dtrader98, I skimmed through this paper quickly. Isn't the conclusion apparent to any layman using a DOM screen? Of what practical use was it to you? Are you designing execution algs? It seems that this paper was more about the appropriateness of the statistical models used to measure liquidity affects. Interesting from an academic perspective or maybe for those writing execution algs... -Lou
Not at all. In fact, it is counter to what most perceive to be primary drivers of large volatility (not so much volume of large incoming orders being executed, but more so, the profile of the order book distribution). My personal experience is that unless you have access to the actual MM order flow (which I definitely do not), the DOM at high frequencies tends to be a distorted and obscured version of actual DOM. The paper is useful in the sense that it gives an idea of how to answer some questions posed earlier regarding the relationship between tick rate and actual liquidity. I am assuming you have defined tick rate as the frequency of occurrence of ticks/unit time here (with no regard to directional magnitude, nor temporal frequency). In this case, as indirectly related to the paper's argument, the magnitude of tick rate would not necessarily be proportional to available liquidity. I.e., liquidity is more a function of true DOM profile then measuring the magnitude of tick rate. Regarding trading algorithms, while I do design and utilize them, I do not focus so much on very high frequency execution as it tends to be cost prohibitive with respect to constant negative commission bias as well as slippage. However, I always find discussions of underlying market minutia useful, since as is below is often as is above.
I find this discussion of tick-rate per time unit as a measure of liquidity quite interesting. In my personal experience, 'time' is a confounding variable. In fact, I do everything I can to remove it. I find that prices move relative to information -- whose closest proxy we have is either volume, or better, significant (and I leave the definition of that word up to the readers as an exercise) price movement. While a liquidity measurement may serve as a good initial screen as to what securities are 'tradable,' do people find that it helps define anything else? I would speculate that a security with a lower liquidity level more likely to trend. I could see this being similar to my empirical studies that stocks with low float are more apt to trend -- because it takes less shares changed to create a significant movement. In companies with large floats, it takes an enormous amount of buying pressure to begin and continue a trend. Perhaps securities with low tick-rates would be similar -- a rapid increase in tick rate may cause a flurry of short-term buying... I am babbling...
dtrader98, I understand your point now. If you're using tick rate as a proxy for liquidity you may have a misspecification error since DOM dynamics are more closely related to liquidity. Personally, I look at tick data and rate from an information theoretic perspective. Thanks for your insight. -Lou
I personally do not know what is the exact mechanism to determine ticks and how they are generated. (Actually I would like to know these details.) In any case, from a trading perspective is important only to determine if that kind of information can improve a trading strategy. So in any case experiments are needed. Having observed empirically that clearly there is no proportionality between tick rate and volumes of liquidity (across instruments), one hypothesis which comes to my mind is that a large component of that signal is generated electronically, in a way that is probably "related" to the ongoing "activity" for the given security, but where the range of tickrate (min-max) used, or other distributional characteristics, are probably specific of the instrument or security (or of the algorithm to make the market). If the information is useful to determine increase or decrease of activity, and therefore for trading purposes, it must be "rescaled" appropriately for each security. And this might pose the interesting problem of the suitable indicator to detect those (relative) accelerations/decelerations. (just an hypothesis) Tommaso
Blulou, I've looked at some of your prior posts/ideas with interest. I like some of your the ways you are approaching markets. A few questions, particularly regarding hurst type of analysis. I've often looked at hurst as no better than simple visual TA (which in my opinion, is not very good at all). However, you seem to embrace it. 1) Have you actually backtested your hurst trigger based entry/exits over many sets of out of sample data? If so, what kind of success rate do you see (simple average percent winners/hit rate is fine). To make the experiment more dependent on hurst; how about simple enter on >some value like enter long h>.7 exit long h<.6? 2) What window period do you use for each log(r/s) vs. log(n) data pt? If fixed, why do you choose that length? If adaptive, how do you scale? Also, how many pts. do you use to calculate your hurst slope? I ask because, in my brief attempts to look at this, it seems to me there is a fleeting dependency on the slope as new information comes in. The estimate is also subjective, as there is no optimal length of window or pts accepted in practice. And each guess will give a different answer. It would be useful to look at a scatterplot of hurst par n periods forward vs hurst par at n. I might try to run this, but I am expecting not to find much linearity nor any type of simple fit. Could you show a plot of this over a few hundred to thousand trials of data? I've also seen arguments where hurst exp is related to slow decay power law. If this is true, I would expect any price series that has greater than hurst exponent than .6 or so to be able to simply be modeled by AR process (there is direct equation relating them as well). However, if you ever try to look at correlogram of simple price changes, it almost NEVER fits a slowly decaying exponential process (unless you are looking at squared returns., i.e. garch). Correlogram of prices fits quite well, but the noise terms renders it relatively useless. Anyways, curious to hear your thoughts on this. Tks. P.S. I was a die hard random walker, but after many years of looking at this, I have made some personal discoveries that put some kinks in that argument.
dtrader98, Actually, I think you're correct for the most part regarding the limitations of using the Hurst exponent. For small N, I've found that if you smooth a VHF indicator it's monotonic w/Hurst. Except Hurst is easier to interpret, i.e., random vs. non-random. Technically, I'm not using a true Hurst exponent - maybe the R/S method is closer to that. The method I'm using comes from the following paper, http://www.complexity.org.au/ci/vol05/sevcik/sevcik.html This approach converges to Hurst in the limit. I can't recall if R/S is the same. I wouldn't begin to defend the use of Hurst insofar as all of the theoretical powers that have been ascribed to it. I use Hurst as part of a filter bank to screen out tick data series for irregularity and to extract information. My filter bank uses 3 other algorithms + Hurst. So, Hurst by itself is not sufficient. I find that for N >= 512 that it's pretty good at sorting out what looks like random data to the eye. I also use it for smaller N, but in that case, I don't think it's any better than something like a smoothed VHF or ADX. I can''t say what my hit rate would be with Hurst alone. I just know that it contributes at the margin. I've never looked at the filter bank stats in aggregate but I think a hit rate of 58% +/- 3% is in the ball park. If I had to guess how much of that 58% is from Hurst and could not be replicated by using something like a smoothed VHF, maybe I'd get 5% of the 58% from Hurst. I'm not exactly following scientific method here. I don't have the resources for that. I've never even produced a correlogram for trading purposes. I know this is sloppy but knowing how to interpret this stuff seems to be sufficient at this point. When I come across a problem I need to solve I'll search in databases such as JSTOR, skim through what I've found, and, if I like it enough try to figure out how to use it. Hell, I'll call/e-mail the prof, hire a coder, and maybe even change the algorithm. Whatever it takes if I think it looks promising enough. PM/e-mail me if you have further questions. I'm always interested in comparing notes when it comes to using quant techniques. -Lou
Could you clarify what your acronym VHF is? I assume you mean very high frequency indicator. By the way, that link you referenced is excellent; I wanted a way to simplify FDI on stochastic series with some meaning -- this approach is perfect for that. He also finds that the asymptotic limit of FDI converges after about 256 iterations. I might try to look at this on market series if I have time. I understand that you are using it in combination. Just thought it would be interesting to isolate hurst on its own merits. Thanks for the comments and I'll keep your PM comment on hand. -------------------------------------- sorry to side-rail the thread, carry on...
Evolving discussion is not side-railing. The academic nature of the topic has kept the trolls at bay. By all means, continue. I certainly will continue keenly reading from the side-lines until I feel the need to chirp in. The discussion and papers have been quite enlightening.
dtrader98, You're welcome. VHF is vertical horizontal filter, I think. You should be able to find it on any of the technical analysis sites. -Lou