Monday 22 September 2014

Prediction failures

The flat payoff curve problem hit iPredict's vote share markets.

A contract paying $0.01 for every percentage point National got was trading at $0.44 just prior to the election: far below National's outcome of 48%. At the same time, iPredict's contract paying $1 if National were to win more than 44% of the vote was trading well over $0.80. What gives? You can tell convoluted stories trying to rationalise it around improbable underlying probability distributions, but it was just flat payoff curves.

I owned 250 contract of Vote.National. I paid an average of $0.4376 for those contracts. I will gain just under 5 cents per contract, for a gain of $10.60. To get that almost eleven bucks, I had $109.40 locked up in those contracts for a year. Sure, a 10% return seems fine, but it's a pretty heavy carrying cost where the opportunity cost is contracts elsewhere providing greater return (and more risk).

If we look across the "Pays $1 if National gets over x%" contracts, we get a fair odds line of 46.5% prior to the election. A contract paying $1 if National did better than 46.5% was trading around $0.50 in the last couple days before the election. That's still much worse than National's eventual outcome, but not nearly as far out as the 44% used in the Herald's comparison. I'd been warning folks round the office for at least the last month not to trust the continuous contracts because of the flat payoff curve problem and to rely instead on the ranged $1 contracts.

Why were the latter so far out too? That will be a fun one for the iPredict crew to look at. The fair odds line on the Greens was between 13.5 and 14%; the Conservatives were pegged between 4.5 and 5%. I don't know whether people were hedging (then why not short PM.National instead of the vote markets?), or failed to adjust thoroughly for that young Green and Internet/Mana voters might disproportionately fail to turn out.

How did I do?

  • Well over a year ago, I'd shorted ACT's winning an electoral seat. I hadn't cleared that short position after Seymour was nominated because the prices had moved to what seemed then a fair price: I don't trade to reverse past errors, only when prices currently look wrong. So I lost $26 there;
  • Up $171 on PM.2014.National, trading whenever prices seemed to have over-reacted either way;
  • Up $5.61 (so far) on Vote.2014.Nat, having bought at around the $0.42-0.44 range. I bought a pile at $0.4475 back in March when prices seemed offensively out of whack with the $1 contracts, but then gave up trying to bring some sense to that market;
  • Up $5 shorting that Act would win 2 or 3 electoral seats;
  • Up $5 shorting that Russell Norman would be the next finance minister;
  • Up $76 shorting Labour, trading whenever the market seemed to have overshot on either side;
  • Up $45 shorting that the Conservatives would cross 5%.
  • I was up on trading on Internet Party contracts, but I'd closed out all those positions well before the election and so am not entirely sure how it all netted out.
But I lost a couple of side-bets. 

Back in January, I'd reckoned that an Internet Party could cross 5% if there were an Internet Party and if there were a big Snowden release in the middle of the election campaign. I bet a round or two of beers on it with one loyal reader of this blog. I no longer believed the Internet Party likely to cross 5% when Kim DotCom sabotaged the Snowden event by releasing an email that everyone but him believed to be a forgery, but that sure wasn't a condition of the bet. I also owe BK Drinkwater and KiwiPollGuy $20 each. 

No comments:

Post a Comment