How to understand an edge

I wanted to bring up a very basic concept that seemed to throw off a more senior member in a distant thread; which is the concept of edge and expectation. A certain book had published tables on win and loss rates for a study the author had done. The poster mentioned that the edges were 'very weak' therefore, why on earth would they publish such a book. To my astonishment, several other posters chimed in with agreement.

I don't want to show the exact table because I'm trying to avoid sounding like an antagonist. But the basic concept was that the author had run millions of trials regarding the outcome of betting that a certain pattern would generate an edge. His initial hypothesis on a certain pattern was that it would generate positive moves following its appearance. He shared several basic observations in tabular format. Following are some of the metrics...
1) Win Rate
2) Loss Rate
3) Avg Gain
4) Avg Loss

The above were also tabulated for gain following several days of the pattern occurrence. So, we have everything we need to calculate basic mathematical expectation or edge.

Now here is an example of the IMO, egregious conclusion drawn by our more senior poster. Over 14+ million observations, the author tabulated
a 42% win rate with an average gain of 3.37% and an average loss rate of 58% and an average loss of 3.44%. The initial perception of a neophyte might be to calculate the expectation as E[x]=Aw%*Aw+Al%*Al or, .42*.0337-.58*.0344=-.0058 per trade or -.58% per trade and conclude it is a losing edge or system.

My initial reply was that a 58% win rate was respectable, to which the senior member was quick to reprimand me for misreading the table, which clearly showed it was a 58% loss rate. Now here's where I draw the contention and it would be a shame for others to draw the same misleading conclusion: It is only a 58% loss rate, if you are viewing it from the perspective of long only. You can easily take the same one day pattern and reverse direction to get both a positive hit rate and a positive expectancy.
i.e. the above equation becomes (+).58*.0344 (-).42*.0337=(+).0058/trade.

What we are looking for is expectation in any system being something other than zero (and ultimately positive). In terms of hit rate, we want something at the very least to be better than 50/50 (the market bias alone has this characteristic with positive drift, therefore relative to the market you need to compare biases, but that's another story).

Now assuming you agree with my contention on edge semantics, is .58%/trade weak? That depends on trading frequency and friction.
I would argue given a relatively high amount of capital to friction per each trade and a decent trade frequency, it is not weak at all. Is it statistically significant? For 14million trials, I would say so. While I'm not a big fan of many basic TA books, I applaud the author for his work here and do not at all consider it a waste of publishing time. I will take this type of book any day over the deluge of books available with zero stats and pure subjective commentary.

2c

Are you sure that the offered pattern can be "reversed"? If it uses SL or TP levels and not just "close on time" it most probably can't be "reversed".

As of how good is 0.58%/trade - it depends on the percentage of what is it? Position value, account balance? If the first one - then that's quite good.

What original thread are you refering to I wanna read it ?

I'm fairly certain that the particular study was triggered from end of close from day1 to dayn (of course there is no SL and TP factoring, which would cause large variance in results), in which case the theoretical pattern result is reversible. I'm also guessing it was likely performed across equity instruments. While one could argue that you couldn't have captured the same exact reverse trade, you could make the same argument that the author couldn't have captured the exact EOD data. However considering that he sampled over 14 million data points, I think such points are secondary and that likely at the very least the hit rate over such a large body of data is fairly reliable; the only minor difference on reversing friction and its impact on net expectation would likely come out awash as the average gains should be comparable. If anyone were to run the same back-test type of study, you would simply use the reverse signal (open/short) on each daily result following the trigger signal (or better yet, just flip the pattern entry signal and use the existing cl-cl data; in which case the back-test methods are identical, but signals reversed. You will seldom find studies/common back-test gear dealing with friction, etc.. (some academic studies include a small percent, but again it's an estimate) but that's a bit outside of the scope of the important issues here. PS. My example was on one of his poorer study results that I came across, many were in the 65% + type hit rate, which I have found similar success rates in some of my own work.

The .58% is an average cl-cl return following a trigger signal; so it is, IMO, quite good over such a large sample no matter how you want to measure it. However, as previously mentioned, if your positions are relatively large relative to friction (slippage/comm), and trading is relatively frequent, it is very good IMO as well. The percentages are relative to entry position amount per trade.

