Designing Markets for Monthly Sales
Joey Crampton February 23rd, 2007
I am attempting to design prediction markets based on monthly consumer goods sales. The purpose of the markets would be to predict future month’s sales. Each type of consumer good would have a separate market. The monthly sales of these goods vary greatly, for example one type could sell 50 units a month while another could sell 25,000 per month. Sales increase (could be over 100%, but rarely) and decrease (up to 100%) by large amounts. I am interested in any market design ideas Midas Oracle contributors may have.
Hello, Joey.
First of all, this is an excellent type of market to run. I’ve blogged a bit about it here:
http://blog.mercury-rac.com/20.....n-markets/
A number of markets could potentially work for you. A CDA is certainly the most flexible, but could cause significant issues if the people involved don’t understand trading. (In which case, training could help mitigate that.) Robin Hanson’s MSR (and David Pennock’s DPM) both solve a liquidity issue and are generally easier for novice traders to understand, but aren’t quite as flexible as the prices change based on an algorithm.
Perhaps the biggest issue you need to examine is your organisation. Do you have (or will you have) enough people from different parts of the organisation with different viewpoints and information sources? What will motivate them, and how will the data be used? There is a significant cultural aspect that needs to be considered if you want to make a prediction market successful.
I’m happy to talk to you in detail if you wish on-line or off-line. You can find my contact details at my website (http://www.mercury-rac.com).
Good luck with your project!
Jed, to make a link clickable, leave a space before and after the link.
Like this: ( http://www.mercury-rac.com ).
I will look into Robin Hason’s combinatorial MSR, when I have 5 minutes. I am saying that regarding Jed Christiansen’s comment on “flexibility”.
My flexibility comment is just a simple observation. Let’s say that one month before the next presidential election, the leading candidate keels over from a heart attack.
With a CDA model, prices can made immediate, and massive, step changes.
With an MSR model, prices cannot make immediate step changes, though with enough money and number of people trading, the same price should result.
That’s what I mean about flexibility. Remember that the MSR model was initially (as I understand it) designed so that traders could effectively trade on a huge, interconnecting number of variables. For that purpose, a pricing algorithm was required so that traders didn’t have to price each individual combination of variables separately. It’s still very flexible and responsive, but it just can’t do the step changes that a CDA model can do.
I hope this explains what I meant a little more fully.
Not sure if I would describe that as flexibility, but in such a case, the MSR price would be less correct for however long it took to push the market down, allowing previous longs to get out at better prices. This should still be rather fast, and with an obvious 90% to 0% sort of event, a quick admin could make the market looser, allowing it to act more like a CDA. This sort of practice makes the MSR more opaque though, in the sense that admins could influence prices by changing parameters, which is one of the viable criticisms of HP’s BRAIN.
Not to say that CDAs are necessarily more transparent. Play-money bids and offers could easily be manufactured by an admin “god” account. Perhaps then opacity is not a viable criticism of BRAIN. After all, if one hits oneself in the head with a hammer, we don’t say that it was a bad hammer.. unless the hammer is actually somehow prone towards finding its way to heads.
Thank-you for the welcome and all the comments.
Jed, sorry I forgot to mention this is intended to be a public market.
My concerns are more about how to design the PM with markets for goods that vary so greatly in sales, but all the sales still add up to the whole industry’s sales.
Why would a trader bother to trade in a market where the monthly change in sales could be maybe be 100, when there are markets where the monthly change in sales could be 1000 or 10,000?
A couple of ideas have been tossed around so far:
1. Use percentages for the increase/decrease. The problem here appears to be that the sales could increase by more than 100%, especially when the good has just been released to the public.
2. Use a multiple for the smaller markets where a market that trades in the 100s would be multiplied by 1000, a market that trades in the 1000s would be multiplied by 10. This may be confusing for traders and won’t completely balance out the markets.
Your thoughts on these options or other ideas would be greatly appreciated.
