might therefore be reflected in share prices only gradually over time. Plots of the CAR give an informal indication of the true length of the event window, but F-statistics provide a more rigorous test. Amihud et al. (1997) and Berkman & Eleswarapu (1998) use a dummy variable approach to calculate the AR but do not test the length of the event window. Moreover, they constrain the event dummy parameters to be equal across large segments of the event window and so impose a specific time pattern on the AR.14
The standard methodology does not impose any pattern on the AR, and their procedures appear unnecessarily restrictive, especially as there is no reason to believe that their constraints correspond to the true AR process.
Efficient and unbiased estimation of Equation (1) or (4) in daily data might be affected by thin trading, which was endemic in Mumbai when BOLT was introduced. Even the most liquid A shares in our data did not trade on average on 23% of possible trading days when the BSE was open (Table 2). For the B shares the comparable figure is 34%. Thin trading will tend to bias the OLS estimates of βk in Equation (1) and Dimson (1979) proposes instead estimating:
Table 2. Trading on the Mumbai Stock Exchange: September 1993 to August 1996 The years run from 1 September to 31 August. Closed days include weekends, holidays, strikes, payment crises, and technical breakdowns. aFor six A firms no prices are available for part of 1993–1994. For one of these firms no prices are available for part of 1994–1995. These firms are excluded from the statistics relating to 1993–1994, 1994–1995 and to the whole period 1993–1994 to 1995–1996. Hence, in Panel 2 the number of observations for 1993–1994 and for the full period is 63, whereas that for 1994–1995 is 68.
1993–1994 to 1995–1996
Panel 1: Number of weekdays the BSE was open for trading
Open days (fraction of total weekdays)
Panel 2: Descriptive statistics for number of trading days of 69 A stocksa
Average number of trading days (fraction of open days)
Panel 3: Descriptive statistics for number of trading days of 83 B stocks
Average number of trading days (fraction of open days)
If non-trading is not too severe, OLS estimation of Equation (5) will deliver unbiased estimates of βk (βk = ∑jβk,j) using relatively small values of h.15Dimson (1979) also shows that there is little difference between this and other methods of dealing with thin trading, in terms of the unbiasedness of the estimated βk.
Non-normality of the returns is also a possible problem. Mills et al. (1996) advocate robust estimation of the market model to deal with the endemic excess kurtosis of stock returns. However, Cable & Holland (2000) point out that the main application of the model in event studies is in the calculation of AR for a large number of stocks and subsequent cross-sectional averaging of these AR to evaluate the impact of the event. They show that averaging the AR across a portfolio typically restores normality for both robust and OLS estimators of the underlying market model. As this is precisely our application of the model, we proceeded with OLS estimation. In summary, we use the market model, estimated by the Dimson method applied to the event dummy approach. Therefore, we estimate:
To reduce the risk of ex post selection bias (Brown et al., 1995), we follow Amihud et al. (1997), who argue that the equilibrium model for evaluating market microstructure reforms should be estimated on data from the post-event period. A further reason for post-event estimation is that the NSE started trading on 4 November 1994, just before the introduction of BOLT. Shah & Thomas (1996) conclude that competition from the NSE had a negative impact on liquidity on the BSE. Using pre-event data to estimate the market model could, therefore, bias the CAR upwards. For the estimation period we experimented using 12 months (March 1995 to February 1996) and 18 months (March 1995 to August 1996). However, as the two periods gave very similar results, we only report the results for the 18-month estimation period.16
We used an event window of t = T −5,…, T + 50, where T is the starting date of BOLT. We used a 5-day window prior to BOLT as we were interested in the effect of BOLT per se and not in the effect of the announcements that preceded its introduction. The 50-day post-BOLT window concludes on the date that the transfer of stocks to BOLT was completed (Table 3). This is a relatively long event window, but a shorter window could lead to the omission of any cumulative learning effects that took place during the transfer process. As all the most liquid A shares and many of the more liquid B shares were transferred to BOLT on its opening day,17 it could be argued that most of the effects of the transfer would occur in the early part of the window. However, a long event window should shed light on the permanency of the effect of automation on stock prices (Amihud et al., 1997). Moreover, the transfer process itself could create a shadow over trading in shares that were transferred earlier to the electronic system. Dealers would need time and resources to adjust to the new system and to trade with the old and new systems simultaneously. These factors could affect the speed with which any gains from the new system were realized, especially given the endemic thin trading in Mumbai at that time. However, we did use F-tests on the estimated event dummies (γk,h) to check the length of the event window.
Table 3. Time periods used in the study Event day denotes the day relative to the day of Bombay on-line trading (BOLT) introduction (T = 14 March 1995). Open days denote the number of weekdays that the BSE was open for trading. Closed days denote the number of weekdays that the BSE was closed due to holidays, strikes, payment crises, technical breakdowns, etc. Weekdays denote the number of open days + the number of closed days. Total days denote the number of weekdays + the number of weekend days. 30 August 1996 and 29 November 1996 both fall on a Friday; hence, they are the last trading days of the respective months.
Number of days in period
Panel 1: Event period and the estimation periods used for the event study
12 months period
T − 5
7 March 1995
T + 223
29 February 1996
18 months period
T − 5
7 March 1995
T + 344
30 August 1996
T − 5
7 March 1995
T + 50
6 June 1995
Panel 2: Pre- and post-BOLT periods used for the comparison analysis
T – 345
1 September 1993
T – 6
6 March 1995
T + 51
7 June 1995
T + 401
29 November 1996
For choice of market index to estimate the market model we faced a trade-off between using an index of a few highly traded shares and one that is more broadly based but that includes more thinly traded shares (Table 1). The BSE Sensitive Index (SENSEX) consists of 30 A shares of large, well-established companies, accounting for approximately 20% of market capitalization and 40% of turnover on the BSE. The BSE100, BSE200, and BSE500 are more broadly based indices of A and B shares. We report results using the BSE200 because this index provides the best compromise between coverage and liquidity. As a check, we computed our results using the SENSEX index (available upon request) and the results for both indices were essentially the same.
