Topic: 2014 USD/mBTC Price Prediction Contest - page 18. (Read 21792 times)

If you halve the entry fee (50 mBTC) I will join. You can halve the reward if you want. I want to participate because I like forecasting, so for me it is not about the prize.
I understand that things like this take a lot of timle to calculate, organize, publish, etc.

But I guess with a 100 mBTC you will have less than 10 participants.
(I have also a guessing game running. Not much participants and the entry fee is just 10mBTC)

The current bitcoin price is the wisdom of the crowd. 

There are not so many markets for bitcoin derivatives (tell me if you find any). It is tricky to organize them, the following reasons come immediately to my mind:

- Instead of speculating with going long, you should just buy bitcoins. Much less counterparty risk, guaranteed effect in the economy, like an option that never expires.

- Shorting would be a useful option for many (eg. ones who don't want to trigger capital gains, or not relinquish control of their coins even for the duration of being less than 100% long) but the harder part is to find a matching long. If it is done with actual loaning of coins and selling them to any buyers in the market, it works.

- Betting against the book (MPEx) requires the book have an enormous stash of coins or good risk mgmt (preferably both).


Derivatives would give a lot of important information to the markets though. With this competition I was hoping to construct this information artificially as an aggregate of the guesses of the participants. To make it possible to lower the handling fee, the system should be totally automatized and parametrized. Is there anyone who sees what should be done and knows how to do it?

Exempli gratia:

- web interface with username,passwd login

- prediction tab, which lists each maturity (1,2,.. months) in the columns and each price range (30 in total) in rows

- in each cell you can add probability by "+", reduce by "-" or write a value. Auto-adjustment to total 100.00%.

- API, with which you can auto-update the probabilities.

- Every day the system checks what are the current probabilities and they will be matched against each other at maturity. You don't need to update them every day of course, not even every week or month but only as you want. (In detail: when the month closes and points are awarded, they will be awarded differently from each day, starting perhaps up to a year in advance. So every day you forecast every future month. Every day will count as 1/360 of you score for that month.

- Publishes all aggregate info in a public sheet which both the contestants and market players can read

How much you asking? 

Okay, you want to use the wisdom of the crowd to get a bitcoin price prognosis.  Excellent idea!  You may get a pretty darn accurate prediction, if you get enough contestants.  Unfortunately, I am  not likely to be one of them, because

* It would take me a lot of time to understand how the heck this contest works.
* I'm kind of too cheap to shell out for it.

You might get quite a few more contestants if you dumb it down.  Put up a page with a histogram you can easily manipulate through a GUI to adjust the location of the peak (or peaks), the standard deviation, and how fat the tails are.  Maybe you can do that without sacrificing the underlying sophistication of the math behind your contest.  Lots more work, of course, but it might be worth it, if it garners you hundreds of contestants instead of a few mathletes.

I need to get at least one other player in the first round (apart from sgbett) to make the scoring formula work I am sure that more will follow in the other rounds when they see how it works and hopefully this gains the academic popularity that I am aiming for.

One way to do the classes would be to make them equi-probable:

1259   or more
836 - 1258
677 - 835
600 - 676
531 - 599
490 - 530
450 - 489
411 - 449
358 - 410
357 or less.

It may not be immediately obvious, but according to Bitcoin's historical price developments, it is equally probable that a 30-day period starting at 490 will end up in any of the ranges above (there is an equal number of instances in each of them in the historical data).

Okay that makes sense actually. Good luck, I'm curious to see how it works out

My secret aim is not to make profit (for either me or you) but to get intelligent and statistics/probability/math/Bitcoin-oriented people to enter thought-out predictions so that the aggregated data of the predictions could be used in making informed trades by both me and all who read this. This would go out to the world as an indication that Bitcoin folks really can forecast its going up (so that it is not random shouting "CCMF!12") and also will be really interesting how accurate the next bubble top forecasting and the following decline will be.

The rationale behind a high fee is to reduce the number of entrants, increase the quality of forecasting, make administration easier and enable me to offer a competitive prize.

Simplifying the method can be done without affecting the essentials. 7 categories would be a bit gamey, I'd like 20 more. The historical range (alone) for Bitcoin is -65%...+765% per month and needs to be adequately captured with the choices.

I think it would be better if this were made much simpler and the fee to enter were either nothing or something very small. Definitely not 0.1btc, that's a lot. That means all you need is 21 people to enter for you to make a profit.

Or what about like a bitcoin lottery or something.  You are well trusted in the bitcoin community and if like 10-50 of us would send like .01-.1 to a address you control, at the end of the month there could be a random selected winner (to which you could take a small % for your hosting).
Would be fun.

I think you need to make promotion contest (no entry fee)  for few month and only then put on serious one.

Otherwise system doesn't look intuitive and "easy to get on" - as it now, you will get only "math" guys to participate. Indeed administrative with current system would be quite a nightmare, so you need to make some restrictions as well (so you an just copy-paste results into one big spreadsheet to get all results).

For example, restrict price ranges to 7 (not less and not more) but still use that geometric average thing.


At this moment audience is not ready - only elitists will enter.
Need more attention for regular Joes like myself!

There is a possibility to make the scoring simpler:

- When it is known which slot wins, the contestants' probabilities for that slot are summed up. Everybody gets points according to what is their "share" of the correct guesses (perhaps scaled to the average of 1).

In the example far above, the players A-J get points in proportion to their guess (5% gets 10 times the points of 0.5%, and nobody can get negative points).

The system here described is actually just using simple average instead of geometric average in scoring.


