PredictIT

Forecasting Swing States with Economic Data

James Bailey1 Comment

Ray Fair at Yale runs one of the oldest models to use economic data to predict US election results. It predicts vote shares for President and the US House as a function of real GDP growth during the election year, inflation over the incumbent president’s term, and the number of quarters with rapid real GDP growth (over 3.2%) during the president’s term.

Currently his model predicts a 49.28 Democratic share of the two-party vote for President, and a 47.26 Democratic share for the House. This will change once Q3 GDP results are released on October 30th, probably with a slight bump for the dems since Q3 GDP growth is predicted to be 2.5%, but these should be close to the final prediction. Will it be correct?

Probably not; it has been directionally wrong several times, most recently over-estimating Trump’s vote share by 3.4% in 2020. But is there a better economic model? Perhaps we should consider other economic variables (Nate Silver had a good piece on this back in 2011), or weight these variables differently. Its hard to say given the small sample of US national elections we have to work with and the potential for over-fitting models.

But one obvious improvement to me is to change what we are trying to estimate. Presidential elections in the US aren’t determined by the national vote share, but by the electoral college. Why not model the vote share in swing states instead?

Doing this well would make for a good political science or economics paper. I’m not going to do a full workup just for a blog post, but I will note that the Bureau of Economic Analysis just released the last state GDP numbers that they will prior to the election:

Mostly this strikes me as a good map for Harris, with every swing state except Nevada seeing GDP growth above the national average of 3.0%. Of course, this is just the most recent quarter; older data matters too. Here’s real GDP growth over the past year (not per capita, since that is harder to get, though it likely matters more):

RegionReal GDP Growth Q2 2023 – Q2 2024
US3.0%
Arizona2.6%
Georgia3.5%
Michigan2.0%
Nevada3.4%
North Carolina4.4%
Pennsylvania2.5%
Wisconsin3.3%

Still a better map for Harris, though closer this time, with 4 of 7 swing states showing growth above the national average. I say this assuming as Fair does that the candidate from the incumbent President’s party is the one that will get the credit/blame for economic conditions. But for states I think it is an open question to what extent people assign credit/blame to the incumbent Governor’s party as opposed to the President. Georgia and Nevada currently have Republican governors.

Overall I see this as one more set of indicators that showing an election that is very close, but slightly favoring Harris. Just like prediction markets (Harris currently at a 50% chance on Polymarket, 55% on PredictIt) and forecasts based mainly on polls (Nate Silver at 55%, Split Ticket at 56%, The Economist / Andrew Gelman at 60%). Some of these forecasts also include national economic data:

Gelman suggests that the economy won’t matter much this time:

We found that these economic metrics only seemed to affect voter behaviour when incumbents were running for re-election, suggesting that term-limited presidents do not bequeath their economic legacies to their parties’ heirs apparent. Moreover, the magnitude of this effect has shrunk in recent years because the electorate has become more polarised, meaning that there are fewer “swing voters” whose decisions are influenced by economic conditions.

But while the economy is only one factor, I do think it still matters, and that forecasters have been underrating state economic data, especially given that in two of the last 6 Presidential elections the electoral college winner lost the national popular vote. I look forward to seeing more serious research on this topic.

Prediction Markets As Investments

James Bailey4 Comments

Supporters of prediction markets tend to emphasize how they are great tools for aggregating information to produce accurate forecasts. If you want to know e.g. who is likely to win the next election, you can watch every poll and listen to pundits for hours, or you can take ten seconds to check the odds. This is great for people who want information- but how do prediction markets fare as investments for their actual participants?

Zero Sum

The big problem with prediction markets as investments is that they are zero sum (or negative sum once fees are factored in). You can’t make money except by taking it from the person on the other side of the bet. This is different from stocks and bonds, where you can win just by buying and holding a diversified portfolio. Buy a bunch of random stocks, and on average you will earn about 7% per year. Buy into a bunch of random prediction markets, and on average you will earn 0% at best (less if there are fees or slippage).

Low Liquidity

Current Kalshi order book for “Will June 2024 be the hottest June ever“. Betting $200 on either outcome could move the price by 5 cents (so move the estimated probability by 5pp).

