Retail Order Flow Segmentation

Transcription

1 Retail Order Flow Segmentation by Corey Garriott and Adrian Walton Bank of Canada 3rd Annual Conference on Financial Market Regulation May 13, 2016 Discussion by Matt Ringgenberg (Currently) (As of July 2016!)

2 1 Overview 2 Comments Comment 1 - Definition of Treatment Firms Comment 2 - Identification Assumptions Comment 3 - Alternative Hypotheses Comment 4 - Minor Comments Suggestions 3 Conclusion

3 U.S. Equity Markets have Changed Dramatically Many recent changes to U.S. markets: Decimalization, regulation ATS, regulation NMS, increase in algorithmic trading, increase in high-frequency trading, proliferation of order types and trading venues Market quality and trading behavior have also changed dramatically More frequent, but smaller, trades More trading venues = fragmented trading Different liquidity? Different risks? Net impact and causes of changes unclear

4 This Paper: Segmentation of Retail Order Flow Matters Authors focus on segmentation of retail orders Examine creation of NYSE Retail Liquidity Program (RLP) Use Harbrouck information share and DiD regression Find segmentation improves liquidity and price efficiency Overall, the paper: 1 asks an important question 2 is well-motivated by theory 3 is relatively carefully executed 4 presents new evidence Big pictures elements are all there = I m going to be extremely tough on the authors because I like the topic My main concerns are about identification

5 Comment 1 - Endogenous Definition of Treatment Firms Authors use two main methodologies: 1 Hasbrouck information share 2 Difference-in-difference (DiD) regression No critique of the Hasbrouck analysis; results are careful, interesting, and solid Worried about the DiD Authors define treatment firms as NYSE listed securities with at least 1% RLP volume This is endogenous! They are selecting treatment firms most likely to benefit from treatment

6 Comment 2 - Worried about Identifying Assumptions I agree with the DiD as a methodology here, but not the setup DiD not a panacea, still requires strong identifying assumptions 1 Assumes RLP = exogenous shock to firm liquidity 2 Assumes control firms are a valid counter-factual for treatment firms if RLP hadn t existed First, macro trends likely impact liquidity Second, would like to see parallel trends and placebo tests Third, control group likely contaminated

7 Figure 1 - RLP by other market centers Control sample is NYSE Arca listed securities (i.e., stocks not eligible for NYSE RLP) But many other RLPs were started during sample period

8 Figure 1 - RLP by other market centers Control sample is NYSE Arca listed securities (i.e., stocks not eligible for NYSE RLP) But many other RLPs were started during sample period BATS RLP NASDAQ RLP

9 Comment 2 - Changes to Identifying Assumptions Would like to see more careful discussion of identifying assumptions What would violate assumptions? Exogenously define treatment and control Include a time fixed effect y i,t = α + βd(i, t) + δ i + δ t + ɛ i,t Show evidence of parallel trends / placebo Make sure window size is appropriate

10 Example of parallel trends and placebo Example (gratuitous self cite) from Haslag and Ringgenberg (2015) paper on fragmentation ROA Volatility ROA Volatility Year COGS/Assets Volatility COGS/Assets Volatility Year Treated Control Treated Control Return Volatility Return Volatility Year CF/Assets Volatility CF/Assets Volatility Year Treated Control Treated Control See Bertrand, Duflo, and Mullainathan (QJE 2004)

11 Example of parallel trends and placebo Example (gratuitous self cite) from Haslag and Ringgenberg (2015) paper on fragmentation Distribution of Placebo t-stats Draws Fraction Return Volatility Fraction ROA Volatility Fraction COGS/Assets Volatility Fraction ROA Fraction COGS/Assets See Bertrand, Duflo, and Mullainathan (QJE 2004)

12 Comment 3 - Need to Explore Alternative Hypotheses Authors focus most of discussion on information asymmetry story (Zhu (2014)) But RLP segmented orders AND allowed for sub-penny quoting Some results could be due to sub-penny quoting See Ye and Yao (2014) Segmentation is a type of fragmentation See Haslag and Ringgenberg (2015) Possible heterogeneity in outcomes due to network externalities Overall, welfare conclusions are very difficult

13 Comment 4 - A Few Minor Comments Overall, paper is really well written with clear text and tables However, language is imprecise in spots Authors commonly use listed and traded on interchangeably After NMS, this is no longer true! There is some evidence that retail traders are informed (e.g., Kelley and Tetlock (2013), (2015)) What about Retail Interest Indicator in TAQ quotes database?

14 Important paper, but I would like to see several changes Overall, I like the topic and paper, but I would like to see several changes: 1 Add a time fixed effect 2 Exogenously define treatment and control firms 3 Make window smaller to avoid contamination 4 Add placebo and parallel trend tests 5 Think about alternatives (tick size & fragmentation) 6 Examine Oster (2015) test of omitted variable bias

15 Conclusion - Important question and interesting results Overall, paper has a lot of potential Important question Well motivated Novel results Proper data Identification? Not convinced yet. Many suggestions...

16 Conclusion - Important question and interesting results Overall, paper has a lot of potential Important question Well motivated Novel results Proper data Identification? Not convinced yet. Many suggestions...

17 Conclusion - Important question and interesting results Overall, paper has a lot of potential Important question Well motivated Novel results Proper data Identification? Not convinced yet. Many suggestions...

18 Conclusion - Important question and interesting results Overall, paper has a lot of potential Important question Well motivated Novel results Proper data Identification? Not convinced yet. Many suggestions...

19 Conclusion - Important question and interesting results Overall, paper has a lot of potential Important question Well motivated Novel results Proper data Identification? Not convinced yet. Many suggestions...