summary causal impact

3
The Causal Impact of Algorithmic Trading on Market Quality : A Summary Rishabh Shukla Policy makers, worldwide, are exploring ways of curbing the use of Algorithmic Trading (AT), which is allegedly used for market manipulation (layering and spoofing, momentum ignition etc.) in general and is deemed to be responsible for ‘flash crashes’ that are brief periods of extremely high volatility. The paper by Nidhi Aggarwal and Susan Thomas (2014) aims to analyse the impact of AT on market quality (of which quoted market depth and spreads are indicators) in electronic limit order book markets to test the hypothesis that AT adversely affects market quality and that AT leads to higher incidence of flash crashes (of which kurtosis of security returns is an indicator). The paper uses the NSE dataset which accounts for 75% of equity spot trading and 100% of derivatives trading in the country during the period of analy- sis,and includes every security traded on NSE in the analysis window. Thus market quality can be analysed at the level of the overall financial system and problems related to fragmented trading can be mitigated, unlike in advanced countries. To address the endogeneity issue (since unobserved factors can affect market quality and AT at the same time), the paper exploits the exogenous event of introduction of co-location facilities ( resulting in reduced latency and increased bandwidth leading to higher use of AT) which were made available at NSE, which directly affected the level of AT but not the market quality. The experimental design adopted identifies pairs of securities using the propen- sity score matching algorithm; which are otherwise identical in characteristics ( found using observable covariates) but differ in the level of AT. The securities which undergo a large change in AT activity after co-lo are referred to as the ‘treated group’ and those which do not show significant change in the level of AT after co-lo (but were similar to the treated group before co-lo) are referred to as the ‘control’ group.In order to control for unobserved factors like changes in macroeconomic conditions, periods with similar level of market volatility are chosen in the pre and post co-lo regime. A difference-in-differences (DiD) regression is used to estimate the effect of AT (the treatment) on the market quality (the outcome). Market Quality indicators like transaction costs, market depth, market risk, efficiency and kurtosis are used 1

Upload: rishabh-shukla

Post on 10-Sep-2015

214 views

Category:

Documents


1 download

DESCRIPTION

Causal Impact of AT

TRANSCRIPT

  • The Causal Impact of Algorithmic Trading on

    Market Quality : A Summary

    Rishabh Shukla

    Policy makers, worldwide, are exploring ways of curbing the use of AlgorithmicTrading (AT), which is allegedly used for market manipulation (layering andspoofing, momentum ignition etc.) in general and is deemed to be responsiblefor flash crashes that are brief periods of extremely high volatility.

    The paper by Nidhi Aggarwal and Susan Thomas (2014) aims to analyse theimpact of AT on market quality (of which quoted market depth and spreads areindicators) in electronic limit order book markets to test the hypothesis thatAT adversely affects market quality and that AT leads to higher incidence offlash crashes (of which kurtosis of security returns is an indicator).

    The paper uses the NSE dataset which accounts for 75% of equity spot tradingand 100% of derivatives trading in the country during the period of analy-sis,and includes every security traded on NSE in the analysis window. Thusmarket quality can be analysed at the level of the overall financial system andproblems related to fragmented trading can be mitigated, unlike in advancedcountries. To address the endogeneity issue (since unobserved factors can affectmarket quality and AT at the same time), the paper exploits the exogenousevent of introduction of co-location facilities ( resulting in reduced latency andincreased bandwidth leading to higher use of AT) which were made available atNSE, which directly affected the level of AT but not the market quality.

    The experimental design adopted identifies pairs of securities using the propen-sity score matching algorithm; which are otherwise identical in characteristics (found using observable covariates) but differ in the level of AT. The securitieswhich undergo a large change in AT activity after co-lo are referred to as thetreated group and those which do not show significant change in the level ofAT after co-lo (but were similar to the treated group before co-lo) are referredto as the control group.In order to control for unobserved factors like changesin macroeconomic conditions, periods with similar level of market volatility arechosen in the pre and post co-lo regime.

    A difference-in-differences (DiD) regression is used to estimate the effect of AT(the treatment) on the market quality (the outcome). Market Quality indicatorslike transaction costs, market depth, market risk, efficiency and kurtosis are used

    1

  • as dependent variables for a given security i at time t. The coefficient(3) ofthe interaction term ATico lot provides the estimate of the treatment effect,which stands for greater AT intensity in the post co-lo period. A significant valueof 3 implies that AT has an impact on the market quality.Since a decrease intransaction costs and market risk, and an increase in depth, increases the mar-ket quality; the hypothesised value for 3 is zero. The alternative hypothesisthen being 3 < 0 and 3 > 0 respectively for transactions costs/market riskand market depth.

    The coefficients for QSpread and Impact Cost (Transaction Costs Indicators)are found to be negative and significant implying that higher level of AT sig-nificantly reduces transaction costs. In case of market depth indicators, the co-efficient is positive and significant, implying that higher AT intensity increasesshares available for trade in the market.

    The hypothesis for efficiency measures like variance ratio and kurtosis is that3 = 0 with the alternative hypothesis being that 3 < 0. If AT intensity doesimprove price efficiency, we expect |V R1| and kurtosis to be close to zero. Thekurtosis coefficient estimate turns out to be positive and insignificant implyinggreater probability of extreme price movements intra-day due to AT. For everysecurity i, frequency of price movements greater than a threshold price relativeto last days trading price is tested. The coefficient is negative and significantonly at the 5% threshold and insignificant at 2% and 10%, implying that higherAT intensity either reduces the probability of extreme price movements or isthe same as for low AT intensity securities.

    In order to verify the robustness of the results that the experimental designyields, a placebo is simulated in which a subset of the treatment group is taken,which is unaffected by the intervention i.e. has a low level of AT activity andis matched with low level AT activity securities before intervention. The DiDestimates for such a treatment and control group should not be significantly dif-ferent from zero. This hypothesis is rejected less than 5% of the time implyingthat a change in the market quality does not take place without a change inthe level of AT intensity. Thus the experimental design does not suffer from theplacebo effect.

    The sensitivity to match design is tested by dropping each of the selected co-variates one at a time to yield a new dataset to run the DiD regression on. Theestimated 3 coefficients are analysed for change in sign due to the modificationof the matching co-variates implying a change in the impact of market qualitydue to the level of AT. The estimates for few Depth parameters and Kurtosisdo change implying weakly established causality for these indicators.

    The paper concludes that higher AT intensity reduces intra-day liquidity for se-curities and leads to either fewer flash crashes or has no effect at all. Thus thereare more benefits than costs to securities that attract higher level of AT activity.

    2

  • 3