4.2. Cumulative Abnormal Returns: Results and Implications
The market model regressions are shown in Table 4. It is reassuring that the mean of the αs is close to zero for both A and B shares. Likewise, the mean of the βs is not significantly different from unity for both A and B shares as we would expect in a large portfolio. We also calculated the cross-sectional means of the t-values, but a more precise check on the significance of the coefficients can be made by testing the null that the t-values exceed their (5%) critical value. This is done: first, using a t-test on the absolute values of the t-values themselves and, second, by counting the t-values that are at or above the critical value and computing the associated probability based on the binomial distribution. These tests confirm that the αs are not significantly different from zero, but that the βs are significant, as we would expect. The F-statistics are also significant and, as these include the 56 event dummies, this provides important evidence that there were significant AR in the event window associated with the introduction of BOLT.
Table 4. Regression diagnostics for the market model, event dummy method Summary statistics and tests from estimation of the market model using the Dimson method to control for thin trading with five leads, five lags and 56 event dummies are shown. The market index used is the BSE200. The means, medians and SDs are the cross-sectional statistics for the respective parameters given in the first column. In addition: t-test on the mean statistics: |μ| = Z5%. To construct cross-sectional t-tests, we compare the firm-specific estimates of tt and F from the market model with their respective 5% critical values. μ = mean tt or F-value from the market model; Z5% is the relevant 5% critical value of tt or F for the corresponding regression. The null and alternative hypotheses are: , . Interpretation of the alternative depends on the tail at which the t-statistic falls: • if t > 0, implies that the t-test/F-test for insignificance can be rejected at above 5% level; • if t < 0, implies that the t-test/F-test for insignificance cannot be rejected at 5% level; • the critical values (Z5%) refer to a two-tailed t-distribution in the case of α and β, or the F-distribution. The critical values against which the means are compared are as follows:
Degrees of freedom
350 − 68 = 282
d.f.1 = 282; d.f.2 = 67
Upper tail: 0.05
t-test on the mean statistics: |μ| = Z5%
Binomial: (<z5%: ≥ z5%)
Binomial: (<z5%: ≥z5%). For the binomial test, (x:y) gives the number of observations for which the absolute value of the t- or F-statistic is less than (x) or greater than or equal to (y), the relevant critical value given in note 1. The null hypothesis is that and are equally likely. The test returns the probability mass function, which is the probability of getting x:y under the null hypothesis. *** show the significance levels associated with rejecting the null, and represent the significance level at 1%.
Panel 1: Sample of 69 A shares (350 observations)
Panel 2: Sample of 83 B shares (350 observations)
The cross-sectional average CAR for A and B shares are shown in Figures 1 and 2. Both reacted positively to the new trading system, but the average impact on the B shares was approximately twice as large as on the A shares. On the first day after BOLT was introduced, A shares jumped by 1.59% on average; and the average CAR in the whole event window was 4.56%. The price increase appears to have been permanent as the average CAR in the last 10 days of the event window was 4.48%. The average B share jumped 2.09% on the first day after the event, and the average CAR for the whole event window was 12.85%, or more than twice that on the A shares. These are dramatic figures for A shares and, more particularly, for B shares. Of course, B shares make up a substantially lower proportion of the market than A shares. However, if we take the lower figure of 4.5% for all shares and apply this to the total market value of quoted securities on the BSE as of March 1995 (Rs 4355bn), this implies an increase in value of some Rs 196bn, or approximately 2% of India’s 1994–1995 current GDP at factor cost (Rs 9434bn). This is a rough estimate but on any reckoning it is a substantial impact.18
To evaluate the CAR further, we extend the approach of Amihud et al. (1997) and study the event window in segments. We test the cross-sectional significance of the CAR at days 0 and 1 and at 10-day intervals from days 0 to 50 by computing t-tests for the null that the mean of each CAR is zero (Table 5). We also compare the incidence of positive and negative CAR using the binomial. For the A shares, all the CAR apart from day 0 are significant and the null of equally probable positive and negative CAR is decisively rejected. For the B shares, all the CAR, t-statistics, and fractions of positive CAR are uniformly higher than they are for A shares. These results provide solid evidence that the introduction of BOLT had a significant positive and permanent effect on share values, and that the effect was particularly marked for the less liquid B shares.
Table 5. Tests of the CAR: Event dummy estimation results Summary statistics (mean, median, and SD) for the estimated cumulative abnormal return (CAR) for various event days, together with t-tests and binomial tests on these statistics are shown. The market model was applied to estimate the CAR, using the Dimson method to control for thin trading with five leads, five lags, and 56 event dummies. The market index is the BSE200. t-test: μ = 0. t-statistic associated with the mean CARt. t-statistic = E(CARt)N1/2/SD(CARt), where N is the number of observations. The null and alternative hypotheses are: H0: μ = 0, H: μ ≠ 0, where: μ = mean CARt. Binomial: (+:−). For the binomial test, (x:y) gives the number of observations for which CARt is positive (x) or negative (y). The null hypothesis is that positive and negative observations are equally likely. The test returns the probability mass function, which is the probability of getting x:y under the null hypothesis. The asterisks show the significance levels associated with rejecting the null. Significance at ***1%, **5%, and *10% levels.