I am guessing this is play-money, in which case you should try to make payoffs more continuous to discourage traders from concentrating on the high %-potential products, which may have a small net effect on the overall industry, and thus not be of especial interest. With real money that problem is diminished. It would also be diminished if you have more lax margin requirements for less volatile markets, as happens with established futures markets. In that case, essentially the volatility of the different markets could be equalized with respect to each trader’s free cash. This would be simple to implement with play-money.
Another option is to have 1) an index market to project overall industry sales, alongside 2) separate binary product markets that represent the % of that product’s sales within the industry. Those binaries should be much less volatile. Even if a product’s sales goes up by 500% in a given period, it may have only gone up 5% relative to total industry sales. This method also encourages traders to buy or sell the full set of products for less or more than 100%.
By the way, another use of internal corporate prediction markets in addition to the ones Jed listed is to test and evaluate employees.
Yes, I intend this market to be play-money.
Not sure I understand how I could make the payoffs more continuous, actually not even sure what exactly you mean by it. Would you mind explaining it a bit further?
Making the margin requirements more lax is an interesting idea, although it may be difficult for new traders to understand. Does anyone know if any other PMs that are currently doing this?
I like the idea of having % of products sales within the industry. It does encourage traders to trade all in all the markets of the industry. Which I think is a good thing. I’m not aware of any PM out there that work there markets this way, are you? Chris do you know of any?
“continuous payoffs”
http://pancrit.blogspot.com/20.....s-and.html
“margin requirements” See TradeSports-InTrade.
“do you know of any?” No.
The two prediction exchanges (betting exchanges) that specialize into predicting revenues is HSX and the SIM exchange. I would study them in depth and spot what designs have been popular with their traders, which is the key.
I guess since things have calmed down for a second I can bring myself to comment on play-money markets…
Although I did suggest an industry index, this is not what I meant by “continuous”, oh Keeper of Keywords. I meant continuous payoffs as opposed to rank-order payoffs. In other words, in play-money markets the goal of traders is often to make it into the top 10 — for example. The other traders are forgotten and get nothing. This encourages risk seeking, as traders buy longshots hoping to “get rich or die trying”. Likewise, why buy a sure thing for 95%, as that 5% gain isn’t going to do much for you in terms of the overall contest? So discontinuous “rank-order” payoffs promote risk-seeking, which promotes the favorite-longshot bias. To be fair, this is somewhat difficult to remedy in play-money markets. One method is to allow users to set-up “leagues” where they can compete against a small group of friends, so that their overall ranking becomes less important to them.
I agree that having different margin requirements for different products would be confusing to most users, which is part of the reason why I favor the idea of maintaining an industry index market alongside separate binary %-of-industry-sales markets (including an “other products” contract). In that case, as with some Intrade accounts, you can sell many different contracts and only have margin frozen corresponding to a single contract, so long as you sell equal quantities of each contract. For example, you want to short-sell products X, Y & Z. X is projected at 5% of industry total sales, Y at 2%, and Z at 1%. Instead of having 95% + 98% + 99% margin frozen, you only need 92% margin frozen, which corresponds to your worse-case scenario of product Z going to 100% and the others going to 0%. While this sounds complicated, I don’t think users will mind since they will always have more margin than they naively thought they would. Most importantly, this will help to mitigate any longshot bias, as traders can efficiently sell all longshots, or sell all products combined, down to 100%.
“This encourages risk seeking” — I agree with you. But, the question is: more than with real-money prediction markets???
“the top 10″ — Exchanges should rank everyone, unless people who don’t want to be ranked.
“leagues” — Rich idea, man. Why don’t they hire you???
For your last-paragraph idea, the would be great if the event derivative traders we know came here and gave us their assessment.
Sure, everyone could be ranked but moving from #792 up to #714 on page 8 of the leader board doesn’t have much utility does it? My “top 10″ comment was a reference to your recent exchange with Eric Z. Front-page visibility (or prizes for the top X traders) is the main payoff, which is one reason these markets encourage more risk-seeking than real-money markets. With real money, each dollar gain has (log) utility. With play money, the utility function is less continuous. See the papers I linked to last week. Also, with play money, if a trader’s account goes to zero from making risky bets, it is probably not difficult to set-up a new account and repeat the process until a longshot hits.