This zero sum problem is close to inevitable based on how prediction markets work. They currently have one other big problem, though it is not inevitable, and is getting better as they grow: liquidity. There are some stocks and bonds where big institutions can buy or sell millions of dollars worth without moving the price. But in markets like Kalshi or PredictIt, I personally move prices often by betting just hundreds, or sometimes even just tens, of dollars. Buying at scale means getting worse prices, if you can even buy at all. PredictIt has a bet limit of $850 per contract for regulatory reasons. This definitely excludes institutional investors, but even for individuals it can mean many markets aren’t worthwhile. Say an outcome is already priced at 90 cents, the most you can make by betting it happens is about $94. That’s not nothing but its also not enough to incentivize lots of in-depth research, especially given the risk of losing the $850 if you are wrong and the opportunity cost of investing the money in stocks or bonds. Kalshi in theory allows bets up to $25k, but most of their markets haven’t had the liquidity to absorb a bet anywhere near that (though this could be changing).

Easy Alpha

Given these negatives, why would anyone want to participate in prediction markets, except to gamble or to generously donate their time to create information for everyone else? Probably because they think they can beat the market. Compared to the stock market, this is a fairly realistic goal. Perhaps because the low liquidity keeps out institutional investors, it isn’t that hard for a smart and informed investor to find mispricings or even pure arbitrages in prediction markets. This seems to be especially true with political prediction markets, where people often make bets because they personally like or dislike a candidate, rather than based on their actual chances of winning; that is exactly the kind of counterparty I want to be trading with.

I’ve been on PredictIt since 2018 and earned a 16% total return after fees; this was on hundreds of separate trades so I think it is mostly skill, not luck. Of course, even with this alpha, 16% total (not annual) return over 6 years is not great compared to stocks. On the other hand, I tended to put money in right before big elections and take it out after, so the money is mostly not tied up in PredictIt the whole time; the actual IRR is significantly better, though harder to calculate. On the other other hand, the actual dollar amount I made is probably not great compared to the time I put in. On yet another hand, the time isn’t a big deal if you are already following the subject (e.g the election) anyway.

Uncorrelated Alpha

The other big positive about prediction markets is that there is no reason to expect your returns there are correlated with your returns in traditional markets. Institutional investors are often looking for investments that can do well when stocks are down, and are willing to sacrifice some expected returns to get it. In fact, there may be ways to get a negative correlation between your prediction market returns and your other returns, hedging by betting on outcomes that would otherwise harm you. For instance, you can hedge against inflation by betting it will rise, or hedge against a recession by betting one happens. If you are right, you make some money by winning the bet; if you are wrong, you lose money on the bet but your other investments are probably doing well in the low-inflation no-recession environment.

Going Forward

Prediction markets have long been in a regulatory grey area in the US, but with the emergence of Kalshi and the current CFTC, everything may soon be black and white. Kalshi has won full approval from the CFTC for a variety of markets, but the CFTC is moving to completely ban betting on elections (you can comment on their proposal here until July 9th).

One great place to discuss the future of prediction markets will be Manifest, a conference hosted by play-money market Manifold in Berkeley, CA June 7-9th. It features the founders of most major US predictions markets and many of the best writers on prediction markets. I’ll be there, and as I write tickets are still available.

Two Types of News: Elections vs Crashes

James Bailey3 Comments

Some events are like elections: it was obvious that some big political news would break on Election Day, we just had to wait to find out what exactly would happen. Others are like market crashes: you might know in principle they’re a thing that can happen, but you don’t really expect any particular day to be the day one happens, so they seem to come out of the blue. As it turns out, for one of the largest crypto exchanges the day of the crash also happened to be Election Day.

FTX.com is facing a bank run sparked by competitor Binance tanking the price of the token that backed some of their assets. Customers are having issues withdrawing their money, Binance has withdrawn its offer to bail out FTX by taking them over, and bankruptcy seems likely. Supposedly this doesn’t affect Americans using FTX US, but I’d be nervous about any funds I had there, or indeed with funds in any centralized crypto exchange or stablecoin (Tether and even USDC seem to be having issues holding their pegs). All this was especially shocking because many considered FTX founder Sam Bankman-Fried one of the most trustworthy people in the often sketchy world of crypto. He was always meeting with US regulators and lawmakers, and seems not to be motivated by greed; he had already begun to give away his fortune at scale.

After any surprising event like this, some people claim it was actually obvious and they saw it coming (despite usually never having said so beforehand), while others start looking back for warning signs they missed. The most interesting one is something that shocked me when I first heard it March, but I never considered the risk it implied for FTX until the crash:

Going forward, red flags to watch out for seem to be topping a list of youngest billionaires (as Elizabeth Holmes also did) and buying naming rights to a stadium.