Panel 1: 69 A stocks
Panel 2: 83 B stocks
t-test: μ = 0
t-test: μ = 0
Next, we use F-tests on the estimated γk,h for each security to help check the length of the event window (Table 6). We tested the joint significance of all 56 event dummies and then the 49, 40, 30, and 20 dummies most distant from the event day. For the A sample (Panel 1) the mean F-statistic is lower than the critical F-value in all cases, but the difference is not large enough to reject the null that they are equal, except for the 20 event dummies furthest from the end (days 31–50). We also tested the joint significance of the dummies in non-overlapping groups of 10 beginning at day 50. Apart from the most distant group (days 41–50), the mean F-statistics are all significantly less than the relevant 5% critical values, implying that the effect in each 10-day period was relatively small. Put together, these results reinforce the conclusion that the overall impact of BOLT on A shares was positive and significant, but it built up slowly, with many of the daily AR being insignificant. For sample B (Table 6, Panel 2), the mean F-statistic is above the critical value, implying rejection of the null that the dummies are insignificant in all but four cases. In three of these cases, the mean F-statistic is significantly lower than the critical value, implying rejection of the null of significance. However, these cases consist of the two blocks of dummies furthest from the event day, and the joint test of these dummies. These results suggest that BOLT had a significant positive impact on B shares, which built up more quickly than for A shares, with the most significant impact occurring during the first 30 days of BOLT.
Table 6. Descriptive statistics and tests of the F-statistics for the event dummies The market model was estimated using the Dimson method to control for thin trading with five leads, five lags, and 56 event dummies. The market index used is the BSE200. Table 6 shows the results of F-tests on the groups of the event dummy coefficients (γh) as indicated in the first column. Fr,s gives the degrees of freedom (r,s) for the relevant F-statistic; F5% is the corresponding critical value of the F-statistic at the 5% level. Mean, median, SD are the cross-sectional summary statistics for the F-values. t-test: μ = F5%. To construct the cross-sectional t-tests we compare the firm-specific estimates of F with their respective 5% critical values. The null and alternative hypotheses are: , . Interpretation of the alternative hypothesis depends on the tail at which the t-statistic falls: • if t > 0, implies that the F-test that the dummies are jointly insignificant can be rejected at above the 5% level; • if t < 0, implies that the F-test that the dummies are jointly insignificant cannot be rejected at the 5% level; where μ is the F-value and F(5%) is the relevant 5% critical value of F for the corresponding regression. Binomial: (<F5%: ≥ F5%). For the binomial test, (x:y) gives the number of observations for which the F-statistic is less than (x) or greater than or equal to (y) the relevant critical value given in column 3. The null hypothesis is that and are equally likely. The test returns the probability mass function, which is the probability of getting x:y under the null hypothesis. The asterisks show the significance levels associated with rejecting the null. Significance at ***1%, **5%, and *10% levels.
Panel 1: 69 A stocks
Panel 2: 83 B stocks
t-test: μ = F5%
t-test: μ = F5%
Binomial: <F5%: ≥ F5%
H0: γ−5 = γ−4 = ⋯ = γ50 = 0
−5 to +50 (56)
H0: γ2 = γ3 = ⋯ = γ50 = 0
+2 to +50 (49)
H0: γ11 = γ12 = ⋯ γ50 = 0
+11 to +50 (40)
H0: γ21 = γ22 = ⋯ γ50 = 0
+21 to +50 (30)
H0: γ31 = γ32 = ⋯ γ50 = 0
+31 to +50 (20)
H0: γ41 = γ42 = ⋯ γ50 = 0
+41 to +50 (10)
H0: γ31 = γ32 = ⋯ γ40 = 0
+31 to +40 (10)
H0: γ21 = γ22 = ⋯ γ30 = 0
+21 to +30 (10)
H0: γ11 = γ12 = ⋯ γ20 = 0
+11 to +20 (10)
H0: γ1 = γ2 = ⋯ γ10 = 0
+1 to +10 (10)
5. Liquidity, Efficiency and Volatility on the Mumbai Stock Exchange
According to Glen (1994), a positive price reaction would be expected if BOLT created improvements in fundamental market attributes, such as LEV. The differential reaction of A and B shares could be explained by differential changes in any of these attributes (Allen et al., 2001). Therefore, we next examine the changes in LEV by comparing market performance in the 18 months prior to BOLT with that in the first 18 months of electronic trading, excluding the event window itself (see Table 3). We choose this 18-month time period first to match up the calculations with those of the event study and second because the extent of thin trading prior to BOLT was such that the power of the statistics would be considerably reduced if we were to significantly shorten the comparison period.
Electronic trading contributes to market liquidity through its impact on fairness, speed of execution, access, and costs (Harris, 2003; Jain, 2005). BOLT should improve fairness and speed of execution of trades because the matching and trading rules built into the system are implemented objectively and are transparent. The previous manual floor trading system was subject to variations in the interpretation of rules and to abuses such as front-running (Harris, 2003). Electronic trading also removes the physical constraints of floor trading and provides equal access to all traders (Allen et al., 2001). Pre-BOLT access to the BSE by investors outside Mumbai was severely compromised by the need for additional intermediaries, poor communications, and higher transactions costs (Shah & Thomas, 1997).19
We use two measures of liquidity: trading frequency and estimated bid-ask spreads. Trading frequency is a proxy for the more commonly used number of trades measure of liquidity.20 The bid-ask spread has been widely used to measure liquidity since the early contribution of Amihud & Mendelson (1980), a more recent example being Elder et al. (2005). Trading frequency is the proportion of open days a share was traded in any given period. The bid-ask spread for each share is calculated from its observed daily high and low price using the method proposed by Corwin & Schultz (2009). This uses the insight that daily high prices are invariably buy orders and daily low prices are usually sell orders. Hence, the high–low ratio on any day reflects the bid-ask spread and the fundamental volatility of a stock. In addition, the part of the high–low ratio due to volatility will increase proportionate to the length of the trading interval, whereas the part due to the bid-ask spread will be constant over different intervals. Assuming geometric Brownian motion, Corwin & Schultz (2009) show that the daily bid-ask spread for a share (SPt) can be estimated as , where and γ = [ln(Ht,t+1/Lt,t+1)]2; Ht, Lt are the observed daily high and low prices of the share and Ht,t+1, Lt,t+1 are the high and low prices taken over the 2 days t, t + 1.21
Average trading frequency for A shares increased significantly between the pre- and post-BOLT periods, and positive changes in trading frequency dominate negative changes (Table 7). For B shares, the average trading frequency also increased significantly, but the difference is not as large or as significant as for A shares. However, comparing A and B shares directly, the increase in A shares’ liquidity was not significantly larger than that of B shares. In addition, there is no evidence of an association between each firm’s listing flag (A or B) and the magnitude of the change in its share liquidity. Evidence from the estimated changes in bid-ask spreads is more decisive (Table 8). There is a significant decline in the average spread between pre- and post-BOLT for both A and B shares. Furthermore, the decline is significantly larger for B shares as evidenced by the t-statistics when we compare the changes between A and B shares. Overall, there was a significant improvement in share liquidity following the transfer to BOLT, particularly for B shares, and this would help explain the positive price reaction to the new system as well as the stronger reaction of the less liquid B shares.
