Prediction market error bars

I just started reading this fascinating paper about the U. of Iowa prediction markets. Here is their approach to determining market accuracy:

1) Observe: an efficient market w/ a risk-free rate of 0 moves as a random walk w/ mean 0

2) Demonstrate empirically that the Iowa prediction market indeed moves as a random walk

3) Analyze the microstructure of the walk, ie what is the distribution of returns?

4) Using (3), you can predict error bounds based on time just by analyzing the walk (do a Monte Carlo simulation or whatever).

This lets you make statements like "At X weeks out, the distribution of final prices is Y". You can then link this w/ statements about the accuracy of closing prices to give overall error bounds for the predictions.

There is one big assumption which is sort of a weak link, namely that future walks will look like past walks. You can certainly imagine that different markets might have different levels of volatility. But you can try to account for that by a) modeling the random walk as a function of these covariates, and b) disclaiming that your bounds are only accurate if the future looks like the past, which is an inherent weakness of any forecasting method.

Share this

- stochastic volatility and

- stochastic volatility and fat tails are a feature of any financial market (October 1987 was a 27 standard deviation event), so any prediction of error bounds will be basically worthless.

Another thought is to the extent that the prediction markets are viable, liquid markets, there will be participants taking positions for hedging purposes so that one might have to take into account other markets as well. For example, if the energy markets were concerned over event X in, say Iran, one might use the prediction market to hedge the event risk against a position in energy or airline stocks.