Heh, about “leagues”, that is a standard in the toolbox of play-money prediction market organizers. I heard Dave Perry mention it last summer, and Bo Cowgill talked about it at the Yahoo event in December.
[...] Designing Markets for Monthly Sales - 11 [...]
Hi Joey,
NewsFutures’ Competitive Forecasting model would be a natural for your sales forecasting application. Widely varrying ranges across product would’t matter one bit, nor would lack of liquidity (many products, few traders). Plus it’s very easy to understand and play. Contact me directly at ejss@newsfutures.com if you’re interested in learning more.
And guys, I hate to steel the fire from Perry and Cowgill, but NewsFutures has featured “clubs” where traders compete with their friends since 2000.
[...] Designing Markets for Monthly Sales - 13 comments [...]
What’s the generic term for “Competitive Forecasting”? What’s the category?
And how do Market Scoring Rule compare with “Competitive Forecasting”?
Robin Hanson’s work showed that MSR was better than Scoring Rules.
Competitive Forecasting is quite a bit different than CDA, MSR, or DPM. They really can’t be compared directly.
What is exactly “Competitive Forecasting”, for Christ’s sake? Averaging predictions, scoring rules, what?
From the NewsFutures blog ( http://newsfutures.wordpress.c.....dow-jones/ ) which discusses Competitive Forecasting:
“participants repeatedly enter and review their forecasts as high-low pairs of values and get scored, in the end, on how accurate and precise their forecasts turned out to be.”
and
“They can just enter their forecasts directly with no extra mental gymnastics. It also yields richer prediction data than a mere sequence of prices and trading volumes.”
This can be translated as: scoring rules (but done differently than typically described here since a high-low pair is involved) and a more complete data set (since instead of implying an individual traders’ prediction, it’s explicitly stated).
If NewsFutures’ Competitive Forecasting is Scoring Rule, then I prefer the Market Scoring Rule of Inkling and WSX since Robin Hanson showed in his papers that it’s more efficient.
And I’m still waiting for some prediction market solution provider to show me some combinatorial MSR. Could be interesting to see.
I want to see it at work, as opposed to on the “paper”, literally.
What are hi-low pairs? I joined the Global Risks Prediction Market and I don’t see any pairs. It looks to me like CDA.
There doesn’t seem to be a function to reduced requirements margin for shorting either. From the User Guide “When you go short, you immediately pocket the amount of the sale, but the market also preemptively withholds from your account the maximum potential amount that may be required for the buy-back - in this case, v$100 per contract.”
Is there something I’m missing here?
Joey, the Global Risks site features both CDA trading and Competitive Forecasting (”CF”). The predictions on “Avian Flu vs Dow Jones” are all CF…
Our platform allows for CDA contracts between 0 and any maximum. That enables the market designer to calibrate the cost of shorts, which is basically MaxPayoff - marketPrice. This, however, is different from letting people buy or short “on margin”.
In CF, however, the min and max values for a prediction variable can be anything, even negative numbers, it doesn’t matter. So, for instance, it’s easy to set up a prediction for a “percentage change in sales” say from -30% to +140%. Try that in a prediction market, it gets real messy real fast to have negative share prices…
If you want more details, don’t hesitate to contact me directly: ejss@newsfutures.com
The mechanism design (e.g., market design) should fit both the users (e.g., traders), who want fun, and the manager of the prediction exchange and its corporate sponsor, who expect accurate predictions.
OK, I checked out the “Avian Flu vs Dow Jones” using CF, quite interesting, different from any other market I’ve seen thus far. I like the charts, It’d be interesting to see what a chart looked like with a couple of thousand forecasts.
Is there any way for traders to know how many v$ they are risking and can earn?
So there aren’t opening prices on these markets? Just a closing price at the end of the forecasting period which determines the payouts? How are the payouts determined if the traders aren’t risking anything (I made a forecast and the v$ in my Avian Flu account are still at v$0)?
Structuring your market
http://blog.mercury-rac.com/20.....ur-market/