Algorithmic trader experiment

The control treatment has 8 human traders per market. The MO-Algo treatment has 7 human traders and 1 market order algorithm. The MO algo uses the current expected asset value and accepts all existing limit orders in the book that imply a profit. The LO-Algo treatment has 7 human traders and 1 limit order algorithm. The LO algo always has an active buy and sell limit order in the book. The price will undercut competing limit orders as long as the undercutting price implies a profit. If there are no competing limit orders, then the LO algo sets a price implying a profit determined by a markup parameter.

2022-03-03

2022-12-31

Main outcome variables on market level:

- Welfare, as measured by sum of profits of all traders
- Price efficiency, as measured by share of transaction prices between the willingnesses to pay of both trader types,
- Liquidity, as measured by a time-weighted bid-ask spread.

On an individual level, the main outcome is the individual trader profit in a round.

More detail in the pre-analysis plan.

- Welfare, excluding the high endowment trader (which is the algorithmic trader in the algo treatments)
- Number of trades
- Number of trades within 1 sec of news release

On an individual level, the secondary outcomes are the number of trades in a round by that trader and the share of trades implying an expected loss given the information at the time of trading by that trader (a measure of traders' mistakes).

See hidden.

In a round, a group of 8 traders receive their endowments in units of an asset and cash in ECUs. They then trade for 100 sec via a double auction. A trader either has a high type, who has an asset valuation of 10 ECUs more than the low type, or the low type. At some time during trading, a public news release either shifts the asset value upwards or downwards.

Each session has 1 practice round which does not count, then 20 rounds of trading that do count. In each round, endowments are reset, and new random variables are drawn (asset value realizations, news time); in fact, these random variables have been pre-drawn in matlab and we use the same sequence of realizations in all sessions and treatments to reduce noise. We pay the payoff of one of these 20 rounds, chosen at random.

7 of the 8 traders receive 4 units of the asset as initial endowments. 1 trader, the high endowment trader, receives 12 units of the asset. 5 traders with an endowment of 4 have the high valuation type, and all others, including the high endowment trader, have the low type. Trader types and endowments are randomly assigned in the practice round, and then constant throughout the experiment. In the trading algorithm treatments, the algo is always the high endowment trader.

Three treatments: Control, MO-Algo, LO-Algo, see above.

Within each treatment, we either start with 10 rounds where the public news release time is known and communicated before trading starts (e.g., news is released 40 sec after trading starts), or with 10 rounds where the public news release time is not precisely known (it is only known that it occurs in the interval 40 sec to 60 sec after trading started). The last 10 rounds will have the other regime. Each order is used for 5 sessions, making 10 sessions per treatment. Hence, we vary whether the precise news time is known within-subject. But the trading algorithms are between-subject.

We use MS excel's random number generator to assign treatments to sessions.

We randomize by trader group/session.

Yes

10 per treatment, 30 overall (we have 3 treatments). A cluster is a group, consisting of 8 traders.

On market level, an observation is one round. 3 treatments, 10 groups per treatment, each with 20 rounds. 3*10*20=600 rounds, 200 rounds per treatment.

10 groups, 10*8=80 subjects in control
10 groups, 10*7=70 subjects in LO-Algo
10 groups, 10*7=70 subjects in MO-Algo

We aim for a power of 0.8 and an alpha of 0.05. We have 20 rounds per cluster (group), so this is 20 observations on market level per cluster. The main outcome variable of interest is welfare, measured as the sum of payoffs of all traders. Relative to the intial endowments, welfare can increase up to 200 ECU or can decrease up to 200 ECU, depending on the number and kinds of trades made. Suppose welfare is uniformly distributed in [0,80], i.e., between 0% and 40% of maximum gains. This implies an SD of about 25. We assume the intracluster correlation to be 0.5. In a previous experiment (Corgnet, DeSantis and Porter (2020)) it was lower than that, sometimes it is higher, so this is our best guess. Under these conditions, how many clusters per treatment would we need if we want to be able to detect a treatment difference of 25? According to Stata, we need 9 clusters per treatment, but we will make it 10 to be sure. The command: power twomeans 0 25, m1(20) m2(20) sd(25) rho(0.5) cluster

We develop a novel experimental paradigm to study the causal impact of trading algorithms on informational efficiency, liquidity, and welfare. In our design, public information about the asset value is revealed during trading, which gives algorithms a reaction speed advantage. We distinguish market-order (aggressive) and limit-order (passive) algorithms, which replace human traders from the baseline markets. Relative to human-only markets, limit-order algorithms can improve welfare, although human traders do not benefit, as the surplus is captured by the algorithms. Market-order algorithms do not significantly change welfare, though they do lower human traders' profits. Both types of algorithms improve price efficiency, lower volatility, and increase the share of profits for unsophisticated human traders. Our results offer unique evidence that non-exploitative algorithms can enhance welfare and be beneficial to unsophisticated traders.

Corgnet, Brice and DeSantis, Mark and Siemroth, Christoph, Algorithmic Trading, Price Efficiency and Welfare: An Experimental Approach (April 14, 2023). Available at SSRN: https://ssrn.com/abstract=4419304.

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4419304