The profit calculations are fixed!

I did some greedy search for a combination that had the most suspicious 4-tuple counts:

• make 40 random choices and assume those ones are human made

• count how often each possible 4-tuple appears in the possibly human made strings in total

• calculate the entropy of this distribution

• flip 1-10 of the random choices

• check if the new assumption has a lower entropy

• keep the changes if yes

• continue search for 10000 steps

• store final combination and repeat everything 1000 times with new initial guess

• find 1000 more combinations each that locally minimize different entropies, like entropy of 3-tuple counts

• average the x*1000 combinations to get a price for each pair

This performed okayish, had 24 of 40 correct. But on some of them it was pretty sure so it made 1400M profit from 3500M investment.

The strategy I used:

• Measure how well combinations compress when all chunks put into one string

• For each pair, check which string appears more often in combinations that compress best

• Bet that that string is human generated

This worked very well at first, but then I took a couple of big loses, and overall didn't make much profit. Still not sure whether this strategy was any good to begin with.

Edit: Maybe I did make a lot of profit actually. Hard to tell due to bugs.

@roma "didn't make much profit" 😁

@nanob0nus I did't know 😅

I got really into this at the beginning, and then it just took too much of my time, so I checked out.

Retrospective on the market:

- Wow, this was heavy. My next markets will be substantially lighter than this and take a shorter time. I think even this market would have been better if it was a lot faster to maintain interest and excitement.

- I think the market was way underconfident towards the end. There was a comment that got 18 of the last 20 predictions right - I would have thought that after getting 9/10 correct of markets 21-30, traders would have caught on and said "hey, those predictions are actually quite good, let's use them", but this didn't really happen.

- If I think about the success of predictors based on that comment, I think they performed quite well actually. Like, getting 18/20 right towards the end is not bad!

- Interestingly, the predictions Just Before Resolving Anything were (very slightly) worse than trivial! Could it be that the predictors fell prey to the same biases that I did? (My memory from looking at the comments at the time suggested this, but haven't thought about this much)

- Not sure if the missing trader bonuses affected anything.

---

I have the data of predictions Just Before Resolving The Nth Market for each N (except for N = 2, for which I forgot to do this). Raw data here. Note: I copied them manually, so there might be typos. You should be suspicious if the prediction of a given market changes a lot (>10%) after a single resolution.

I haven't done really any data analysis on them yet. If you have ideas, implement them or send them my way!

See the below comment for the data

47 51 48 43 56 50 43 48 62 54 50 49 60 53 44 43 53 50 57 46 51 55 55 61 61 56 40 42 45 55 46 41 54 43 56 49 49 55 58 58

?

51 46 54 51 41 48 60 55 49 49 60 53 44 49 55 50 57 46 50 55 56 62 61 56 40 42 45 55 50 43 51 43 55 49 49 47 52 55

38 61 45 40 47 61 55 44 46 60 55 44 50 57 50 55 46 48 55 56 68 63 56 43 42 45 55 46 43 54 42 55 49 49 47 52 58

59 44 38 47 66 56 43 46 64 55 44 39 58 50 58 46 46 55 59 68 63 58 39 39 48 55 55 46 49 36 52 49 46 42 52 59

35 38 47 68 52 54 46 64 55 44 39 58 50 58 46 46 55 59 68 63 58 39 39 48 55 55 46 49 36 52 49 46 42 52 59

27 49 69 60 59 46 64 55 44 39 60 50 58 46 46 62 59 68 63 58 39 39 48 62 55 46 49 36 52 49 46 42 52 59

51 65 60 59 46 64 55 44 39 60 50 58 46 46 62 59 68 63 58 39 39 48 62 55 46 49 38 47 49 46 42 52 59

60 60 59 46 64 55 44 36 60 50 58 46 41 62 59 68 63 58 39 39 48 62 55 46 49 38 47 49 46 42 52 59

51 60 46 64 55 44 36 60 50 58 46 41 62 59 68 63 58 39 39 48 62 57 48 47 36 47 47 46 41 52 61

63 46 57 55 44 36 60 50 58 46 41 62 59 68 63 58 39 39 48 62 57 48 47 36 47 47 46 41 52 61

45 55 55 44 36 60 50 58 46 41 62 59 50 63 58 40 39 48 62 57 48 47 38 47 47 48 43 52 61

64 57 44 36 52 50 58 46 41 62 59 50 63 58 40 39 48 62 57 48 47 38 47 47 48 43 52 61

