Trading fast and slow: Colocation and liquidity Jonathan Brogaard Björn Hagströmer Lars Nordén Ryan Riordan Market Microstructure: Confronting Many Viewpoints #3 December 8th, 2014 Market Colocated traders Key points of the paper 1. Fast traders have – – – – 2. Introduction of 10G colocation at NASDAQ OMX Stockholm - 3. Higher order-to-trade ratios Higher market making presence Better liquidity timing (better effective spreads) Better ability to trade on short-lived information Who is buying the fastest connectivity? (mostly market-makers) What happens to market liquidity? (it improves) What is driving the liquidity improvement? – Market-makers avoiding being adversely selected – Inventory management (relaxed inventory constraint) Adverse Selection Hypothesis Fast traders have a short-term informational advantage Fast traders trade actively on news adversely select traders who do not have time to revise stale quotes (Biais, Foucault & Moinas, 2014; Cartea & Penalva, 2012; Foucault, Hombert & Rosu, 2013; Martinez and Rosu, 2013) News traders get faster Adverse selection costs increase Fast liquidity providers use speed to avoid being picked off (Jovanovic & Menkveld, 2012; Hoffman, 2014; Aït-Sahalia and Saglam, 2014) Market makers get faster Adverse selection costs decrease Inventory Hypothesis Aït-Sahalia & Saglam, 2014: • The inventory constraint of market makers depends on the accuracy of the signal on future trade flows • Faster market makers have better control of their inventory, as they can cancel quotes quickly when inventory builds up Market makers get faster Inventory costs decrease Current empirical evidence on trading speed Empirical studies on colocation events find improved liquidity but increased volatility (Boehmer, Fong & Wu, 2012; Frino, Mollica & Webb, 2013) Studies of AT/HFT show: Informed (Brogaard et al. 2013, Hendershott and Riordan 2009) Supply liquidity (Menkveld 2013, Malinova et al. 2013) … Empirical studies on trading system upgrades find mixed results Positive effects: Boehmer, Fong, and Wu (2014); Frino, Mollica, and Webb (2014); Riordan & Storkenmaier (2012) Negative effects: Hendershott & Moulton (2011); Gai, Yao & Ye (2013); Menkveld & Zoican (2013) How is this paper different than other papers? Previous papers classify traders by Exchange-defined HFT flag (Hagströmer and Norden, 2013; Brogaard et al., 2013) Trading behaviour (Kirilenko et al., 2011; Hasbrouck and Saar, 2013; Malinova et al., 2013) We identify groups based on the exchange services (colocation) they “consume”, i.e. self selection We study the behaviour & impact of these colocated/fast traders (basic, 1G, and 10G) that results from being fast Remaining agenda Data Descriptive statistics on colocated traders Who upgrades Liquidity effects Mechanism Data Colocation history and trader classification Feb 8, 2010: INET introduced Basic colocation Mar 14, 2011: Premium Colocation 1G introduced as add-on to Basic Sep 17, 2012: Premium Colocation 10G introduced Trader group N Fast vs. Slow Event study No colocation 80 NonColo NonColo Basic colocation 13 SlowColo Premium colocation 1G 11 Premium colocation 10G 12 Colo 10GColo We identify trader groups based on the colocation services they “consume”, i.e. self selection Allows investigation of traders from different speed segments Data Proprietary data from NASDAQ OMX Stockholm Data on trading entity level and colocation status Stocks in the OMX S30 index (30 largest stocks in Sweden) NASDAQ OMX order books (no MTFs) 2012 AUG SEP OCT Pre Post Aug 20 – Sept 14 Sept 17 – Oct 12 Sep 17: Nasdaq OMX introduces Premium Colocation 10G Thomson Reuters Tick History (TRTH / SIRCA) Event study on liquidity Robustness wrt index futures and consolidated order book Descriptive statistics: What fast traders do What fast traders do: Total volumes Limit orders 17% 55% 27% Trades 2% 22% 19% 56% 4% NonColo BasicColo NonColo BasicColo PremiumColo 10GColo PremiumColo 10GColo What fast traders do: Quotes and trades 57.39 10.70% 11.70% 28.98 6.05 8.17 Order to Trade Ratio 0.60% 0.20% BBO Presence BBO Presence: % of time which trading entities have orders posted at the best bid and offer in the limit order book What fast traders do: Trading performance Volume-weighted average effective spread across all trades NonColo Colo Spread (bps) 4.00 3.00 2.00 1.00 0.00 Active Trading Passive Trading What fast traders do Panel A: Trading activity Number of trading entities Share of all limit orders Share of all cancellations Share of all trades Share of all SEK trading volume Active trades per stock-day Passive trades per stockday Panel B: Trading behavior Order-to-Trade Ratio Liquidity Supply Ratio BBO Presence Inventory Crosses Zero Slow vs. Fast NonColo Colo 80 36 17.0% 83.0% 14.2% 85.8% 55.8% 44.2% 58.9% 1944.3 2103.4 41.1% 1685.1 1526.0 6.05 52.0% 0.6% 1.0 41.26 47.5% 8.2% 6.7 Segments of colocation BasicColo PremiumColo 10GColo 13 11 12 1.5% 26.9% 54.6% 1.2% 24.0% 60.6% 3.6% 18.9% 21.7% 3.7% 17.5% 19.8% 79.2 805.1 800.9 181.7 566.5 777.8 8.17 28.98 57.39 69.6% 41.3% 49.3% 0.2% 0.4 10.7% 6.9 11.7% 9.4 Inventory Crosses Zero: the number of times a trading entity changes between having long and short positions in a stock-day Who uses Colocation? High Frequency Traders? HFT is always Algorithmic Trading (AT) – but AT is not always HFT Typical properties of HFT: – Fast turnover – Low Intraday inventory – End the day neutral – High Volume (SEC, 34-61358, Concept Release on Equity Market Structure) HFT is a mixture of the use of technology and trading strategies (do they differ?) Are colocated traders different than other HFT classifications? Number of accounts Trades Hagströmer and Nordén (2013) HFT Definition NonColo & NonHFT NonColo & HFT Colo & NonHFT Colo & HFT 53 27 20 16 46.4% 9.4% 23.1% 21.1% 78 31 7 51.4% 15.1% 33.6% Kirilenko et al. (2011) HFT Definition NonColo & NonHFT Colo & NonHFT HFT* *Due to the small number of firms in this HFT category, we are unable to disclose their distribution across NonColo and Colo accounts. This is to comply with the NASDAQ OMX policies on participant confidentiality. Who upgrades? Who upgrades? Probit on all colocated trading entities: 𝑈𝑝𝑔𝑟𝑎𝑑𝑒𝑖 = 𝛼 + θΩ𝑖 + 𝜖𝑖 Explanatory variables measured before the upgrade Upgrade mostly associated with market-making characteristics, not news-trading - More likely to post quotes at the best bid and offer - Higher Order-to-trade ratios - Active trades are uninformed (no additional price impact) - Higher % of trades supply liquidity - Provide liquidity when it is more expensive But it is not a perfect bifurcation - Still use a lot of active trades - Inventory management does not appear different Liquidity effects What happens to liquidity? Depth at Depth at Quoted BBO 0.5% Spread (MSEK) (MSEK) (bps) NASDAQ OMX 0.761 Pre Post 8.980 0.822*** 9.757*** NonColo Realized Effective Spread Spread (bps) (bps) Effective Spread (bps) Price Impact (bps) 4.517 4.206 3.86 0.38 4.267 4.405** 4.126*** 3.86 0.28*** 4.152*** Depth at BBO - the average MSEK volume posted at the BBO Depth at 0.5% - the MSEK trade volume required to change the price at all and by 0.5% Quoted spread –half the difference between the best offer and best bid price scaled by the spread midpoint Effective spread – the difference between the trade price and the spread