Table 7. Trading frequency before and after BOLT aSix firms were excluded from sample A due to there being no pricing data in the early part of the pre-Bombay on-line trading (BOLT) period. bTrading days for firm k indicates the number of days in which there was a trade in the firm’s shares on the BSE. cTrading frequency is the number of trading days as a proportion of the number of days that the BSE was open during the period. Pre-BOLT includes 340 open days, whereas post-BOLT includes 351 open days. Panel 2 of Table 3 gives details for each period. dChange in the trading frequency of firm k is calculated as the difference between the stock’s trading frequency: DTFk = TFpost,k − TFpre,k. et-test on the mean change tests the null hypothesis that there is no increase in trading frequency. The null and alternative hypotheses are: H0 : μ = 0, H : μ > 0; μ = mean DTF. The test is of the form: t-statistic = E(DTFk) N1/2/SD(DTFk), where N is the number of observations. Asterisks show the significance level for a one-tailed test. fBinomial test on change: (+:−) compares positive and negative changes in trading frequency. The null hypothesis is that positive and negative changes are equally likely. The test returns the probability of getting (+:−) under the null hypothesis. The asterisks indicate the significance level of rejecting the null. gThe F-test is of the form: F-statistic = σ2B/σ2A ∼ F(NB − 1, NA − 1); σ2B and σ2A are the respective variances of the A and B samples, and NA, NB are the respective number of observations. hThe t-test is of the form: . The pooled variance is: . iThe t-test is of the form: . jThe chi-squared test is a contingency table test of the median calculated by classifying all scores: first, as being above, at or below the median of the combined sample, and second as belonging to sample A or B. The test is of the form: ; O is the observed value and E is the expected value of each cell and is given by its row total times its column total divided by the total number of observations. Significance at ***1% and **5% levels.
Panel 1: 63 A stocksa
Panel 2: 83 B stocks
Trading days: pre-BOLTb
Trading frequency: pre-BOLT (TFpre)c
Trading days: post-BOLTb
Trading frequency: post-BOLT (TFpost)c
Change in trading frequency (DTF)d
t-test on mean change:eμ = 0
Binomial test on changef (+:–)
Panel 3: Comparison of the change in trading frequency of A stocks versus B stocks
σ2A = σ2B = σ2
σ2A < σ2B
μA = μB (assuming σ2A = σ2B)
μA ≶ μB
μA = μB (assuming σ2A <σ2B)
μA ≶ μB
Listing flag and size of change are not associated
Listing flag and size of change are associated
Table 8. Estimated bid-ask spreads before and after BOLT aSix firms were excluded from sample A due to there being no pricing data in the early part of the pre-Bombay on-line trading (BOLT) period. bEach day’s bid-ask spread for any security (SPt) is calculated using the formula: ; where ; γ = [ln(Ht,t+1/Lt,t+1)]2; Ht and Lt are the observed daily high and low prices of the share; and Ht,t+1Lt,t+1 are the high and low prices taken over the 2 days t, t + 1. The average spread for each firm is the time mean of the daily spreads calculated over the pre- and post-BOLT periods respectively. The summary statistics reported (mean, median, etc.) are calculated from the cross-section of the individual firm averages. cChange in average spread of each firm is calculated as the difference between the post-BOLT average spread and the pre-BOLT average. dt-test on mean change tests the null hypothesis that there is no decrease in the average spread. The null and alternative hypotheses are: H0: μ = 0; H : μ < 0, where μ = mean SP. The test is of the form: t-statistic = E(SP)N1/2/SD(SP), where N is the number of observations. Asterisks show the significance level for a one-tailed test. eBinomial test on change: (+:−) compares positive and negative changes in the average spread across firms. The null hypothesis is that positive and negative changes are equally likely. The test returns the probability of getting (+:−) under the null hypothesis. The asterisks indicate the significance level of rejecting the null. fThe t-test is of the form: . The pooled variance is: . gThe t-test is of the form: . Significance at ***1% and **5% levels.
Panel 1: 62 A stocksa
Panel 2: 80 B stocks
Average spread (SP): pre-BOLTb
Average spread (SP): post-BOLTb
Change in Average spreadc
t-test on mean change:dμ = 0
Binomial test on changee (+:–)
Panel 3: Comparison of the change in the spread of A stocks versus B stocks
μA = μB (assuming σ2A = σ2B)
μA ≶ μB
μA = μB (assuming σ2A < σ2B)
μA ≶ μB
5.2. Weak-form Efficiency
Electronic trading promotes rapid dissemination of information. Improved transparency, information flows, and enhanced liquidity should all improve the price discovery process. As B shares are less heavily traded, there should be less public information available for them than for A shares. Therefore, we might expect improvements in the information content of prices following BOLT to be particularly marked for B shares. To check this, we conducted weak-form efficiency tests using the Ljung–Box Q-statistic:
Here, is the sample autocorrelation at lag τ, m is the maximum lag, T is the sample size, and QLB ∼ (Diebold, 1998). The Q-statistic is a portmanteau test and, therefore, has low power against alternatives that involve intermittent spikes in the autocorrelation function. In general, we would not expect to encounter high-order autocorrelations in daily share returns, and we would argue that the test is, therefore, adequate for the present purpose. However, as a check on its robustness, we computed the Q-statistic for maximum lag lengths of m = 1,…, 5.
