<i>I'm curious as to the fundamental price movement the C2 system you have is trying to exploit. If it's a pair's approach, then you're doing something I haven't seen before because you don't trade a correlation. If its an indicator based approach then you're either doing some sort of vola. breakout/trend ID approach or a mean reversion approach. </i> Yeah, I'm sure if you saw the program you'd instantly understand Is there something I'm not explaining with volatility based overbought oversold conditions? What is a pairs system? When one pair gets out of whack, we sell short one and buy the other, right? Do you really think I need to calculate the correlation between QID and QLD? What do you think the answer is? .. .. .. .. I think it would be described statistically as perfectly negatively correlated, which is really what a quant based pairs system should be based on. I see a lot of yahoos including some here and on c2 that think positively correlated pairs are the way to go, when actually this is probably one of the worst combinations to start from. Take my word for it, a pairs system needs to find two nearly perfectly correlated pairs to work well. And, assuming these are perfectly negatively correlated, ie:correlation of near -1, then all you have to do when the long is overvalued is go long the inverse with 100% of equity. It's not that confusing. Why hasn't anyone picked up on this? Why the hell would I waste commissions by shorting QID and going long QLD when I can get the exact same effect by simply going long QLD? Put another way, when the pair is "overbought", why would I short QLD and go long QID when I can save myself the timing issues and just go long QID? Really, do you think it's not a pairs system? I'm saying I cared very little about finding a pair. I already knew which pair I wanted to perfect, and I didn't even have to calculate it's correlation because you already know. There's a mixture of indicators with a basic pairs shell, but the real nobel prize winning idea was to eliminate the static levels that plague long term performance of pairs models. I have developed a unique method of adapting for volatility so that essentially everything has been "normalized." That is, the formulas by themselves, though with different values each time, are the same fundamental equation. I might be mentioning too much about the system, but in reality it's based on statistical z-scores, and these are "normalized" to fit a bell curve. The results however, are definitely positively skewed. I know we're sharing files, but I hate to say I won't be sending you this one.
Beau, Looking at a few trades it appears that your model attempts to capture a form of mean reversion over the length of several days. The title of the system suggests pairs/scalping so I was under the impression that you were looking for quick stat. arb. opportunities between QID/QLD. Hence my comments were aimed at correlations between some of the QQQQ components and if you've tested your model on any of those components relative to QID/QLD. Mike
I was going to start explaining to TraderZones what you mention above. But I don't think I'm dealing with someone with a quant. background and my efforts would likely be wasted on closed ears. What you mention above is what I've developed to some degree. Its not the S&P system you're familiar with, its a stocks model that is adaptive and volatility based mean reversion (intraday only). The variables aren't really normalized per se as all the variables fluctuate based on recent volatility. The trick is forecasting vola. with good accuracy for the next trading day and triggering trades around vola. extrema. The real work is in the volatility model, which is the heart of this system. Anyhow - that's another discussion, but the below report is the model under simulated performance at 0.02 per side trading costs/$50k per position (i.e. no sizing algo), max. $1.5mil exposure at any given time. In real life the system gets about 85% of the sim. Initial capital 300000.00 300000.00 300000.00 Ending capital 4875429.45 4875429.45 300000.00 Net Profit 4575429.45 4575429.45 0.00 Net Profit % 1525.14 % 1525.14 % 0.00 % Exposure % 0.10 % 0.10 % 0.00 % Net Risk Adjusted Return % 1479132.85 % 1479132.85 % N/A Annual Return % 6.10 % 6.10 % 0.00 % Risk Adjusted Return % 5916.76 % 5916.76 % N/A -------------------------------------------------------------------------------- All trades 13430 13430 (100.00 %) 0 (0.00 %) Avg. Profit/Loss 340.69 340.69 N/A Avg. Profit/Loss % 0.68 % 0.68 % N/A -------------------------------------------------------------------------------- Winners 7602 (56.60 %) 7602 (56.60 %) 0 (0.00 %) Total Profit 10383497.04 10383497.04 0.00 Avg. Profit 1365.89 1365.89 N/A Avg. Profit % 2.73 % 2.73 % N/A Max. Consecutive 28 28 0 Largest win 36704.62 36704.62 0.00 -------------------------------------------------------------------------------- Losers 5828 (43.40 %) 5828 (43.40 %) 0 (0.00 %) Total Loss -5808067.59 -5808067.59 0.00 Avg. Loss -996.58 -996.58 N/A Avg. Loss % -1.99 % -1.99 % N/A Max. Consecutive 21 21 0 Largest loss -17969.80 -17969.80 0.00 -------------------------------------------------------------------------------- Max. trade drawdown -17969.80 -17969.80 0.00 Max. trade % drawdown -35.88 % -35.88 % 0.00 % Max. system drawdown -88519.38 -88519.38 0.00 Max. system % drawdown -21.51 % -21.51 % 0.00 % Recovery Factor 51.69 51.69 N/A CAR/MaxDD 0.28 0.28 N/A RAR/MaxDD 275.04 275.04 N/A Profit Factor 1.79 1.79 N/A Payoff Ratio 1.37 1.37 N/A Standard Error 737177.67 737177.67 0.00 Risk-Reward Ratio 0.08 0.08 N/A Ulcer Index 6.54 6.54 0.00 Ulcer Performance Index 0.11 0.11 N/A Sharpe Ratio of trades 3.82 3.82 0.00 K-Ratio 0.0070 0.0070 -1.#IND Anyhow, the EC is fairly smooth, this was tested on 5min data from 1982-2009 for about 700 symbols, it performs well across every market but daily implementation is a nightmare. Some days I'll have upwards of 300 trades and fill maybe 3/4 of them if I'm lucky. The EC improves as a function of vola, but during those times it also has the worst DDs. Also, I wouldn't share this model with anyone either - its my fund's core strat. and its taken years to get it right. I'll only share the decent stuff on C2, the great stuff stays with me and me only Mike
You're correct about the mean reversion, and I, too, did not want to get into quant with anybody because I doubt we're talking to that pedigree out here. You and I might be on that level, but it's beyond the scope of this forum to talk about those. Basically it's looking at QID and QLD's closing prices, but what happens after that, I can only say that I'm using standardized z scores that are adaptive for volatility for find critical values. It's actually calculating these critical values that are part of the magic, and, as I said, these values are not static as in the flawed backtests and systems I've seen elsewhere. I don't think we're actually working off that different of information. You might want to see what happen to your system on SSO.
Ditto to the implementation. WL's good at finding systems, not too good at actually being able trade, though. LOL.
Probably wouldn't trade that one. Win percentage usually needs to be around 70% to be viable in each year, otherwise I bet there are some down years. The 0.68% might not actually be enough to overcome transactions costs, or even taxes when you think about it.
The RAR to max dd is an interesting stat to look at, but I think it's much more valuable to look at APR to max dd.
Not sure where you came to the conclusion about the 70% figure. The win rate is arbitrary IMO as it hides the avg. loss stat. Besides, being right 70% of time is useless if you have several large losses. I tend to look at the sample size as the most valuable statistic. This system has generated over 14k trades in this particular portfolio, and the portfolio was picked arbitrarily. The point being that this concept works across all portfolios I've ever tried it on - regardless of the products traded. I essentially have over 200k trades to qualify this "edge" with. More importantly, the 0.68% assumes quite a bit of slippage as the average trade size is around 5k shares. I currently trade at 0.003 per share and my transaction costs amount to about 4% of total net profit, which is lower than all my other systems. Slippage rounds out to about 0.01 per share as I trade liquid products mostly. Also, no - there are no down years. Again, my point is the method in which one needs to qualify a true edge. The fact remains that it works in all markets under all conditions over a very large sample set. This allows for easy scaling and proves robustness. I would be interested in seeing results from a system that is intraday/long-only that has over 10k trades in the sample set with similar figures. I've found profit factors over 1.5 to be very difficult to achieve intraday with a large trade sample set. Mike