midpoint prevailing prior to trade NonColo Effective Spread - the same measure conditional on a NonColo trader being involved in the trade Effective spread = 𝑞𝑡 𝑝𝑡 − 𝑚𝑡 = 𝑞𝑡 𝑚𝑡+∆ − 𝑚𝑡 + 𝑞𝑡 𝑝𝑡 − 𝑚𝑡+∆ Price impact Realized spread Liquidity improves in the equity market before and after the upgrade Up Next: What is a good control for time series variation? Control group – OMX 30 futures Depth at Depth at BBO 0.5% (MSEK) (MSEK) OMXS30 index futures 0.034 0.477 Pre 0.039*** 0.492 Post Quoted Spread (bps) Effective Spread (bps) Price Impact (bps) Realized Spread (bps) NonColo Effective Spread (bps) 1.406 1.392 1.461 1.477 1.19 1.04 0.31 0.52 1.461 1.477 Liquidity improves in the futures market before and after the upgrade Up Next: Full difference-difference analysis Liquidity improvement ln(ELiqit) - ln(FLiqit) = a + bPostt + gXit + qi + eit Panel B: Difference-in-Difference Analysis Quoted Spread (bps) Effective Spread (bps) Price Impact (bps) Realized Spread (bps) NonColo Effective Spread (bps) 0.055*** (3.121) -0.007 (-1.596) 10.000 (1.043) -0.017** (-2.396) 0.000 (0.011) 0.319 (0.019) -0.025*** (-6.286) -0.011* (-1.718) 36.584 (1.483) 0.078*** (14.850) -0.011*** (-2.629) 38.087 (1.323) -0.139*** (-9.121) -0.010 (-0.970) 33.073 (1.125) -0.033*** (-20.513) -0.006 (-0.768) 20.629 (0.760) Yes 1200 0.027 Yes 1200 0.007 Yes 1200 0.025 Yes 1200 0.029 Yes 1198 0.091 Yes 1200 0.020 Depth at Depth at BBO 0.5% (MSEK) (MSEK) NASDAQ OMX Post -0.058** (-2.363) Turnover -0.008 (-0.783) Volatility 14.359 (0.618) Stock FEs Yes N 1200 0.021 Adj. R2 Even in the full diff-in-diff specification, liquidity improves Up Next: Is this due to migration of order flow from other exchanges? Liquidity improvement: Consolidated order book ln(ELiqit) - ln(FLiqit) = a + bPostt + gXit + qi + eit Panel B: Difference-in-Difference Analysis Quoted Spread (bps) Effective Spread (bps) Price Impact (bps) Realized Spread (bps) NonColo Effective Spread (bps) - -0.016** (-2.313) -0.002 (-0.281) 6.434 (0.238) -0.033*** (-8.823) -0.009 (-1.238) 37.434 (1.222) 0.091*** (5.323) -0.009* (-1.684) 40.045 (1.180) -0.166*** (-4.223) -0.009 (-0.968) 34.923 (1.089) - - Yes 1200 0.007 Yes 1200 0.034 Yes 1200 0.036 Yes 1200 0.148 - Depth at Depth at BBO 0.5% (MSEK) (MSEK) Consolidated Order Book Post -0.064** (-2.424) Turnover -0.010 (-0.688) Volatility 24.735 (0.555) Stock FEs Yes N 1200 2 0.025 Adj. R Mechanism Inventory Management One channel through which speed may influence liquidity is inventory costs. To better understand the effect of trading speed on inventory management consider how 10GColos change their inventory management behavior after upgrading. Focus on Inventory crosses zero and BBO Presence Inventory Crosses Zero BBO Presence 10GColo Pre Post 13.316 9.807*** 0.129 0.130 SlowColo Pre Post 4.830 4.191*** 0.074 0.070 Inventory held longer by all traders; BBO Presence changes are small Up Next: Full difference-difference analysis Inventory Management Full difference in difference analysis 𝑦𝑖𝑡𝑒 = 𝛽1 𝑃𝑜𝑠𝑡𝑡 + 𝛽2 10𝐺𝐶𝑜𝑙𝑜𝑒 + 𝛽3 𝑃𝑜𝑠𝑡𝑡 10𝐺𝐶𝑜𝑙𝑜𝑒 + 𝛾𝑋𝑖𝑡 + 𝜃𝑖 + 𝜖𝑖𝑡 , Inventory Crosses Zero -0.034 (-0.232) 8.767*** (27.553) -2.978*** (-15.334) -0.728** (-2.472) 0.385*** (6.543) BBO Presence Stock FEs Yes Yes N Adj. R2 23483 0.062 23390 0.062 Post 10GColo Post*10GColo Turnover Volatility -0.004*** (-2.576) 0.052*** (12.975) 0.004*** (3.152) 0.001 (0.523) -0.000*** (-2.939) In the full diff-in-diff specification, 10GColo are more stable market makers Up Next: Does inventory influence liquidity? Inventory and Spreads How is inventory management related to market liquidity? Comerton-Forde et al. (2010) find strong evidence showing a positive link between