For both A and B shares there is a decrease in the Q-statistics between the pre- and post-BOLT eras (Table 9). The changes are significant for B shares but not for A shares. However, for both groups, the mean Q-values are typically much larger than the medians, suggesting that the means may be driven by a few large Q-statistics. Therefore, we focus attention on the binomial tests that compare the number of Q-values below the critical value with those that are at or above it. These show that there are always significantly more Q-statistics below the critical value than those that are at or above it, before as well as after BOLT. This suggests that the BSE might have been weak-form efficient for some of the sample before the introduction of BOLT. However, the mean and median changes in Q-values between the pre- and post-BOLT periods are all negative, and the binomial tests show that negative changes in Q were more likely than positive changes, both of which suggest an improvement in efficiency for A and B shares. In addition, the magnitudes of the changes in Q are larger for B shares than for A shares, and the mean change in Q is significant for B shares but not for A shares. Therefore, there is strong evidence of an improvement in efficiency for B shares but less so for A shares.
Table 9. Efficiency: Return predictability for A shares before and after BOLT Summary statistics for autocorrelations and Ljung–Box tests using one to five lags of the daily returns on each share, before and after Bombay on-line trading (BOLT) are shown. aSix firms were excluded from the sample of 69 A firms. This is due to missing pricing data for part of the first period. bQLB. The Ljung–Box Q-statistic: ; where T is the number of daily returns, ρ2(τ) is the autocorrelation at lag τ, and m is the maximum lag. Under the null hypothesis that the time series of returns is white noise, QLB is distributed as χ2m. The change is defined as: dQLB = QLBpost-BOLT − QLBpre-BOLT. cThe Pre-BOLT period includes 340 open days, whereas the post-BOLT includes 351 open days. Panel 2 of Table 3 gives details for each period. dt-test on QLB. The t-test on the mean QLB is defined as: with degrees of freedom 62 (A) or 82 (B). The critical value for the appropriate chi-squared with which the mean QLB is compared is given in the χ2(5%) column. The null and alternative hypotheses are: and . Interpretation of the alternative hypothesis depends on the tail at which the t-statistic falls: • if t > 0, implies that we reject efficiency at above the 5% level; • if t < 0, implies that we cannot reject efficiency. et-test on the change in QLB. The t-test is defined as: t = μN1/2/SD(μ). The null and alternative hypotheses are: H0: μ = 0, H: μ ≠ 0, where: μ is the mean value, and N is the number of observations. fBin: Binomial test on QLB. For the binomial test, (x:y) gives the number of observations for which QLB is less than (x) or greater than or equal to (y), the relevant critical value. The null hypothesis is that and are equally likely. The test returns the probability mass function, which is the probability of getting x:y under the null hypothesis. The asterisks show the significance levels associated with rejecting the null. gBin: Binomial on the change in QLB. The binomial (x:y) gives the number of observations for which dQLB is negative (x) and the number of observations for which dQLB is positive (y). The null hypothesis is that negative and positive changes are equally likely. The test returns the probability mass function, which is the probability of getting x:y under the null hypothesis. The asterisks indicate the significance level of rejecting the null. h|ρ(1)| is the absolute value of the autocorrelation at lag 1. Significance at ***1%, **5%, and *10% levels.
t-test on QLBd: μ = χ25%, t-test on changee: μ = 0
Bin.: QLB(<χ25%:≥χ25%)f; change: (−:+)g
Panel 1: 63 A stocksa
QLB max lag = 1b
Change in QLB
QLB max lag = 2
Change in QLB
QLB max lag = 3
Change in QLB
QLB max lag = 4
Change in QLB
QLB Max lag = 5
Change in QLB
Panel 2: 83 B stocks
QLB Max lag = 1
Change in QLB
QLB max lag = 2
Change in QLB
QLB max lag = 3
Change in QLB
QLB max lag = 4
Change in QLB
QLB max lag = 5
Change in QLB
Finally, the sample autocorrelations at lag 1 for both A and B shares are statistically and economically significant in both the pre- and post-BOLT eras. There is a decline in the autocorrelation coefficient for A and B shares, as between pre- and post-BOLT, but once again, the decline is larger for B shares. Therefore, overall, we conclude that there was an increase in efficiency following BOLT, which was concentrated particularly in the less liquid B shares. Clearly, this finding is also in line with the larger CAR found for the B shares.
Electronic trading might be expected to improve the dissemination of information, to reduce the preponderance of uninformed investors, and to thereby reduce return volatility (Admati & Pfleiderer, 1988). However, other aspects of market microstructure tend to increase volatility. Easley & O’Hara (1991) show that, in a market where dealers coexist with a public order book, returns are likely to be more volatile than in a pure dealer market. As BOLT included the introduction of an order book alongside dealers in the BSE, we might expect an increase in volatility. Perhaps the best documented empirical regularity between volume and volatility is that of a positive correlation between volume and volatility (Bessembinder & Seguin, 1993). Therefore, increased volume and liquidity following BOLT might lead to an increase in return volatility. In short, even if the direct effect of electronic trading is to reduce volatility, indirect effects might produce an increase. However, if volatility decreased following BOLT, given that we have already established that liquidity and efficiency increased, this would be a further and strong indication that it was the move to electronic trading per se that reduced volatility.