57 43 36 47 57 60 46 41 65 59 59 62 58 40 39 48 64 57 48 47 38 47 47 45 43 65 57

45 35 47 60 60 46 40 65 59 59 62 58 40 37 48 64 57 48 47 38 47 47 45 43 65 57

31 47 64 63 46 40 65 59 52 57 56 39 37 48 57 60 58 47 36 48 51 45 43 49 59

44 64 63 49 40 65 59 48 57 52 39 37 49 57 60 48 47 36 48 51 45 43 49 59

64 68 50 36 65 59 48 52 49 39 37 49 57 60 48 47 36 48 51 45 43 49 59

70 47 36 65 59 48 49 49 39 37 49 57 60 48 47 36 48 51 45 37 49 59

47 36 65 59 48 49 49 39 37 49 57 60 48 47 36 48 51 45 37 49 59

36 68 61 46 47 47 42 35 49 69 60 48 43 36 45 55 45 37 58 59

68 61 46 47 47 42 35 49 69 64 48 43 36 45 58 41 37 58 59

64 48 47 47 42 35 49 69 64 48 43 36 45 63 41 37 58 59

40 47 47 42 35 49 69 64 48 43 36 45 63 41 37 63 60

45 45 42 36 49 69 67 48 43 36 45 66 37 37 63 60

49 41 35 56 69 67 51 38 35 40 66 37 37 63 63

52 35 56 69 67 51 47 35 40 66 37 37 63 63

35 62 69 67 51 47 35 40 66 37 37 63 63

62 69 67 51 47 35 40 66 37 37 63 63

60 70 51 47 32 40 66 37 37 57 63

68 57 49 34 41 59 38 38 57 61

57 49 34 41 59 38 38 57 61

44 34 41 61 38 36 57 63

34 41 61 38 36 63 63

60 59 38 36 62 63

59 42 40 72 63

42 44 72 63

43 69 63

60 63

64

@Loppukilpailija Thanks a lot for the market!

The amount of data here was a bit overwhelming and that made the market less catchy. For example I realized quite late that @GGtr s guess is so good, since I had to check that manually after each new outcome. Also manifolds overview features on these mass markets are still not ideal, e.g. the total profit count is total bs.

The daily resolutions stopped me from getting overconfident, since we all experienced some blowbacks. @capybara on the very first one, me on 7,10,11,17 and @roma on 33.

@nanob0nus

e.g. the total profit count is total bs.

I'm actually not sure how much I've made or lost 😅

Feels like I've made 500-1000.

Manifold shows just below the answers: "You made Ṁ155 in profit!".

Manifold doesn't show me in "Top traders" although other numbers there are way below 155.

"Top traders" definitely wrong. The 155 number might be right, but I'm not sure.

@roma

https://discord.com/channels/915138780216823849/1173315755517939862/1178401097887527032

someone somewhere maybe is aware of this and hopefully tries to fix this soon. Or it is just fixed for new markets and this one stays defect

@nanob0nus The total profit calculation seems to assume that all the markets are still at the final market price before resolution. And the profit leaderboard just takes this wrong value from the last resolved market.

@nanob0nus Crazy how broken it is! Hope they'll fix old markets too

Final update and concluding thoughts tomorrow!

(Market was closed for a moment, sorry about that. Added a few days more - this will end during this week anyways.)

My previous prediction set 01100110111100010011 did fairly well (got 9/10 right so far), let's stick to it and predict 1100010011 for the last 10. Too bad there are some uncertain ones there.

@GGtr so close

39 is the first wrong one

10 markets left! The derivative market has something about whether the market is right about the rest of the pairs and whether someone correctly guesses the rest of the pairs. I'll leave you a day to make predictions.

Let's have a day-long break now that we are half-way through. Next update tomorrow.

My predictiones for the second half: 01100110111100010011 (not all of them very confident)

This is based on:

At this point we can see that OP is trying too hard to switch exactly half the time: all the artificial strings switch between 11 and 13 times, which is far too neat - random ones switch between 9 and 17 times. So betting on the string that switches closer to 12 times works pretty well.

We can see that chunks of 4 are distributed unevenly: there's too few 1111 and 0000 and too many 0110. Can assign scores based on that.

@GGtr This was a great prediction, congratulations! (Strange that the market didn't reflect the quality of this one.)

@Loppukilpailija My guesses after 18 reveals:

1000101110111000011000101000001000010011
1000101110111000011000101000011000010011
1000101110111000011001000000000000010011
1000101110111000011001000000001000010001
1000101110111000011001000000001000010011
1000101110111000011001000000001010001011
1000101110111000011001000000001010010001
1000101110111000011001000000001010010011
1000101110111000011001000000001110010001
1000101110111000011001000000011010010011
1000101110111000011001000000011010010111
1000101110111000011001000000111010010011
1000101110111000011001000010111010010001
1000101110111000011001000010111010010010
1000101110111000011001001000001000010010
1000101110111000011001001000001000010011
1000101110111000011001001000001100010001
1000101110111000011001001000001100010011
1000101110111000011001001011111000010010
1000101110111000011001010000111010000011
1000101110111000011001010010111010000001
1000101110111000011001011010111010000010
1000101110111000011001011010111010000011
1000101110111000011001011010111010001010
1000101110111000011001100000001000010001
1000101110111000011001100000001000010011
1000101110111000011001100000001010010011
1000101110111000011001100000001111110001
1000101110111000011001100000011010010011
1000101110111000011001100010011010010001
1000101110111000011001100010011010010011
1000101110111000011001101000001000010011
1000101110111000011001111010111010000001
1000101110111000011001111010111010000010
1000101110111000011011010000011110110011
1000101110111000011101001010101000010010
1000101110111000011101100000001010010011
1000101110111000011101100000001110010001
1000101110111000011101100000011010010011
1000101110111000011101100010011010010001
1000101110111000011101101000001000010011

Tried to use AI for text completions and based on how many tokens were predicted I'd determine which is which. I didn't have much luck, mainly because it never actually responded with the right format. I tried most of the OpenAI models.

Good god I'm losing so much money

Would be interesting to calculate log score of the final predictions.

@Irigi I will do that - in fact, there's a market about whether the final predictions are better than trivial on the derivative market: https://manifold.markets/Loppukilpailija/what-will-happen-on-a-market-on-dis

EDIT: I'm also interested in seeing how the quality of predictions for [markets that have not been resolved yet] moves over time. I haven't yet done any plots or anything, but I have taken screenshots of the market probabilities each time I resolve one.