market-maker inventory and spreads. To show such a link for our dataset and 10GColos we perform an intraday version of their analysis. Inventory and Spreads Aggregate Invt-1 (1) 0.001*** (3.503) High Aggregate Invt-1 (2) -0.193*** (-5.710) 0.194*** (5.727) Mean Abs(Invt-1) (3) (4) 0.247*** (2.591) 16.752 (1.058) 0.049 (1.009) 0.001 (0.166) 17.642 (1.073) 0.050 (1.032) 0.001 (0.161) 17.123 (1.064) 0.049 (1.001) 0.001 (0.167) 0.027 (0.344) 0.166*** (6.046) 17.827 (1.080) 0.049 (1.000) 0.001 (0.163) Stock Fes Yes Yes Yes Yes N Adj. R2 603423 0.243 603423 0.244 603423 0.243 603423 0.244 High Mean Abs(Invt-1) Returnt-1 Turnover Volatility 10G Colos inventory influences spreads, especially when inv. is large Up Next: Emphasize inventory constrained times Inventory Management when Constrained Aït-Sahalia and Saglam (2014): fast market makers submit two-sided quotes when their inventories are within an upper and lower bound. – When inventory is outside the bounds, in contrast, they only submit quotes on the opposite side of their inventory position. A related strategy for inventory-constrained market makers is to post orders asymmetrically around the current midpoint quote, in order to adjust the execution probabilities (known as leaning against the wind). We formulate a test of the asymmetric quoting effect by studying presence at the best bid and offer prices separately and conditional on the inventory of the individual trading entity. Inventory - the number of shares accumulated in that stock-day up to the time of each minute-by-minute randomized snapshot used in the BBO Presence When a trading entity has a long position, a quote at the best bid implies a chance of expanding the position, while a limit order posted at the best offer price represents a chance of reducing the position. How 10GColo liquidity supply depends on inventory 1 minute snapshots: Inventory level and quote presence Leaning against the wind (Menkveld and Hendershott, 2013) Reduce = presence at the best offer (bid) conditional on a long (short) position 0.4 Presence 0.35 0.3 Expand = presence at the best bid (offer) conditional on a long (short) position 0.25 0.2 1 2 3 4 5 6 Inventory deciles Reduce Expand 7 8 9 10 How 10GColo liquidity supply depends on inventory 1 minute snapshots: Inventory level and quote presence Leaning against the wind (Menkveld and Hendershott, 2013) Reduce = presence at the best offer (bid) conditional on a long (short) position 0.4 Before Presence 0.35 0.3 After Expand = presence at the best bid (offer) conditional on a long (short) position 0.25 0.2 1 2 3 4 5 6 7 8 9 10 Inventory deciles Reduce Expand Reduce post Expand Post Inventory Management when Constrained Quote Asymmetry, defined as the difference between Reduce and Expand presence. Focus on 10th decile: close to inventory constraint Quote Asymmetry with constant constraint Quote Asymmetry with changing constraint Inventory Constraint Level 10GColo Pre Post 0.182 0.090*** 0.182 0.117*** 8.837 8.990*** SlowColo Pre Post 0.059 0.039*** 0.059 0.038*** 8.988 9.051 Both types of Colos decrease their asymmetric quoting in the post period Up Next: Full difference-difference analysis Inventory Management when Constrained Full difference in difference analysis 𝑦𝑖𝑡𝑒 = 𝛽1 𝑃𝑜𝑠𝑡𝑡 + 𝛽2 10𝐺𝐶𝑜𝑙𝑜𝑒 + 𝛽3 𝑃𝑜𝑠𝑡𝑡 10𝐺𝐶𝑜𝑙𝑜𝑒 + 𝛾𝑋𝑖𝑡 + 𝜃𝑖 + 𝜖𝑖𝑡 , Post 10GColo Post*10GColo Turnover Volatility Stock FEs N Adj. R2 Quote Asymmetry with constant constraint Quote Asymmetry with changing constraint OLS -0.022*** (-8.435) 0.119*** (6.289) -0.071*** (-5.253) -0.001* (-1.771) 0.000 (-1.020) Yes 9580 0.099 OLS -0.023*** (-11.706) 0.119*** (6.311) -0.042*** (-12.745) -0.001 (-1.212) 0.000 (-0.181) Yes 