Market economics can be applied to solving many problems in computer science, especially regarding resource allocation. E.g. A Multi-Agent System for Controlling Building Environments

An allocation is Pareto Efficient if no one can be made better off without making someone worse off.

The utility or margin in a trade is the difference between the price paid and the price the buyer/seller was willing to pay.

Supply and demand curves change constantly in a Continuous Double Auction as trades are made and traders leave the market.

Vernon Smith pioneered experimental economics.

Metrics for markets

Smith’s Alpha

Root mean square deviation of transaction prices around the theoretical equilibrium price, as a percentage (lower is better).

Allocative Efficiency

Total utility earned by all traders, divided by the theoretical maximum possible utility (surplus), expressed as a percentage. Measures how effective the market is at extracting gains. Can never be over 100%

Single Agent Efficiency

Profit earned by an agent, divided by its expected profit if all trades took place at equilibrium price. Measures how well an individual agent performs. Can be over 100%.

---

Gode and Sunder found that most of the intelligence is in the market itself, and not the traders.

Cards

Q: What does an allocation being Pareto Efficient mean? A: If no one can be made better off without making someone else worse off.

Q: What is the utility or margin in a trade? A: The difference between the price paid and the price the buyer/seller was willing to pay.

Q: What are the three main metrics for markets? A: - Smith’s Alpha

• Allocative Efficiency
• Single Agent Efficiency

Q: What is Smith’s Alpha? A: Root mean square deviation of transaction prices around the theoretical equilibrium price, as a percentage (lower is better). Quantifies volatility.

Q: What is Allocative Efficiency? A: Total utility earned by all trades, divided by the theoretical maximum possible utility (surplus), expressed as a percentage.

Q: What is Single Agent Efficiency? A: Profit earned by an agent, divided by its expected profit if all trades took place at equilibrium price.

Q: What did Gode and Sunder find about market efficiency? A: Most of the intelligence is in the market itself, not the traders.

Q: What is a ZIU trading bot? A: Zero Intelligence Unconstrained. Just generated random bid/offer prices. Useless.

Q: What is a ZIC trading bot? A: Zero Intelligence Constrained. Generates random bid/offer prices, but won’t trade at a loss. Surprisingly human, but there are conditions where they won’t equilibrate.

Q: What is a ZIP trading bot? A: Zero Intelligence Plus. Have a profit margin they won’t trade below, which they adjust using a learning rule. Can succeed in markets where ZICs fail. If trades are taking place below current price, they’ll reduce their margin, and the inverse if it’s above. The amount by which the margin is adjusted is determined by a learning rule.

Q: What is the learning rule for ZIP? A: Widrow-Hoff with momentum. It aims slightly above the value it’s adjusting up to, or slightly below the value it’s adjusting down to. It does this through both a percentage and an absolute, which are randomly generated for each function call.

Q: What is Kaplan’s Sniper Trader? A: It does nothing until it sees that the bid-offer spread is sufficiently small, or the offer is less than the smallest transaction price in the previous period, or the market closes soon. Then it jumps in and “steals the deal”. Only works if other traders aren’t using a sniper. They don’t adapt to market activity. They don’t know the equilibrium price, so will snipe any deal.

Q: What is a Gjerstad-Dickhaut (GD) trader? A: Computes a belief function using recent market activity to estimate probability of a bid or offer being accepted.

Q: What is the equation for a GD trader’s belief function? A: $f(p)=\frac{able(p)+ole(p)}{able(p) + ole(p) + rbge(p)}$ where $able(p)$ is the number of accepted bids $\le p$ , $ole(p)$ is number of offers made $\le p$, $rbge(p)$ is number of rejected bids $\ge p$.

Q: What is the difference between GD and MGD (Modified GD)? A: GD interpolates with a cubic spline for prices that don’t exist in the history list. MGD shows a probability of zero for prices above or below historical maximum/minimum.

Q: What is the Adaptive-Aggressive (AA) trading algorithm? A: Uses a moving average to estimate equilibrium price, and Smith’s alpha to quantify volatility. Makes bid/offer based on these plus current aggressiveness. A more aggressive agent places a bid/offer more likely to be accepted. Aggressiveness updates similar to margin of ZIP. Underlying aggressiveness function also considers market volatility, in a more volatile market a small change in aggressiveness leads to a large change in behaviour.