In contrast to this crash, the election happened right when we all expected, and at least largely how I expected. Like markets, I underestimated Democrats a bit; polls overall were impressively accurate this year, though they of course missed on some particular races. Votes are still being counted, and as of now we don’t even know for sure which party will control Congress (PredictIt currently gives Democrats a 90% chance in the Senate and a 20% chance in the House). But here are some early attempts to assess forecast accuracy. As I said, some polls were quite good:

Some polls weren’t so good, which means its important to weight better pollsters more heavily when you aggregate them. Some attempts at that were also quite good:

Oddly, some no money (Metaculus) / play money (Manifold Markets) forecasting sites seem to have done better than the real-money prediction sites:

A Dragonfly’s View of Election Day 2022

James Bailey3 Comments

This is my last post before the US midterm elections on Tuesday, so I’ll leave you with a prediction for what’s coming.

Who is the best predictor of elections? Nate Silver at FiveThirtyEight has had a pretty good run since 2008 using weighted polls. Ray Fair, an economics professor at Yale has a venerable and well-credentialed model based on fundamentals. I typically favor prediction markets, because they incorporate a wide range of views weighted by how willing people are to put their money where their mouth is, and traders are able to incorporate other sources of information (including predictors like FiveThirtyEight). But which prediction market should we trust? There are now many large prediction markets, and the odds often differ substantially between them.

When there are many reasonable ways of answering a question or looking at a problem, it can be hard to choose which is best. Often the best answer is not to choose- instead, take all the reasonable answers and average them. Dan Gardner and Philip Tetlock call this approach Dragonfly Eye forecasting, since dragonfly’s eyes see through many lenses. So what does the dragonfly see here?

Lets start with the US House, since everyone covers it.

  • FiveThirtyEight’s latest forecast shows that Republicans have an 85% chance of taking the House; it shows a range of possible outcomes, but on average predicts that Republicans win the popular vote by 4.3% and take 231 House seats (substantially over the 218 needed for a majority)
  • The Fair Model predicts that Democrats will win 46.6% of the two-party vote share (leaving Republicans with 53.4%). This has Republicans winning the popular vote by 6.8%, a moderately bigger margin than FiveThirtyEight. The reasoning is interesting; the economy is roughly neutral since “the negative inflation effect almost exactly offsets the positive output effect”, so this is mainly from the typical negative effect of having an incumbent party in the White House.
  • Prediction markets: PredictIt currently gives Republicans a 90% chance to take the House. Polymarket gives them 87%. Insight Prediction also gives them 87%. Kalshi doesn’t have a standard market on this, but their contest (free to enter, 100k prize) predicts 232 Republican seats.

Its a bit tricky to average all these since they don’t all report on the same outcome in the same way. But the overall picture is clear: Republicans are likely to do well in the House, with an ~87% chance to win a majority, expected to win the popular vote by ~5.55% and take ~232 seats.

The Senate is closer to a coin flip and harder to evaluate.

  • FiveThirtyEight gives Republicans a 53% chance to win a majority (51+ seats for them; Democrats effectively win if the Senate stays 50-50 since a Democratic Vice President breaks ties for at least 2 more years). The most likely seat counts are 50-50 or 51-49, but confidence intervals are pretty wide and 54-46 either direction isn’t ruled out.
  • The Fair Model doesn’t make Senate predictions, only House and Presidential predictions.
  • Prediction markets: PredictIt gives Republicans a 70% chance to win a Senate majority, probably with 52-54 seats. PolyMarket gives Republicans a 65% chance, as does Insight Prediction. Kalshi predicts 53 Republican seats.

Overall we see a much higher variance of predictions in the Senate; a 17pp gap between the highest (70%) and lowest (53%) estimates of Republican chances, vs just a 5pp gap for the House (90% to 85%). This shows up with the seat counts too; everyone agrees there’s a substantial chance Republicans lose the Senate, but if they do win, it will probably be by more than one seat. The average estimate is ~52 Republican seats. FiveThirtyEight and PredictIt agree that the closest Senate races will be Georgia, Pennsylvania, Arizona, Nevada, and New Hampshire (though they rank order them differently), so those are the races to watch.

Forecasts for governors aren’t as comprehensive, but FiveThirtyEight predicts we’ll get about 28 Republican (22 Democratic) governors, while PredictIt expects 31+ Republicans; I’ll split the difference at 30. Everyone agrees that Oregon is surprisingly competitive because of an independent drawing Democratic votes. The biggest difference I see is on New York, where PredictIt gives Republican challenger Lee Zeldin a real chance (26%) but FiveThirtyEight doesn’t (3%).