10062 0.113 WLS -0.015*** (-14.379) 0.058*** (7.584) -0.043*** (-19.961) 0.000 (-0.494) 0.000 (-0.317) Yes 9580 0.040 WLS -0.019*** (-23.585) 0.058*** (7.816) -0.046*** (-36.610) -0.001 (-1.184) 0.000 (-0.001) Yes 8759 0.138 10G Colos asymmetric quoting decreases more after the upgrade Inventory Constraint Level OLS 0.062* (1.829) -0.145*** (-5.818) 0.091*** (3.717) Yes 1468 0.230 Conclusions We provide new insightful summary statistics for colocated firms – Higher order-to-trade ratios – Higher market making presence – Better liquidity timing (better effective spreads) – Better ability to trade on short-lived information The colocation upgrade is associated with Improved market liquidity Overall and for NonColos Is not a shift of liquidity across markets Results suggest the improvement in liquidity is driven by fast traders’ improved inventory management More Summary Stats Who upgrades? Probit on all colocated trading entities: 𝑈𝑝𝑔𝑟𝑎𝑑𝑒𝑖 = 𝛼 + θΩ𝑖 + 𝜖𝑖 Explanatory variables measured before the upgrade Upgrade associated with market-making characteristics, not newstrading Probit (1 = 10G) Number of Active Trades (1000s) Number of Passive Trades (1000s) Liquidity Supply Ratio BBO presence Active Price Impact (bps) Passive Price Impact (bps) Active Effective Spread (bps) Passive Effective Spread (bps) Order-to-trade ratio Inventory Crosses Zero # of trading entities (N) 0.020 -0.046 7.237 16.11 -0.139 1.894 1.266 1.244 0.007 0.074 29 t-stat Marginal Effect (2.23) (-2.95) (2.31) (2.28) (-0.50) (2.09) (1.51) (2.18) (2.37) (0.972) 0.008 -0.018 3.012 6.425 -0.055 0.756 0.505 0.496 0.003 0.029 Information Processing To understand how speed influences adverse selection costs we evaluate how 10GColos react to news We specify a probit regression to investigate whether those who upgrade impose more adverse selection costs on other traders in their active trading or do they use their speed to avoid being picked off in their passive trading (or both). 𝑇𝑟𝑎𝑑𝑒𝜏 = 𝛽1 𝑃𝑜𝑠𝑡𝜏 + 𝛽2 𝑁𝑒𝑤𝑠𝜏 + 𝛽3 𝑃𝑜𝑠𝑡𝜏 𝑁𝑒𝑤𝑠𝜏 + 𝛾𝑋𝜏 + 𝜃𝑖 + 𝜖𝜏 , Trade - 1 if trade τ (with τ=1,…,N) is by a 10GColo entity, and 0 if by a SlowColo 𝑓𝑢𝑡 𝑁𝑒𝑤𝑠𝜏 = 𝑟𝜏 𝐷𝜏 - lagged returns from the index futures market multiplied by the direction of trade indicator D Post - 1 for observations after the event and 0 otherwise. Xτ - Lagged Volatility (the average squared one-second return), - Lagged Volume (expressed in 0.1 MSEK) - Depth at BBO (expressed in 0.1 MSEK) - Quoted Spread (basis points). - Size, the 0.1 MSEK value of the trade Information Processing: Probit Analysis Active Trading Post News News × Post Lagged Volatility Lagged Volume Depth at BBO Quoted Spread Size Stock Fixed Effects N Psuedo R^2 Probit (1 = 10G) -0.074** 212.907*** -85.509 5.894 0.052*** 0.092*** 0.001 -0.543*** Marginal Effects -0.030 84.774 -34.047 2.347 0.021 0.037 0.0005 -0.216 Passive Trading Probit (1 = 10G) 0.0580*** -99.271** -144.599*** 4.048*** -0.005 -0.047*** -0.012*** -0.473** Yes Yes 1,100,026 0.025 1,264,206 0.013 Active trading on news unchanged, Passive trading avoids news trades Up Next: How is inventory management changing? Marginal Effects 0.023 -39.600 -57.682 1.615 -0.002 -0.019 -0.005 -0.189 Decomposing the spread into adverse selection costs and inventory costs • Using the Sadka (2006) price impact model 𝐷𝜏 = direction of trade 𝜀𝛹,𝜏 = unexpected direction of trade 𝐷𝜏 = signed trade volume 𝜀𝜆,𝜏 = unexpected signed trade volume ∆𝑝𝜏 = 𝑧0 𝜀𝛹,𝜏 + 𝑧1 𝜀𝜆,𝜏 + 𝑐0 Δ𝐷𝜏 + 𝑐1 Δ𝐷𝑉𝜏 + 𝑦𝜏 1 𝐸𝑆𝑡 = 𝑇𝑡 𝑇𝑡 𝑗=1 𝐷𝑗 𝑧0 𝜀𝛹,𝑗 + 𝑧1 𝜀𝜆,𝑗 + 𝑐0 𝐷𝑗 + 𝑐1 𝐷𝑉𝑗 /𝑃𝑗 Adverse Selection costs Inventory costs • Applying the Kim & Murphy (2013) trade aggregation approach • Scaling to observed effective spread