Order Splitting and Interacting with a Counterparty Vincent van - - PowerPoint PPT Presentation

Order Splitting and Interacting with a Counterparty

Vincent van Kervel Amy Kwan Joakim Westerholm

Pontificia Universidad Cat´

• lica de Chile and University of Sydney

IEX ARC 2020 November 2020

Motivation

◮ Virtually all investors have access to dynamic trading algorithms. ◮ Main focus of this paper: order splitting algorithms, and how they interact.

◮ Split up a large quantity into many small trades. ◮ Often have a long horizon, up to several days. ◮ Typically large institutions.

Motivation

◮ Virtually all investors have access to dynamic trading algorithms. ◮ Main focus of this paper: order splitting algorithms, and how they interact.

◮ Split up a large quantity into many small trades. ◮ Often have a long horizon, up to several days. ◮ Typically large institutions.

◮ Main advantage of order splitting is that it allows liquidity to replenish in between trades. ◮ New insight: Order splitting facilitates interaction with counterparties:

◮ It helps detect presence of counterparties. ◮ It signals your trading interest to the market.

Hypothesis

Can strategic liquidity motivated traders “find” a natural counterparty in the market through order splitting? When a large buyer and seller “find” each other ◮ trading costs are lowest, ... ◮ both can trade very large quantities without moving the price. ◮ Circumvents costly intermediation sector.

Hypothesis

Can strategic liquidity motivated traders “find” a natural counterparty in the market through order splitting? When a large buyer and seller “find” each other ◮ trading costs are lowest, ... ◮ both can trade very large quantities without moving the price. ◮ Circumvents costly intermediation sector. But what happens if the traders all want to trade in the same direction? Generally, how do they interact? What are the optimal strategies when other strategic investors are present? We need a model!

Model (one slide)

Model (one slide)

◮ Study a multi-period model with two strategic long-lived traders with a private valuation (alongside noise traders and a market maker). ◮ They gradually learn about each others private valuations. ◮ Optimal strategy: Each investor trades on own private valuation and on expected valuation of the counterparty. ◮ Over time coordinate better and increase trading aggressiveness.

Model (one slide)

◮ Study a multi-period model with two strategic long-lived traders with a private valuation (alongside noise traders and a market maker). ◮ They gradually learn about each others private valuations. ◮ Optimal strategy: Each investor trades on own private valuation and on expected valuation of the counterparty. ◮ Over time coordinate better and increase trading aggressiveness. ◮ Lemma 1 Endogeneous order sizes. ◮ Lemma 2 Detecting a counterparty through price impacts / reversals ◮ Lemma 3 Counterparties and trading costs.

Data

Goal: study interactions between institutional parent orders ◮ Euroclear Finland provides Nasdaq OMX Nordic trade data for 220,000 unique trader identities for 2007. ◮ Match to public trade and quote data from TRTH to get bid, ask and time stamp. ◮ Data can be obtained at Euroclear-Finland through research proposal + NDA. However, since 2009 some trades are netted per account on a daily basis.

One data point: buyer vs seller (Volume)

One data point: buyer vs seller (Volume)

1 2 3 4 5 6 Volume (millions of Euros) 09:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 17:00 Time Seller Buyer All Transactions

Amer Sports OYJ, AMEAS, October 4, 2007.

One data point: buyer vs seller (Price)

16.2 16.3 16.4 16.5 16.6 Price (EUR) 09:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 17:00 Time Seller Buyer All Transactions

Amer Sports OYJ, AMEAS, October 4, 2007.

Institutional parent orders

◮ Identify 26,226 parent orders using definition of Korajczyk and Murphy (2018) and van Kervel and Menkveld (2019).

◮ Aggregate (child) trades by a single institution for each stock and each day, and apply size and directionality conditions to label it parent orders. Can extend over multiple days.

◮ Typical order is about $1 million, 28% participation rate, lasts 4 hours, and consists of 38 child trades. Uses 40% (active) market

• rders, and has implementation shortfall of only 6bps.

Institutional parent orders

◮ Identify 26,226 parent orders using definition of Korajczyk and Murphy (2018) and van Kervel and Menkveld (2019).

◮ Aggregate (child) trades by a single institution for each stock and each day, and apply size and directionality conditions to label it parent orders. Can extend over multiple days.

◮ Typical order is about $1 million, 28% participation rate, lasts 4 hours, and consists of 38 child trades. Uses 40% (active) market

• rders, and has implementation shortfall of only 6bps.

◮ Parent orders are frequently overlapping:

◮ During the life of an order, on average 8.7% of market volume is traded in the same direction, and 15.6% in opposite direction.

Results 1/2: endogenous order size and trading costs

◮ Simple panel regressions of buy and sell order size on counterparty buy and sell volume by other institutional investors.

Results 1/2: endogenous order size and trading costs

◮ Simple panel regressions of buy and sell order size on counterparty buy and sell volume by other institutional investors. ◮ Contemporaneously, a one-stdev shock to log counterparty sell volume correlates with a 16.9% increase in buy order size and a

• 14.5% decrease in sell order size.

◮ In the half-hour preceding a new order, a one-stdev increase in log-sell volume increases buy-size by 40.8%. ◮ Suggests that investors frequently do not end up trading what they initially wanted to trade (see Perold, 1988)

Results 1/2: endogenous order size and trading costs

◮ Simple panel regressions of buy and sell order size on counterparty buy and sell volume by other institutional investors. ◮ Contemporaneously, a one-stdev shock to log counterparty sell volume correlates with a 16.9% increase in buy order size and a

• 14.5% decrease in sell order size.

◮ In the half-hour preceding a new order, a one-stdev increase in log-sell volume increases buy-size by 40.8%. ◮ Suggests that investors frequently do not end up trading what they initially wanted to trade (see Perold, 1988) ◮ Similar regression of implementation shortfall: a one-stdev increase in counterparty volume in the opposite direction reduces trading costs by 5.6 basis points, relative to a sample average of 6.1 bps Caveat: correlation, no causality.

Results 2/2: Detecting a counterparty

◮ Create a measure of price reversals based on the price impact of child trades (similar to the realized spread)

Results 2/2: Detecting a counterparty

◮ Create a measure of price reversals based on the price impact of child trades (similar to the realized spread) ◮ correlates with trading by other institutions both at the parent-order level and at the half-hour level: detection device ◮ Estimate a VAR model at the half-hour frequency:

◮ Participation rate, net counterparty trading, price changes ◮ an increase in the measure predicts future order flow by other institutions (detection), and ◮ increases the size of the current parent order (response)

◮ The learning and response in trading rate confirm unique model prediction.

Conclusion

◮ Empirically, first to study interactions between institutional parent orders:

◮ Large impact on size and trading costs. ◮ Well-known that market conditions matter, and we show counterparty volume has crucial role

◮ Investors dynamically adjust in real time, consistent with detection and interaction with counterparties. ◮ Confirms the theoretical contribution of gradual learning and coordinated trading. ◮ Useful implications for the design of order execution strategies, exchanges and regulators.