Jan

30

An Anomaly is Becoming More Frequent Again, from Victor Niederhoffer

January 30, 2009 |

9 Comments so far

1. Rocky Humbert on January 30, 2009 1:06 pm

   I was a perma-bear in JGB’s during the late 1990’s, and the pain of that experience is etched in my mind.

   The current behavior of the USH9 contract is eerily similar to the JGB from back then. The JGB would repeatedly grind to absurdly low yields over a period of several a few months, and then swan-dive. A difference, however, is that the Nikkei would be comparatively bid during the swan-dive. The S&P is lacking that trait.

   Hanging on my wall is a screen shot from 9/11/98, which was the day when 3-month Yen Libor traded negative at Barclays Bank.

   History may not repeat, but it rhymes.

2. the boy on January 30, 2009 3:00 pm

   Totally agree. everyone was a JGB bear then just like they are now on US bonds.

3. Chris Monoki on January 30, 2009 3:44 pm

   The aftermath of the ‘29 crash left many investors, players and speculators scarred to the extent of failing to adjust to the ‘new era’. It took almost a generation of new players to pick at things that they saw as undervalued, but that their predecessors still deemed too risky.

   No different was after the dot.com bust in 2000. Some colleagues of mine were still thinking Yahoo, Cisco and that ilk and failed to think of anything else. ‘When to get back in the same names’ was their strategy rather than thinking of new new things. They (admittedly myself) missed the run in energy- and housing-related stocks.

   Perhaps we’re entering a similar era, and perhaps I’m with the old crew, incapable of grabbing opportunities because of my heightened risk tolerance. My goal is to avoid that trap, constantly asking, ‘what’s the new new thing?’ Will there be a new market to play in? Is is preferreds? Will a certain sector, industry (re)surface in the commons? I don’t quite know yet, but I’m seeking.

   Keep pressing,
   Chris Monoki

4. douglas roberts dimick on January 31, 2009 4:54 am

   V, have read this article?

   How to compare different loss functions and their risks [in machine learning theory]

   www.c3.lanl.gov/ml/pubs/2005_loss/paper.pdf

   The calculations are beyond me, though the theory itself appears to present an interesting application of stop loss correlating to both classification and density level detection.

   As described in the abstract hereinbelow. I am wonder if surrogate class could be analogous to a price action range setting derived from a computational model of prior (session or multi-session) drawdown?

   dr

   September 7, 2006
   Abstract
   Many learning problems are described by a risk functional which in turn is defined by a loss function, and a straightforward and widely-known approach to learn such problems is to minimize a (modified) empirical version of this risk functional. However, in many cases this approach suffers from substantial problems such as computational requirements in classification or robustness concerns in regression. In order to resolve these issues many successful learning algorithms try to minimize a (modified) empirical risk of a surrogate loss function, instead. Of course, such a surrogate loss must be “reasonably related” to the original loss function since otherwise this approach cannot work well. For classification good surrogate loss functions have
   been recently identified, and the relationship between the excess classification risk and the excess risk of these surrogate loss functions has been exactly described. However, beyond the classification problem little is known on good surrogate loss functions up to now. In this work we establish a general theory that provides powerful tools for comparing excess risks of different loss functions. We then apply this theory to several learning problems including (cost-sensitive)classification, regression, density estimation, and density level detection.

5. the boy on January 31, 2009 7:53 am

   Think alternative energy. The new administration is pushing for a new energy plan for the country. These stocks rose during the oil bull market and have now come way off with oil and the overall market. Most think they are still an oil play, which seems very wrong to me as Mr. President has stated he will not let temporary falls in oil price stop him from making long term changes in energy policy. The ironic thing is that in 20 years its very likely these alternative energy investments will look like they were a bad idea because the investments themselves will help to keep the price of oil down, which in turn makes alternative energy less attractive. But truthfully if you include the cost of funding wars and fighting terrorism in the NPV calculation then the investment is most likely a great one. At any rate whether its beneficial in 20 years or not, this is the area that the current govt seems to want to push, which is important since we are in times when the government is the only making any long term capital commitments.

6. Curmudgeon 6721 on January 31, 2009 11:36 am

   Mr. Dimick, thanks for bringing up that article about Machine Learning Theory. The dream of machine learning theory is that you feed enormous amounts of market data to an algorithm, which automatically sorts the wheat from the chaff and learns to recognize what makes a situation bullish or bearish; the resulting classification algorithm can then be applied to any situation to tell you whether to buy or sell. The algorithm has figured it all out by, in essence, trying every combination of inputs until it found what is best.

