Revision to "algorithmic trader experiment"

The treatments are:
- control (all traders are human subjects)
- market order algo treatment: the medium type trader is played by a market order algorithm (an algo that only uses market orders)
- limit order algo treatment: the medium type trader is played by a limit order algorithm (an algo that only uses limit orders)

These treatments are between subject. In the algo treatments we will vary within subject a greediness parameter of the algos. For the market order algo, the greediness can be 0 or 5, and for the limit order algo, the greediness can be 5 or 10. In these treatments, 10 rounds are then high greediness and 10 rounds are low greediness (with order changed between sessions).

2025-04-09

2025-10-31

Welfare, price efficiency, liquidity

Outcomes on market level:

1. Welfare, measured as the sum of welfare gains,
2. Price efficiency, measured as the share of transaction prices between the asset
valuations of the low type and the high type traders, where the valuation prior to news release
is the expected asset value for the type,
3. Liquidity, measured as the inverse of the time-weighted bid ask spread

4. Number of trades,

5. Number of trades immediately (within 1 sec) after news release.

6. Price volatility, as the variance of the prices

7. Non-Intermediary Welfare: Welfare excluding the medium value trader, measured as sum of welfare gains of all players except the medium value trader

See hidden design.

We will run a market experiment in the lab. A session consists of 11 traders, who will play one practice round (not counting) and then 20 rounds for data collection (which will be incentivized). 5 traders have a low private value (0) for the asset. 5 traders have a high private value for the asset (20). And 1 trader has a medium private value for the asset (10). Traders then exchange units of the asset via double auction for 100 seconds each round.

After 50 seconds of trading, the common value part realization of the asset value is publicly announced ("news release"). The common value can either be 20 or 80, drawn with equal probability.

Traders are initially endowed with 400 ECU cash and 4 units of the asset. The payoff of an asset at the end of the round is the sum of the common value and the private value. Thus, some traders value the asset more than others.

The experiment allows for shortselling, up to 10 shortsales per market/round. And traders can use an interest free loan of 400 ECU during trading.

Ztree determines the random variables, with the exception of the asset common value realizations, which have been drawn via matlab random number generators beforehand, so the same sequence of common value realizations can be used in all sessions.

We cluster at the group level (11 traders).

Yes

9 per treatment, 27 overall (we have 3 treatments). A cluster is a group, consisting of 11 traders.

On market level, an observation is one round. 3 treatments, 9 groups per treatment, each with 20 rounds. 3*9*20=540 rounds, 180 rounds per treatment.

9 groups in control
9 groups in market order algo treatment
9 groups in limit order algo treatment

We use the same power calculation as for the original study: According to Stata, we need 9 clusters per treatment. The command: power twomeans 0 25, m1(20) m2(20) sd(25) rho(0.5) cluster