Overall forecast: moderate red wave, Republicans take the House and most governorships, probably the Senate too. But if they lose anything it is almost certainly the Senate.

These forecasts seem about right to me. Democrats are weighed down by an unpopular (-11) President and the highest inflation in 40 years. This would lead to a huge red wave, but Republicans have their own weaknesses; an unpopular former President lurking in the background, and the Supreme Court making a big unpopular change voters blame them for. This shrinks the red wave, but I don’t think its enough to eliminate it. The effect of Roe repeal is fading with time, and the unpopular Biden is more salient than the unpopular Trump; Biden is the one in office and is more prominent in media coverage. Facebook and recently-acquired Twitter may be doing Republicans a favor by keeping Trump banned through Election day. But if he drags Republicans down anywhere, it will be the Senate, where candidate quality (not just party affiliation) is crucial and his endorsements pushed some weak/weird/extreme candidates through primaries. We’ll also see this “extremist” Trump effect (abetted by cynical Democratic donations to extreme-right candidates) dragging down Republicans in some key governor’s races like Pennsylvania, where Democrats are now 90/10 favorites..

CFTC Orders PredictIt Shut Down- Can Political Betting Survive?

James Bailey8 Comments

Political betting has long been in a legal grey area. It seems that the Commodities Futures Trading Commission wants to make everything black and white, but at least for now it has simply made everything murkier.

PredictIt is the largest political betting site in the US; if you want to know who is likely to win an upcoming election, its the best place to find a quick answer. Prediction markets have two great virtues- they are usually right about what’s going to happen, and if they aren’t you can bet, making money and improving their accuracy at the same time.

PredictIt has operated since 2014 under a “no-action letter” from the CFTC. Effectively, the regulators told them “we’re not saying what you’re doing is definitely legal, but we know about it and have no plans to shut you down as long as you stick to the limits described in this letter”. But last week the CFTC withdrew their letter and ordered PredictIt to shut down by February 2023.

My first question was, why? Why shut them down now after 8 years when all their operations seem to be working as usual? The CFTC said only that “DMO has determined that Victoria University has not operated its market in compliance with the terms of the letter and as a result has withdrawn it”, but did not specify which of the terms PredictIt violated, leaving us to speculate. Did the scale simply get too big? Did they advertise too heavily? Did Victoria University, the official operator, let too much be handled by a for-profit subcontractor? Did some of their markets stray too far from the “binary option contracts concerning political election outcomes and economic indicators” they were authorized for?

PredictIt hasn’t been much clearer about what happened, simply putting a notice on their site. Their CEO did an interview on the Star Spangled Gamblers podcast where he said there was no one thing that triggered the CFTC but did mention “scope” as a concern- which I interpret to mean that they offered some types of markets the CFTC didn’t like, perhaps markets like “how many times will Donald Trump tweet this month”.

The other big question here is about PredictIt’s competitors. In 2021 it seemed like we were entering a golden age of real-money prediction markets, with crypto-based PolyMarket and economics-focused Kalshi joining PredictIt. I looked forward to seeing this competition play out in the marketplace, but it now seems like we’re headed toward a Kalshi-only monopoly where they win not by offering the product users like best, but by having the best relationship with regulators. Polymarket had offered markets without even a no-action letter, based on the crypto ethos of “better to ask forgiveness than permission”; this January the CFTC hit them with a $1.5 million fine and ordered them to stop serving US customers.

If the CFTC doesn’t reverse their decision to shut down PredictIt, then February 2023 will see a Kalshi monopoly. This has led to speculation that Kalshi is behind the attack on PredictIt; their cofounder issued this not-quite-a-denial. But it certainly looks bad for the CFTC that they are effectively giving a monopoly to the company that hires the most ex-CFTC members.

For now you can still bet on PredictIt or Kalshi (or even Polymarket if you’re outside the US). If you’d like to petition the CFTC about PredictIt you can do so here. It might actually work; while the CFTC’s recent actions certainly look cronyistic, they’ve been reasonable compared to other regulators. They’re giving PredictIt no fines and several months to wind down, and even Polymarket gets to keep serving non-US customers from US soil. I’d likely make different decisions if I were at CFTC but the ideal solution here is a change in the law itself, as we’ve seen recently in sports betting. Prediction markets are impressive generators and aggregators of information, and politics and policy are at least as valuable an application as sports. To go meta, suppose we want to know- will PredictIt survive past February? There’s a prediction market for that, and its currently saying they’ve got a 20% chance.