How To Set Up a Prediction Market: HP Example
Here’s the closest thing I’ve found to an explanation of how to set up and conduct a prediction market. This paper, INFORMATION AGGREGATION MECHANISMS: CONCEPT, DESIGN AND IMPLEMENTATION FOR A SALES FORECASTING PROBLEM, by Charles R. Plott of CalTech and Kay-Yut Chen of Hewlett Packard Laboratories, describes how they set up a prediction market for sales forecasts at HP with the following results:
· In 6 out of 8 events for which official forecasts were available the market predictions were closer to the actual outcome than the official forecast.
· The probability distributions calculated from market prices were consistent with actual outcomes.
· The market made accurate qualitative predictions about the direction that the actual outcome will occur (above or below) relative to the official forecast.
I’m separating my notes into two posts. First, the nuts and bolts about how it was done and second, some of the scientific issues.
How it was Done
Typically, the prediction was for monthly sales for a month three months in the future.
· Hesitation to engage employees in an exercise in which they might lose money. Solution: provide a small amount of cash to each participant before the market sessions – this constrains the amount of stakes a participant can have in the market and affects incentives to trade.
· Market has to offer useful information. E.g., if forecasts are not valuable if they are made with horizons less than 3 months then market sessions need to be conducted 3 months before the event to be predicted.
· Relatively small number of participants chosen. Selected specifically from different parts of the business operation because they were thought to have different patterns of information about the targeted event. These patterns of information, including market intelligence, specific information about big clients, and pricing strategies, were in need of aggregation. No public summaries of information available to the participants during the operation of the IAM. The official forecasts were not known until after the IAM closed.
· Participants need to be selected carefully – don’t want to “miss” a person with much information but it might not be efficient to include many people without any relevant information. Little is known theoretically about the information size relative to the market that might be required for effective information aggregation to take place.
· Laboratory experiments have suggested that a small number of uninformed participants provide both market liquidity and a function of adding “consistency” to the market through a process of “reading” and “interpreting” the actions of others. So, around five subjects was recruited from HP Labs (with little or no information) in each experiment.
Preparing the participants
15-20 minute instruction session from: explained the structure of incentives, the market mechanism and the web interface. In addition, the participants were told the goals of the experiment and were told that their participation was important for HP business. Contact information provided and participants were encouraged to call if they encountered difficulties.
Defining the Contracts to be Traded
Most similar to the IEM “winner take all” markets (state contingent securities).Traded a complete set of state contingent contracts (Arrow-Debreu securities). The space of possible outcomes was partitioned into about 10 intervals. Each interval was given a name and with each interval there was an associated security with the same name that traded in a market with that name. Thus the interval 0-100 would be associated with a security named 0-100 that traded in a market named 0-100. The interval 101-200 would be associated with a security named 101-200, etc.
If the final outcome fell in an interval, the corresponding security would pay, say, one dollar per share at the end of the experiment. All other securities would pay nothing. A higher payoff per share would place more value on the share but the payoff per share interacts with the total cost of the exercise and the potential volume of trades and related market liquidity.
How to Start
Each participant is given a portfolio of shares in markets and cash to start. Could start with equal shares in all securities or could start with shares in every other security, alternating which security was first across participants. The unequal distribution of endowments was used to encourage trading by attempting to make sure that the initial endowments of securities did not approximate the ultimate equilibrium.
Web based, double auction markets. Marketscape software (Laboratory of Economics and Political Science at Caltech.)
· All the markets for an event were organized on a single web page for easy access.
· Links to a complete time series of trades available.
· Links available to HP data bases, which allowed participants to review data held by HP.
· A participant could enter a buy offer, a sell offer or acceptance of an offer through the web form on the page. Orders were compared to the other side immediately. If a trade was possible, it was executed and if not the order was placed in an order book. The best offers were listed on the main market web page. The whole book of offers was available for each market at the click of a button.
· Participation was anonymous. However, each participant was assigned a subject ID number for each experiment. During the experiment, the subject ID number of the person who made offers and transactions were public knowledge. Participants had the ability to track behavior of other subjects with in the same experiment if they wished to.
When should market be open?
· In all cases the information was gathered for a week with the markets being open during lunch and in the evening every day. Management did not want participants being preoccupied with the task during the working day when pressing issues needed attention.
· It is desirable to have a schedule (for example, 24 hours for a week) to minimize conflict with other activities.
· It is not desirable to leave market open for long periods because participants will often find a lack of activity in the market and thus lose interest.