Our volatility comparisons began with β coefficients estimated using the market model, and representing firms’ systematic risk. Given the evidence that the market model is incomplete as a model of security risks (e.g. Fama, 1991), we also examined changes in total risk: standard deviations of share returns. This includes firm idiosyncratic risk, which arguably might be particularly affected by BOLT. We estimated βs in the pre- and post-BOLT eras and compared their cross-sectional means and variances over time and between A and B shares (Table 3). In a large portfolio, the cross-section of βs may be expected to have a mean of unity. Deviations from unity may be regarded as sampling variation, and we compared the means mainly as a check on this point. To assess the change in volatility, we compared the cross-sectional standard deviation of βs before and after BOLT. If the spread of the βs (their standard deviations) increased we can assert that the range of risks of stocks increased and vice versa. In fact, the standard deviation of the βs decreased for A shares but increased for B shares (Table 10). However, both A and B shares exhibited a significant decrease in total risk in the pre- and post-BOLT eras. Therefore, volatility mostly decreased following BOLT, with the decrease being slightly more marked for B shares.
Table 10. Volatility before and after BOLT: Systematic risk and total risk Summary statistics for β coefficients and SDs of returns before and after Bombay on-line trading (BOLT) are shown. aSix firms were excluded from sample A due to there being no pricing data in the early part of the pre-BOLT period. bPre-BOLT period includes 340 open days, whereas post-BOLT includes 351 open days. Panel 2 of Table 3 gives details for each period. cβ is estimated from the market model with five leads and five lags and 56 event dummies. The model is as specified in Equation (6) and the market index used is the BSE200. The change in β ≡ βchange = βpost-BOLT – βpre-BOLT. dSD of daily returns in the period. The change in SD ≡ SDchange = SDpost-BOLT − SDpre-BOLT. eThe F-tests in Panel 1 test the null hypothesis that there is no reduction in volatility as measured by the SD of β. fThe binomial test in panels 1 and 2: (x:y) shows positive (x) and negative (y) changes in volatility. The null hypothesis is that positive and negative changes are equally likely. The test returns the probability mass function, which is the probability of getting x:y under the null hypothesis. The asterisks indicate the significance level of rejecting the null. gThe t-test in Panel 2 is a test of the null hypothesis that there is no reduction in volatility as measured by total risk. The null and alternative hypotheses are: H0: μ = 0, H: μ ≠ 0, where μ is the mean change in total risk. The test is of the form: t-statistic = E(change)N1/2/SD(change), where N is the number of observations. Significance at ***1%, **5%, and *10% levels.
63 A stocksa
83 B stocks
Panel 1: Systematic risk: βc
Change (Post − Pre)
Binomial on change (+:−)f
Panel 2: Total risk: SD
Change (Post − Pre)
t-test on mean change: μ = 0g
Binomial on change (+:−)
6. Effects of Liquidity, Efficiency and Volatility Improvements on Cumulative Abnormal Returns
Were the LEV improvements responsible for the increase in value observed in the positive CAR? The evidence for this is, we believe, persuasive, but still circumstantial. The attribution of differential CAR for A and B shares to differential LEV improvements is likewise circumstantial. Therefore, we aim to sharpen this evidence by studying the cross-sectional relationship between individual CAR and LEV improvements. This is important in India where several financial reforms occurred over the period within which we conducted our study, and there is a risk of cross-event contamination. Few microstructure reforms take place in isolation, and we would argue that all such event studies should include a cross-sectional test.22 To test the hypothesis that the LEV changes explain the distribution of CAR, we estimate cross-sectional regressions of the form:
CARt,k is the CAR on the kth stock through day t; Zj,k are the stock-specific LEV that may have produced the changes in share values measured by the CAR; and εk is the regression error. We ran regressions using CAR through days 10, 15, 20, 30, and 50, as it could be argued that later dates in the window would be more subject to contamination by extraneous events. In fact, the results of these regressions were qualitatively similar, and we therefore report only the results using three representative days: the CAR through days 10, 20, and 50.23
Zj,k includes the following variables. First, are the changes between the pre- and post-BOLT eras in: liquidity (DLIQ), which we measure either by trading frequency or the spread; Q-statistic for i lags (DQi); β coefficient (DBETA); and standard deviation (DSD). For liquidity, we ran the regressions using first trading frequency and then the spread. Results using the former were rather more clear-cut, possibly because of a somewhat larger sample.24 However, for comparative purposes, we report the 20- and 50-day regressions using trading frequency, and the 10-day regressions using the spread measure. For the Q-statistic we ran regressions with lag length varying from one to five. The results were qualitatively very similar and we therefore report only the results using five lags (Q5). We included both β and the standard deviation as they correspond to two quite different concepts of risk.
The second set of variables in the Zj,k includes the levels of LEV in the pre-BOLT era (PLIQ, PQi, PBETA, and PSD). To understand the rationale for these variables, consider as an example, firms with low liquidity before the introduction of BOLT. Illiquid shares would not normally be tracked by brokers. If, following the introduction of BOLT, there was an improvement in these firms’ liquidity, there are two possibilities. The first is that trading increases and brokers now find it worthwhile to track the stocks: this creates a secondary improvement in liquidity, implying a significant positive coefficient on PLIQ. Alternatively, the increase in trading might not be sufficient to induce brokers to track the stocks, implying a negative or zero coefficient.
Third, our base hypothesis implies that differences in CAR between A and B shares can be explained by the extent of different LEV improvements as between individual shares, not by a different marginal response of the different groups to the same improvements. However, we also examined whether BOLT had a different marginal impact on A and B shares, especially because it apparently had a larger total effect on (the CAR of) the B shares. A different marginal effect could occur because microstructure improvements were more highly valued by investors in B shares.25 This could be the case if, for example, investors in B shares had more to gain from these improvements compared with investors in A shares. We include a dummy (FLAG) to control for intercept differences in the response of A and B shares:
We also include eight interaction terms defined as: FLAG × Zj,k (Zj = DLIQ, DQi, DBETA, DSD, PLIQ, PQi, PBETA, PSD). If B shares had a stronger marginal response to the reforms we would expect to find these interaction terms all positively signed. It transpired that only the interactions involving DLIQ and PLIQ were significant in the regressions, and, therefore, we report only the results including these variables. However, we checked our results with a Chow test for a split between A and B shares on all the non-interacted variables in the regression.
