Risk Markets And Politics

Sunday, November 26, 2006

Discounting and prediction market prices

The prolific Alex Forshaw asks why 2008.GOP.NOM.HUCKABEE trades above 7. But should you sell even an impossible event for less than 100-(100/((1+(r-rm))^T)+ c? Here, T is the years to expiration, r the risk-free rate, rm the yield on your margin (which is zero for most Intrade traders), and c is the trading fee, which is currently less than or equal to 0.5 + 1/(1+r)^T (a 0.5% price-taker's entry fee plus the present value of the 1% in-the-money fee). The price of 2008.GOP.NOM.HUCKABEE seems more reasonable with this is mind, but traders are only interested in selling the whole set of possibilities for more than 100*(1+(r-rm))^T + Σc, or buying them for less than 100/(1+(r-rm))^T - Σc. This will mitigate any discounting "bias". (Play-money markets also have a discount rate, but it is logically and empirically much less. Play-money markets also show some evidence of yield curve inversion.)

Discounting returns may not bias prices for a number of reasons, including de-biasing by large interest-earning accounts, a short average trade horizon, and gamblers' "irrationality". Insofar as rates (and commissions) are found to affect prediction market prices, then perhaps an "implied probability" or "delta" metric to supplement price is inevitable. In that case, if we interpret price as probability, then the price will be biased, but only because of our interpretation.

Sunday, November 12, 2006

Political Factor Analysis

David Pennock, Lance Fortnow and others on Marginal Revolution and DailyKos explain the "failure" of prediction markets for Senate control as stemming from the assumption that state-level elections are independent events, an assumption that ignores national movements in sentiment.

Is there a way to test this idea? If a general anti-Republican/pro-Democrat wave swept across the country, decreasing the independence of individual elections, it might be expected that those Republicans who best represented the party line would fare the worst compared to what prediction markets and polls projected.

Voting records such as those available at the Washington Post can serve as a simple proxy for how representative members are of their given party. A legislator will be "typical" if their votes nearly always correspond to their party line. Now do the electoral fortunes of legislators who are more typical in this sense in fact exhibit a higher "beta" to general sentiment shifts? If (possibly after trimming special cases) this did not appear to be the case in the last election — if prices were not worse in proportion to how typical the Republican candidate was, then the lack of independence explanation is suspect. In that case, no explanation should be required either, since after all, 70% != 100%.

This is only a first pass, as an issue-level factor analysis should be more enlightening than one that remains at the party-level. This would also allow us to take the challengers' views into account when modeling individual races.

In terms of the "defense" of prediction markets, apparently it needs to be pointed out ad nauseum that the question is not about their absolute success, but their success relative to alternate tools. Empirical findings confirm the superiority of prediction markets to polls, and the logic of motivation is clearly in their favor. In terms of anecdotal evidence, consider these opinions submitted in response to a television survey on confidence in polls conducted on 11/5/06:
I tell the pollsters the exact opposite of what I'm planning to do. It's none of their business anyway.
Carole, Columbus, OH

I don't believe in polls. I keep a penny by the phone. 'Heads I'm a Democrat, tails I'm a Republican.' I like to keep them guessing.
Tom, Seattle, WA
Maybe polls should be more emotive and oblique, and ask questions like, "What issues anger you the most?" and "What issues most affect your day-to-day life?". Prediction markets would then be able to digest this data and better gauge the independence of elections.

In any case, both political factor data and analysis techniques will see development before the 2008 elections. On the latter side, this will include working through ambiguities and accounting for correlations in issue preferences (which constrain party platform "optimization").

[Cross-posted to Midas Oracle]