   It sounds good in theory. But experience shows that the followers of Dr. n who have applied a purely mechanical approach have not done as well as those who discover trading rules by using all the human insight that they are capable of, as well as computer support. The markets are too subtle and too close to random and too subject to change for automatic statistical classification methods to be highly successful in this field.

7. the boy on January 31, 2009 12:41 pm

   Trend following has been around for over 30 years now. Purely mechanical. Has returned around 8% annualy with max drawdown around 20-25%. Very similar to long term returns on stocks, though stocks have the much larger drawdown. Any argument over survivorship bias is equaled by stock index calculation methodology. Haven't seen any of these, even the ones that stank, lose 100% of capital any one time, much less more.

8. Curmudgeon 6721 on January 31, 2009 1:15 pm

   the boy and I are speaking past each other, perhaps because I did not make myself clear. I was speaking of mechanical approaches to rule construction, where the trading rule itself is designed by a computer with no human intervention. Neural networks and SVM are examples of this approach (which I claim is not successful). The tf rules were not constructed by a computer, but by a human being, namely Richard D. Donchian; so they are not mechanically generated rules but human generated ones. All rules, once generated by whatever means, can then be applied mechanically (or not). But that is another subject. I hope I have made myself clear this time. I have nothing to say for or against tf (a subject on which too much has already been said); I am talking about Machine Learning Theory (NN, SVM, …) approaches to financial markets.

9. douglas roberts dimick on February 8, 2009 5:45 am

   Curmudgeon, thank you. See comments at LD’s triangulation article: http://www.dailyspeculations.com/wordpress/?p=3561

   His comment reference to Steven’s Handbook may prove revealing a la your observations on machine learning. As of yet, I have only surveyed the book; it appears, though, to offer applications as to how rules-based architecture may validate (machine logic) program systematics – being execution protocol via market strategy and situation parameter issue calibration and excitation – rather than specific algorithmic formulation of market-strategy positioning.

   Right, no doubt about it: quantitative program theory sounds good. Of course it does; optimization occurs within closed-loop formations. A girlfriend (when I was at Georgetown studying lit) was from RPI and a military industry consultant specializing in AI; she said then (20 years ago) that this same approach was what had flawed and impeded AI work to that point in time.

   Thus, the question: what is the process for articulation, selection, application, and operation of those very rules characterizing said closed-loop systems?

   There was – for me – a sense of affirmation when examining Pitt’s article-cited video of program trading systems as therein recounted by Madoff’s head trader (Josh). His “herd” of animals analogy was entertaining; revealing was his observation on (non)correlation among and within the different sized animals (or funds). The point discerned?

   Markets imitate, perhaps emulate or create organic constructs that, yes, correlate with varying frequency. However, those same exchange processes also assimilate; therefore, an algorithmic program may optimize itself into becoming one if not the determinative dynamic of any given market exchange series, so defined by position size and directional rate(s). As now known and as may be construed from Josh’s overview, a representative program’s correspondent position is nominal when designed to quantify and then correlate an investment that instead becomes the exchange series itself.

   See where Josh talks about one big animal turning around to exit out the door. What appears to be pre-calibrated, separated positions within a finite, quantified market exchange, so becomes an evolving circuitry of input-output synaptic processes, so connecting (like musical chairs) those “automatic statistical classification” programmed positions. Moreover, as Josh so reveals, those closed-loop designs accordingly fail to account (or recalibrate) for the “Quantitative Relativity” of both (a) “their own” respective dynamics (or individual and collective fund positioning) as well as (b) parallel or collateral operatives (or relative size and entry/exit timing of other funds with correlative or contra program excitation).

   In conclusion, given my past year of work with state machine logic, I find that the determinative aspect of rules-based application here is neither, algorithmic nor quantitative, as you so indicate.

   However, I do not agree with your conclusion as to “automatic statistical classification” at least beyond the scope of this lead article, being frequency of one or more anomalies – correlating or otherwise. As I have yet to ramp up from beta to cash, I shall keep any further words on that matter where my few dollars lay – out of the market. dr

   Ps. I apologize for the tardy response – did not see your post.