I'm not here to spark any negative debate or flame wars, I just wanted newbies to understand this concept. I'm not so much a newbie myself (if only on the board), but to see a seasoned senior saying such things with others agreeing, I was a bit aghast.

P.S.S. the ultimate contention was regarding why the publisher would consider publishing such a waste of results, which IMO is far better than many of the subjective candlestick books that do nothing but banter on subjective results. I consider his work closer to crabel, and worthy of merit. Heck, I might even go out and buy a copy now.

Care to share the book's name? So that we could buy it too ?

This point has been extensively discussed on this thread : Why inverting trades doesn't work.

It's quite easy to reproduce this result : find a system that blows up your equity, reverse it and it will still produce a negative equity result.

Yes I read that thread. And figured some posters would attempt to use that as a rebuttal, without thinking it through. Unfortunately, the results of that thread do not apply here. Nor did the conclusion apply to all cases; as mentioned earlier, if you apply SL and TP types of methods then there is no way to apply the logical reverse directly, other than empirical evaluation. In this case, you are only evaluating the close of one day to the close of the next, and reversing the sign of the entry signal only i.e. it is fixed. If you have done any type of backtesting with commercial software, most use this exact method and ignore friction, so I hope you are adjusting somehow to compensate, because as I said you can make the very same argument about using EOD or any other form of backtesting data. You will always have some kind of friction, but since you are only predicting one day ahead ; how likely will the hit rate be off due to spread or friction? Not much (if it did, then why even rely on any backtest software). Ok then, what about the avg gain/loss? Well, certainly you won't get exact EOD execution (you won't ever get the exact sequence of win/loss trades either, but you should have some confidence in millions of observations worth of expectation), however, if you think of the friction as noise relative to your 14 million trial backtest, you should have some entries above and some below the expected EOD value, so that as I mentioned the noise comes out awash over time (and you can also monitor this, but again, it's we're going beyond the scope of the original argument). And for other naysayers, if you truly think friction is going to change the reversal results of a fixed daily percentage change significantly over 14 million trials, then you might as well not buy into any of the forward results either, because the same argument can be applied in that direction as well.

I won't give the book name, but if you use a little sleuthing I'm sure you could find the thread (notice I haven't posted much). However, I'll do you one better-- I'll give you a link to a paper with similar conclusions and comparable results on the same topic with a bit more statistical rigor.

http://papers.ssrn.com/sol3/papers.c...#PaperDownload

The conclusion of that thread was that a system with a real edge can be reversed while a system with "random" entries can not.

And that is a possible answer to your initial question : an edge can be defined by its ability to be reversed.

Have a good day jamjamjam this discussion is starting to get very interesting !

Quoting jamjamjam

Disliked

Yes I read that thread. And figured some posters would attempt to use that as a rebuttal, without thinking it through. Unfortunately, the results of that thread do not apply here.

Ignored

You haven't actually given us anything to look through except your own analysis.

but i guess we'll have to take your word for it since you wont' link to the OP

edit: Also, the whole study means diddley squat if he curve fit the system. He just used a larger sample size than most.

Literally three posts ago I referenced an academic peer reviewed paper that you can go hog wild dissecting. Again, the only reason I don't link to OP is that I (primarily out of respect) am trying to avoid sparking some kind of unnecessary flame war with the OP (which is not my goal at all; quite the opposite); that being said, a little common sense and detective work should get you to the OP. All of the ideas anyone needs to discuss regarding the matter, have been discussed pretty clearly right here in this thread. Let's just use this thread in the newbie community to let the pros and newbies alike chime in, with all the source material right here.

The major sub point of this thread is regarding reversal of particular systems and definition of relative edge; if your contention is that the system is curve fit to begin with, then the whole argument of relative edge reversal becomes irrelevant. That's a completely different discussion (and even more controversial) for another day and another thread.

I'd appreciate if those replying could try to read the thread in it's entirety before commenting on issues that were already addressed; it's really not that long (yet). At the same time, I do appreciate the discourse from everyone who has something interesting to add.