Two doctoral students in Switzerland are running an online test [EN, DE] to see whether people can spot the difference between true stock prices and a random walk.
via Marginal Revolution
Addendum: Both doctoral students are present in the comments section.
via Marginal Revolution
Addendum: Both doctoral students are present in the comments section.
Mahalanobis - am 2005-12-12 15:41 - Rubrik: fun
dsquared (guest) meinte am 12. Dec, 16:27:
I got two-thirds right over fifty tries
On selecting the real from artificial series, that is. I basically used the "trading creates volatility" result from behavioural finance and made the assumption that the real series would show a relationship between price and trading volume that "made sense" but the artificial ones wouldn't. So I looked for a big spike in volume and checked whether it matched up with a large price move or not. Most of the time, one chart would have a big price move on the big trading volume and the other would just have slow sideways action - that one would be the fake data. I would reckon that I could have got it up to maybe 75% if I had full ability to play around with the scale on the volume charts because quite a lot of them had the scale dominated by a single bar.It didn't keep stats on my ability to predict the stock price but looking at it manually, I would be amazed if I did better than chance.
Mahalanobis antwortete am 12. Dec, 23:11:
Rare Events/Jump Diffusion
I identified 18 out of 25 (72 %) artificial price processes correctly (1 - pbinom(18,25,0.5) = 0.0073).Whenever a process looked liked if it followed the SDE
dPt = σdWienert + dJumpt,
I said that it's not artificial. Some of their artificial processes don't look like plain random walks (actually generalized Brownian bridges), but I think my method isn't too bad ;-D.
dsquared (guest) antwortete am 13. Dec, 00:00:
If I was them I would have bootstrapped actual equity returns
In fact, if I was being a real #!@% I would have bootstrapped four- or five-day sequences of returns and then scaled the "down" days to be equal in magnitude to (the "up" days * the total period return on the real chart).I doubt they went that far but I bet the reason it doesn't look quite like a Brownian bridge is that they're drawing from some empirical distribution or other, in which case anyone who is identifying the patterns based on price alone is doing pretty well.
dsquared (guest) antwortete am 13. Dec, 00:02:
PS I'm now up to 80% w/volume
but can't seem to do it at all without! I therefore obviously insist that it isn't cheating at all.
Paul N (guest) meinte am 12. Dec, 19:12:
This is a powerful experiment for people to perform; I've mentioned before how I tried this experiment some years back to convince myself that technical analysis was crap. Like the previous post, I think that using the volume is cheating, I got 3 of 4 right using the volume and I'm pretty confident I could get over half in the long run. I got 2 of 7 right without the volume (!).
Does the number of times you play this game indicate the relative value of your time (i.e. Cowen did it just 3 times), or just how obsessive you are?
HedgeFundGuy antwortete am 12. Dec, 23:08:
I was 2 for 3 w/o volume. I agree that volume is cheating. I figured you can look for fat tails (jumps) and big up-down or down-up patterns that seem to suggest mean reversion but actually are quirky market-on-close prints that are quickly corrected an offer no arbitrage (it's easy to make money on close-to-close prices, too bad you can trade those!).
Daniel (guest) antwortete am 13. Dec, 11:23:
... of course we record whether the trading volume is displayed or not ...In order to construct the random walks, we proceed as follows:
- In the first step, we randmonly draw normal (continuous) stock returns. The mean and standard deviation of the artificial returns correspond to those of the real stock. When estimating the mean and standard deviation, we of course ignore days without trading volume.
- Next, we scale the returns of the random walk series such that the overall one-year return of the computer generated process is equal to the one-year return of the real stock. Standard deviations are not adjusted.
We didn't want to bootstrap returns as the vast majority of theoretical models in finance assume normally distributed returns. Hence, randomly drawing from the normal gives us more degrees of freedom in analyzing the data...
Mahalanobis antwortete am 13. Dec, 16:15:
Thanks for your reply!
Going back to my original question: Don't you think that many participants try to get ≥ 50% right? And that some participants stop playing when, say, guessing correctly 3 times out of 3. Are you aware of any literature on this issue (participants are free to choose the number of rounds)?Btw, I looked at the trading volume most of the time (never throw away information) but its impact on my initial decision was negligible.
Nevertheless, congratulations for this nice site!
dsquared (guest) antwortete am 13. Dec, 16:23:
I bet you are right on this but I don't think it biases the results
I am struggling to remember on this one, but I seem to recall a few examples from sampling textbooks to do with families that keep "trying for a girl" and proofs that they don't affect the overall distribution of sexes in the population. I suspect that the behaviour you describe will also result in people who fail a lot having longer average trials (before they give up in disgust) and that under reasonable assumptions it will average out in the sample. You wouldn't want to use any particular person's record to evaluate their true ability (FWIW I promised myself that I would do fifty trials come what may, although I probably wouldn't have posted a comment here if I'd completely failed) but in the aggregates it should be OK given a large enough sampling size.
Pascal (guest) antwortete am 13. Dec, 16:33:
Nice discussion here
Thanks for all your insights and submitted rounds. where pretty overwhelmed by all the attention it got in the last 36 hours.We both, Daniel and me, don't know of any of this lit but will be very pleased to know more about it. nevertheless, you can be assured that we will address all these questions.
maybe, i'll add a blog or news section to the survey page (or reactivate my own blog for that matter) to keep up with preliminary results. we had reveived quite some positive feedback and interest for this study. especially from alex and tyler at marginal revolution.
thanks again.
tc, pascal
Pavel (guest) meinte am 13. Dec, 13:41:
9 out of 12
I corectly guessed 9 out of 12 pairs of charts (no volume). Perhaps could do better if trained.