Fourth, we controlled for possible spillovers from the NSE (Shah & Thomas, 1996) by including dummies for firms that were quoted on the NSE prior to the introduction of BOLT (NSEPRE), and those that were introduced to the NSE in the event window (NSEEVENT). If there were spillovers, we would expect shares quoted on the NSE to benefit less from the introduction of BOLT than shares that were quoted only on the BSE. Therefore,
However, neither of these variables was significant in any of the regressions, strongly suggesting that our tests are not affected by the establishment of the NSE, and we therefore do not report the coefficients for either NSEPRE or NSEEVENT. Hence, in summary, we report results of the following regression:
for CAR through days 10, 20, and 50. We expect β1 > 0, γ1 < 0, δ1 < 0, θ1 < 0; β2, γ2, δ2 and θ2 to have either sign; the signs of β3, β4 and ϕ will depend on whether there is a differential effect of LEV on A and B shares26; and λ1, λ2 will depend on whether there are spillovers from the NSE.
Up to 24% of the cross-sectional variation in the CAR can be attributed to variations in LEV (Table 11), and this percentage is clearly significant. However, as LEV improvements are the identified sources of the added value in BOLT, it would seem reasonable to argue that our initial estimate of added value, based on the CAR alone, should be scaled back using the correlation coefficient. This would put a more conservative value on BOLT of 0.24 × Rs 196bn or approximately 0.5% of India’s 1994–1995 GDP (Rs 9434bn). Most of the coefficients for the changes in liquidity, efficiency, and β have the expected sign, although efficiency is not significant and β is only marginally significant. The main exception is the sign on DLIQ in the 10-day regression, which, perhaps not surprisingly, has weaker explanatory power than the regressions at 20 and 50 days. The change in total risk is significant but positive, consistent with the argument that trading improvements may be associated with increased volatility. The sign on DSD underlines the importance of these regressions. The mean and median values of total risk decreased (Table 10), but there is a positive association between firm-specific changes in total risk and the firm-specific CAR. A simple before-and-after comparison does not bring out this point. Two of the variables measuring initial conditions are significant, and these are signed as expected: for example, higher initial liquidity creates a positive secondary effect, except for the 10-day CAR.
Table 11. The contribution of liquidity, efficiency, and volatility to the price reaction to BOLT OLS regressions of Equation (9) using the cross-section of 146 A and B shares are shown. aThe dependent variable (cumulative abnormal return [CAR]) is a firm’s CAR at day T + 50, T + 20 or T + 10. The sample size is 142 for regressions using T + 10 data because the spread measure of liquidity is used, and these data are available for fewer firms. The t-statistics are White-corrected for heteroskedasticity. Explanatory variables: LIQi is a measure of share liquidity: for CAR at T + 50 and T + 20, trading frequency is used to measure share liquidity; for CAR at T + 10, the estimated bid-ask spread is used. To facilitate comparison between the regressions using trading frequency and the spread as liquidity measures, the spread is entered with sign reversed. Therefore, DLIQ is measured so that its expected sign is positive in all the regressions. Qτi is Ljung–Box Q-statistic for τ lags; βi is the estimated β coefficient; SDi is the standard deviation of returns. Variables prefixed by D give the change between the pre- and post-BOLT eras. Variables prefixed by P give the level of liquidity, efficiency or volatility in the pre-BOLT era. FLAGi is a dummy that equals 0 for A shares and 1 for B shares. bLMHET is the Breusch–Pagan test for heteroskedasticity. cChow is the Chow test for a structural break between the parameters for A and B shares. In this model, a general Chow test is equivalent to including a complete set of variables interacted with FLAG, and testing whether there is a significant decrease in explanatory power when these interactions are removed. Therefore, each reported Chow test checks only those variables that are not already interacted; for example in column 2 of Table 11: DQ5, DBETA, DSD, PSD. Significance at ***1%, **5%, and *10% levels.
CAR after 50 days (using trading frequency)
CAR after 20 days (using trading frequency)
CAR after 10 days (using spread, with sign reversed)
DLIQ × FLAG
PLIQ × FLAG
χ25%(1) = 3.84
F(6,126) = 2.17
F(4,130) = 2.44
F(6,130) = 2.17
F(6,130) = 2.17
F(6,126) = 2.17
F(4,130) = 2.44
F(6,130) = 2.17
F(6,130) = 2.17
F(6,122) = 2.17
F(4,126) = 2.44
F(6,126) = 2.17
F(6,126) = 2.17
F (zero slopes)
4.7068***F(9,136) = 1.95
5.5935***F(7,138) = 2.08
6.1078***F(7,138) = 2.08
4.8095***F(9,136) = 1.95
5.4572***F(7,138) = 2.08
6.1372***F(7,138) = 2.08
1.8219*F(11,130) = 1.86
2.0792***F(9,132) = 1.95
1.8157*F(7,134) = 2.08
2.0720*F(7,134) = 2.08
No. of firms
Turning to the differences between A and B shares, we see that in the 50- and 20-day regressions, the interaction terms with DLIQ and PLIQ are significant when DLIQ and PLIQ are excluded; whereas DLIQ and PLIQ are significant when the interaction terms are excluded. When all four variables are included, the interaction terms are marginally significant, but DLIQ and PLIQ are not. Given that the FLAG dummy is highly significant we would argue that the interaction terms should be retained, giving substantial support to the (reasonable) conclusion that liquidity improvements were significantly more valuable for the less liquid B shares.27 For the 10-day CAR using the spread measure of liquidity, the regressions have less explanatory power,28 but DLIQ and PLIQ and the interaction terms are weakly significant when they are included together. The Chow tests confirm that there is no evidence of any other significant interactions in the model.
