iPredict accuracy blitzes traditional polls
Monday 28 November 2011
FOR IMMEDIATE RELEASE
iPredict accuracy blitzes traditional polls
The Chief Executive of iPredict, Matt Burgess, has declared the performance of Victoria University’s online predictions market in forecasting the election result to have been “absolutely outstanding” and said refinements would be made to assure even greater accurate in 2014.
Mr Burgess was commenting on iPredict’s initial post-election analysis which shows iPredict’s trading prices were closer to the final result than 41 of the 51 traditional opinion polls published in the eight months before election day.
The analysis also shows that, in the month before the election, iPredict forecast each of the National, Labour and Green party votes to within 1.5% of the final result. During that time, the average iPredict forecast of National’s party vote was 47.10% (0.89% less than won by National on Saturday), Labour 28.60% (1.47% more than the actual result) and 11.1% for the Greens (just 0.48% more than its 10.62% result).
Even more significantly, over the three months before election day, iPredict forecast National’s party vote to within an amazing 0.01% of its actual result, while overestimating Labour’s vote by just 1.67% and underestimating the Greens by just 0.42%.
iPredict also accurately forecast the
results in at least 66 of the country’s 70 electorates
(with Christchurch Central still too close to
There are now nearly 6000 traders registered on iPredict and anyone else can sign up for free at www.ipredict.co.nz and begin trading the more than 500 contracts on offer.
“iPredict’s absolutely outstanding forecasting accuracy for the three major parties has stunned even us, although it is very much consistent with the international literature on prediction markets around the world. As our analysis continues, we will be passing our learnings on to our colleagues running similar markets around the world, including at the University of Iowa where the world’s first predictions market was launched for the US election in 1988.”
Mr Burgess also noted that the market had appeared to slightly underestimate the centre-right National Party’s vote share and slightly overestimate the centre-left Labour and Green parties vote shares.
“This is an important result to us, because it lays to rest concerns that, because predictions markets involve money, they may have a bias towards the centre-right. That appears not to be the case.”
iPredict used a bundle of indexed contracts to forecast the party vote shares of the political parties. This means that, for each share they hold in a party’s vote, traders will receive that amount in cents. That is, holders of National’s party vote shares will receive 47.99c per share.
The contracts were “bundled” in that the sum of the value of payouts will equal $1, as should the value of all the contracts at any given time.
Mr Burgess said iPredict’s initial analysis suggested that indexed contracts appear to be more accurate outside the extremes of prices, as indicated by the extraordinary accuracy of the National, Labour and Green party votes.
For parties with forecast votes below 10%, the picture was less clear as a result of the cost of buying a contract (at less than 10c) being significantly less than the cost of shorting the same contract (at more than 90c).
For learning purposes, adjustments to liquidity and other steps had been taken at various times to gain greater understanding of the extent of that apparent bias and these would be considered by the PhD candidate and other academics analysing the market as part of their research programmes.
“We are satisfied with our performance for more of the small parties, given that we were within 0.5% of the result for the Conservative, Maori and UnitedFuture parties for the month before the election. We aren’t quite as satisfied with our record for the Act, Mana and New Zealand First parties where we were as much as 2.11% out from the actual result, however our initial analysis still indicates we outperformed the traditional polls overall, even for these parties.”
Mr Burgess said iPredict had learned from the issues about low-value indexed stocks, as had been highlighted during the election period by Dr Eric Crampton, senior lecturer in economics at the University of Canterbury.
As a result of the learnings and Dr Crampton’s analysis, a bundle of indexed stocks for the 2014 election had been launched today. This bundle includes just four options – National, Labour, Green and “all others”. As had been done for the Conservative and Mana parties in 2011, additional stocks would be launched on smaller parties’ vote shares, adjusted by a factor of 10 to address the apparent issues for parties with lower levels of support.
The full table of iPredict’s performance
over the last year is as
Party Election result iPredict averages by time period before election iPredict forecast compared with election result
1 month 3 mths 6 mths 1 year 1 month 3 mths 6 mths 1 year
National 47.99% 47.10% 48.00% 47.00% 47.00% -0.89% 0.01% -0.99% -0.99%
Labour 27.13% 28.60% 28.80% 29.40% 29.90% 1.47% 1.67% 2.27% 2.77%
Greens 10.62% 11.10% 10.20% 9.60% 9.00% 0.48% -0.42% -1.02% -1.62%
NZ First 6.81% 4.70% 4.70% 4.70% 4.80% -2.11% -2.11% -2.11% -2.01%
Act 1.07% 3.00% 3.50% 3.70% 4.10% 1.93% 2.43% 2.63% 3.03%
Mana Party 1.00% 2.00% 1.90% 1.70% 1.60% 1.00% 0.90% 0.70% 0.60%
Conservative 2.76% 250% 2.30% 2.10% 2.10% -0.26% -0.46% -0.66% -0.66%
Maori Party 1.35% 1.20% 1.30% 1.50% 1.80% -0.15% -0.05% 0.15% 0.45%
UnitedFuture 0.61% 0.90% 1.10% 1.10% 1.20% 0.29% 0.49% 0.49% 0.59%