The paper titled, “The causal impact of algorithmic trading on market quality”, written by Nidhi Aggarwal and Susan Thomas, systematically identifies the casual relationship between Algorithmic trading intensity and market quality features in the Indian markets. In this post, I will briefly summarize the contents of the paper :

What is the paper about ?

How does algorithmic trading (AT) affect the quality of securities markets ? is something that is always discussed/ analyzed/ debated by old and new market participants. Academic literature on this topic is  predominantly based on US markets. Not surprising as US markets are considered to be one of most technologically sophisticated markets. There are a few problems when one wants to answer the AT impact in US markets, i.e.

  • It is an fragmented market and hence it makes it difficult to understand the causal impact of a specific market microstructure feature.

  • One has to infer whether the order / trade is AT or non AT as the exchanges do not provide any concrete flag indicator. Even if they provide, the sample set of securities shrinks down considerably

  • Endogeneity bias

In the light of above problems, the authors say that their analysis and the dataset used, has three advantages:

  • Co-lo facility introduced by NSE is used as an instrument variable to counter endogeneity bias

  • Matching techniques can be used as there is wide span of data available

  • Since most of the trading happens on NSE, comprehensive coverage of securities are used to create better treatment and control groups for comparison purposes

The dataset used in the paper is all the tick data of equity orders and trades from the NSE between 2008-2013. The variables used in the analysis are AT intensity, quoted spread, impact cost, depth , price risk , liquidity risk, variance ratio and kurtosis. The two time periods used in the analysis are Jan 1, 2009 - Dec 31, 2009 and July 1, 2013 - Aug 31,2013. The challenge in establishing causal link from AT to market quality is that there could be a third variable that is driving both AT and market quality of an instrument. The authors find that large firms saw a significant and uniform increase in the level of AT intensity. But there was a set of medium to small sized securities that showed heterogenous AT intensity after the colo, thus giving authors a good dataset of treatment and control data.

Main steps of the analysis :

  • Start with a sample of 1577 securities listed on NSE in Aug 2013

  • Prune down the set so that securities had at least 50 trades per day during pre co-lo and post co-lo. This resulted in 918 securities. As a part of providing summary statistics, the authors mention that an analysis comparing 2009 and 2013 suggests that non-AT orders still constitute a significant part of orders demanding and supplying liquidity

  • Compute the AT difference between post-colo and pre-colo for all the 918 securities. The securities whose AT activity stands at less than 30% percentile are tagged as control group and the securities whose AT activity stands at more than 70% are tagged as treatment group. This procedure yields 276 securities in each of the treatment and control groups

  • Since this is an observational study, it would be incorrect to directly compute the market quality difference between treatment and control groups. One has to address selection bias. This is done via propensity score method. For a set of covariates, a logit model is used to compute the propensity score. Based on the propensity score, each security in the treatment group is matched with a security in the control group. This procedure yields 91 securities in the treatment group and 73 securities in the control group.

  • Density plots of propensity scores and mean test of various covariates are given to show the effectiveness of matching procedure

  • An additional matching is done based on market-wide, economy-wide differences between pre-colo and post-colo time periods. This whittles down the comparative days to 59 days.

The final dataset used for Difference in Difference-in-Difference regressions is 91 treatment group securities, 73 control group securities , observed on 59 days before and 59 days after co-lo.

Findings :

The results are based on the coefficient of interaction variable in DID regressions

  • Spread, Impact cost and Order imbalance have gone down for treatment group as compared to control group

  • Price risk and liquidity risk have gone down for treatment group as compared to control group

  • Higher AT securities show less persistence in intraday high-frequency returns

  • There is a reduction in extreme price movements for securities with higher AT

  • The final section of the paper includes simulation tests to check the robustness of results. From the same control group, a placebo group is tagged as treatment group and DID regressions are run,  1000 times. The results validate the conclusion from the original dataset that there is no impact of market quality in the absence of change in AT intensity. Variations in matching design are also done in order to check the robustness of the results.

The takeaway from this paper is that AT has been good for the Indian markets as it has decreased impact costs, improved price efficiency, reduced liquidity risk and reduced extreme price movements.