The most important finding here is that there is a reliable, significant cross-sectional relationship between the value of BOLT (measured by the CAR) and most of its conjectured effects in terms of measured improvements in LEV. The results also show that different types of risk may have different effects in the context of market microstructure reforms. In addition, the significant interaction terms strongly suggest that the much stronger reaction to BOLT of the less liquid B shares was due in part to the greater value placed on increased liquidity by investors in these shares. This offers important new evidence that improved liquidity can increase security values, as argued by Lee (2009). Overall, these results underline the need to examine the cross-sectional relationships among CAR and the variables that are thought to determine them in any market microstructure event study.
The aim of this paper was to assess the improvement in the functioning of India’s main equity market, the BSE, following implementation of new technology and automation of the trading system by BOLT. We did this by studying the impact of BOLT on a sample that comprises a substantial proportion of the market capitalization of the BSE and includes the most liquid (A) and less liquid (B) shares. We found that the impact of BOLT, measured by the CAR over a 56-day event window, was substantial: the average CAR for the A shares was approximately 4.5%, whereas that for the B shares was more than twice as much (over 12%). Taking a conservative figure of 4.5% for the whole market, and using only the explainable part of the CAR, the estimated impact of BOLT was equivalent to approximately 0.5% of India’s GDP: a substantial gain.
In the second part of the paper, we sought a more detailed explanation for the CAR in terms of improvements in LEV accompanying BOLT. The liquidity of A shares increased, as did B shares, but significantly more so (on the spread measure). Efficiency also increased, and the increase was greater for B shares. For volatility, the systematic risk of A shares decreased somewhat, but the systematic risk of B shares increased significantly. However, there was a significant decrease in the total risk of both A and B shares, with a stronger decrease for the B shares. Regression analysis provided additional evidence that the CAR could be explained by cross-sectional variations in firm-specific LEV improvements. Overall, the results suggest that BOLT improved the BSE’s market microstructure for both A and B shares, with liquidity improvements being particularly important in contributing to the more marked improvement in the value of the less liquid B shares. There is no evidence of a direct spillover effect on shares that were also quoted on the NSE.
Our main conclusions are as follows. First, we confirm existing evidence that improvements in trading arrangements can have significant positive value in an emerging market. Second, and more importantly, we find that in India liquidity improvements were especially important in contributing to the overall success of the reforms but more particularly in improving the performance of the less liquid B shares. Less liquid shares also tend to be less efficiently priced and exhibit greater idiosyncratic risk, and the changes in both these factors were appreciably more marked for B shares than for A shares. This, in turn, suggests that liquidity improvements may be particularly important in microstructure reforms in emerging markets, which are invariably less liquid than the more developed markets. Clearly, an important topic of further research is to identify more fully the different aspects of microstructure reform that foster liquidity improvements in emerging markets.
Exceptionally, The New York Stock Exchange maintains floor trading in parallel with electronic trading.
The BSE is usually known by its acronym, which uses the former name of the city, Bombay; hence, “BSE”. It is independent of and competes for business with the NSE.
The BSE was established in 1875. In the period covered by the present study it was second only to New York in the number of domestic companies quoted on the exchange.
There were a series of pre-announcements about BOLT, which would make it difficult to pin down a single announcement date. Allowing for announcements would suggest, if anything, that we should revise upwards our estimates of the positive impact of BOLT.
The history and infrastructure of the Mumbai Stock Exchange are set out by the BSE at: http://www.bseindia.com. Unless otherwise stated, our discussion in this section draws heavily on this source.
Badla was reintroduced the year following the BOLT reforms and, therefore, does not affect our analysis.
The NSE began trading debt in June 1994, and equities in November 1994. By December 1995, 1300 stocks were traded and the NSE had become more liquid than the BSE (Shah & Thomas, 2001).
Over time, the order book gradually assumed greater importance, and in 2001 the facility for placing quotes on the system was removed altogether so that BOLT is now strictly an order-driven system.
The distinction between A and B shares is explained below.
In 1995, there were approximately 100 A shares, and more than 5000 B shares. In 1996, B shares were split into B1 and B2 to reflect the greater liquidity of the former. In 1999, a “Z” group was introduced for companies that had failed to: comply with listing requirements, resolve investor complaints, or make arrangements for dematerialization.
Following introduction of BOLT, members of the exchange were permitted to open trading terminals within Mumbai and later in other cities.
Cash volume is often used to measure liquidity, but Jones et al. (1994) argue that the number of trades provides a better measure. In the present study, we cannot use either cash volume or number of trades, as there are insufficient consistently reported data available on these variables for all the shares in our sample, especially but not exclusively in the pre-BOLT period.
Corwen and Schultz (2009) give the formula: Spread=2(eSP − 1)/(1 + eSP). However, the following is easily calculated: Spread ≈ SP for 0 ≤ |SP|. At any likely level of accuracy in the calculation Spread = SP.
Lasfer (1996) carries out a similar test in the context of an event study of dividends.
The results of these regressions are available upon request from the authors.
There are more firms with missing spread data than with missing trading frequency data (Tables 7 and 8). This does not affect the comparison tests, as the degrees of freedom are adjusted accordingly.
We thank an anonymous referee for this point, and for the idea of including interaction terms in the model.
To facilitate comparison between regressions using trading frequency and the spread as liquidity measures, the spread is entered with the sign reversed. Therefore, DLIQ is measured so that its expected sign is positive in all the regressions.
The change in sign of FLAG when the interactions are excluded suggests that FLAG is doing the work of the interaction variables when they are not present, and gives further support to the view that they should be included.
This is also true for the 10-day CAR using trading frequency as the regressor.