In my opinion this has however many downsides:

- Makes the system much more gameable, because anticipation of other players betting heavily a certain range makes you hesitant to bet it, regardless of your estimation of its probability.
- Dull and not enough action, because not possible to get negative points.
- Prone to lucky number strategy, because unbalanced guesses not enough penalized.
- Not elegant to handle cases where people skip rounds etc. (should they get an average score for missed rounds??)

interesting game. I am not sure yet. 100 mBTC seems to be above my budget.

i think i see now.

the width represents the fact that you have bet [P/Slot] on every price within the range so then [Prob] = [P/Slot] * [width]

clear as a bell!

RE: dates....

For this month I must wait til 28th when you'll give us the top bracket, then PM you predictions by 23:59 GMT on 30th? and these predictions are for VWAP on 31st May.

I better fire me up a copy of excel!

Yes. You will always hit at least 0.0001%, because that is the rule. (Stemming from the requirement to have all positive numbers for calculating the geometric average.)

In your example your P(slot) is actually 0.1000% instead of 1%. You would do pretty horribly if the price does not do anything special and falls in the range where people have forecast ~2% per slot. In the test round, the geometric average of 16 predictions in the $0.60/mBTC area was about 1.4%.


Yes. Like I tried to explain, P/slot is what counts, but some might prefer to think of it in larger blocks. I don't care, it is not really that much harder to make a true distribution of 491 slots, but for those who don't use Excel, the option to compress it might come in handy.


It is exactly this kind of questions that we need to make everything obvious for the lurkers also, before the contest starts. But I take refuge in the fact that it is possible to enter the competition after one (or even several) rounds have passed, or re-enter with a new horse if you fail a prediction with the old one

btw I'm totally entering thats why i want to figure out how best to win hehe

I am not *sure* I can game the system. I'm just saying that people will try - Its human nature so I was pointing out what I thought was a fairly obvious way of scoring points.

I see now I missed the width multiplier.

if I rephrase like your example...


then i'm guaranteed to hit, its now just a question of what i score.

Would I be right to assume that its the "P/Slot" figure that is used to award points rather than the  "Prob." figure?

eg assuming geomettric avg was say 2%

I get 1% / 2% - 1 = 0.5 pts

rather than 50% / 2% - 1 = 24 pts

perhaps this was just obvious to everyone except me

You are required to assign a probability to each of the values "slots" from 0 to 491 (cents per mBTC). The probability must be at least 0.0001%.

For your (and mine) convenience, it is possible to group the slots that in your opinion are equally probable, like I did in the example in the OP. If you give eg. 10% probability to the range 0.60-0.69 (which consists of 10 slots), the probability per slot is 10% / 10 = 1%.

You are free to give different probabilities for each of the 491 slots but many will find it better to group them. The grouping does not affect winning chances because the total percentage is divided equally to all the width of the group before applying it in the calculation.


Like explained above, in the "atomic level" your task is to allocate one million "probability units" to (approximately) 491 slots, so that the slot where the "ball" (Bitcoin price after 30 days from the prediction) falls, contains more of your "probability units" than the other players'.

There is only the minimum of 1 PU (=0.0001%) per slot. There is no maximum, so if in your opinion it is wise to put 999510 PU's in the slot labeled as 65 (and leave all the others with the minimum) you are free to do it.

Let's calculate it out: assume that the slot is a very high probability one, and there are 20 forecasters in the game, and the others have given it a geometric average of 4% probability. You go for 99.95%. If you win, you get 20.2734 points. But the probability of winning (wisdom of crowds) was only 4%, which is 1/25. You would need at least 25 contestants to have a +EV.

If you're out of luck and any of the other 490 slots is realized, such as one with the average 1% probability, your guess of 0.0001% would net you -6308.5374 points.

* *

The grouping of slots is a user interface convenience, which you are free to use or not.

The system punishes reckless behavior, which is giving way too little probability for an event which actually has a chance to happen (only if it happens, of course). Insuring against these is easy, just give a little more weight to them.

The system rewards lucky guesses (especially on events that were not deemed probable by the others (see "reckless behavior" above)), but their EV (expected value = long term success) is still negative. Making a lucky number in a tight band (even a single slot) however does not mean that you need to sacrifice your whole prediction, even a few % will win big if the others

If others introduce volatility to the guesses, making a conservative range of estimates yourself can actually give you a positive outcome over the whole range. In the trial round in my thread, my personal estimate will net a positive score if anything below $3000/BTC happens in 2014-5-17. This is because many of the others were playing bull and bear. Whichever happens, the correct forecasters will win big and the others lose big, but I will make a few points regardless. After 8 rounds the hotheads have hopefully made a few bad guesses negating the good ones, and I will win

You are sure that you can "game the system", but I am equally sure that the Prediction Contest has very intricate tactics in many levels of the rules, which make sure that the best forecaster will win! After all, points are awarded relative to others

so does "any number of classes" mean i can have 2 guesses?

e.g.

over (10 x price) : 50%
0 - (10 x price) : 50%

I'm sure i'm not the only person that will enter the competition looking to win by gaming the scoring system

Perhaps you need some way of punishing overly broad ranges (or rewarding tight ranges)?

or perhaps simpler, a maximum range size, or standardised ranges?

Despite your efforts, it's still not very clear for (at least) me.

What if persons used 30 different values for prediction and I used 3?


Bother to give few examples on real values?
like, what if my predictions are:

380 - 33%
440 - 27%
560 - 40%

As I understand, geometric mean would be (0,33*0,27*0,4)^(1/3) = 0.329
And lets say actual price was 486

So what?

Ideas:

- Perhaps also the 2nd and 3rd place should receive some of the prize pot, like in poker tournaments.

- Disqualifying entrants whose point total reaches -100 points

- Bonus for keeping in the black = having a positive score each round