Thanks for the informative thread. As someone with a background primarily in intraday trading, the first thing that I notice in the performance results is the low profit per share (.004357). Does the fact that you guys are not addressing per share related stats here indicate that you are less concerned with those and related issues like overcoming trading costs, efficiency etc, or is it just too early in the process to look at that? If I am looking at this from the wrong angle (a day-trader's perspective), if you would, please show me where you think that my focus should be in analysing performance stats at this stage in the process as it relates to longer term trading systems.
Hi Talon, This is a good start for a system. With some tweaking and additional filters, one can make this into a tradeable, portfolio level system. I added a few of my propriatary filters to this basic idea and tested it on a portfolio of about 600 stocks from 1997 to present. The results are attached. The historical price data is good BTW, its not Yahoo data... I'm not going to give away my filters, but, they basically filter for vola. So, to reproduce something comparable to the below result, do the following: 1. Devise a method to trade when volatility is above a certain threshold for the specific product. 2. Play with the EntryPct and ExitPct variables. Find a combination that "fits" all stocks/products, not just specific ones. The trick is not to be too general or too specific. 3. Trade only higher priced stocks with ADV > 1mil. The below assumes the following: - One way transaction costs = 0.015/share (my estimate for comms + slip.) - $2500/trade position size - Max 30 position open at once. - $50k starting capital. - NO STOPS!!! haha.. Thanks for you input talon. This is a great template for someone looking to develop a tradeable system. Where is Bowo BTW? Be careful, don't give away too much - he'll take this system and then try to sell it on C2 for 5mil, or $200/month... LOL... Mike
Before applying a vol filter, if we use a fixed position size and do not employ stops, we are already taking on more risk in high volatility periods/stocks (of course with higher upside as well). Then as shown in the above example, performance is improved by taking this high volatility bias further with a filter. Is there a fundamental reason for mean reversion systems not to work as well in low vol environments other than the scaling of returns? Or is this just the result of selecting this position sizing scheme (ie. we know the systems works, so we aren't concerned about the increased dollar risk to get the larger winners)? To build a system aimed at consistency above all else (income in low vol periods), does it make sense to backtest/optimize with pos size = 100 / (20 day atr / close); or something similar?
...taking on more risk in high volatility periods/stocks That's exactly where I'd want to take on the most risk! It presents the most ineffiecent pricing! Personally (and this is merely a personal approach), I always develop trading systems with no stops first. I'll then apply stops if the system permits. As a rule of thumb, I don't use stops on intraday systems and use "worst case" stops for longer time frames. In nearly all my research, I've found stops usually do more harm than good. This particular system, as graciously provided by our thread starter, can benefit from the use of stops... I'll leave it up to talon et al. to develop that aspect. My preference is to just use small position size across many products. In essence your position size is your stop. Don't go around throwing all your eggs into one basket. Is there a fundamental reason for mean reversion systems not to work as well in low vol environments other than the scaling of returns? Yes. Think about this for a bit. Mean reversion means the price is not at the mean, right? The further away from the mean we go, the greater the profit opportunity. The trick is developing a measure of what constitutes "abnormal" volatility. Low vola means less "abnormal" price inefficencies... I would say never never never optimize position size. Just don't do it. The edge of your system has nothing to do with position size. When you're trying to validate and quantify the edge, position size should remain constant. After the quantification is done, then and only then *might* one want to mess with position sizing. In general, serial correlation is an issue with mean reversion systems. That is important because any position sizing algorithm will imply correlation and causation, which is a fallacy IMO. Think path dependency here... Mike
Well in backtesting you will see pretty consistently that most technical ideas for stocks do better on the long side than the short side. For a lot of mechanical systems, the best you can hope for on the short side is that they break even after transaction costs. So this raises an obvious question: Why not just trade the longs? You certainly could do that, but there is always the possibility of a longer term bear market (which, by most measures, we still have not see yet), or even of a return to a very long trading range. Most of our system development is done on the last 20 years or so of stock data, and it's possible that we might have a regime shift that would change this bias. So, my vote is to usually trade the shorts as well. In terms of process, remember that any system test is a joint test of all the system conditions. If you really want to understand the market behavior you need to isolate the moving parts. A simple way to do this might be to take the entry condition alone and test it with a fixed time for exit (1 day... 5 days... etc). If you do this, you do want to correct for the upward drift. An easy way to do this is just to collect the trade returns and subtract the average return for the series (for each individual stock). At least you want to make sure that your long entries beat random entries, right? This is not an invitation to a relative performance measure... it still has to make money on an absolute basis! You could also detrend as you suggest. You could well have different long / short entry conditions, but then you are back to adjusting for the long bias of the more recent stock data. In general, I try to keep rules symmetrical, but this may not be the best answer. Yes the distribution looks "real". I would initially be concerned about the handful of very large losses, but keep in mind this is one step in the development process. At this point, if this were a fresh idea, it certainly looks to be worthy of further investigation. I think you are exactly on track here. Don't forget the value of visually checking to see where winning and losing trades cluster on charts. This is extremely time consuming and you frankly have to train yourself to see patterns... this can take days to do right but this is where you have a real edge to employ your human pattern recognition skills. I do think this is a more advanced trading skill... you don't have to do this, but it's something you want to work towards.
Hey TZ, what is your point? Are you suggesting that I intend to launch a website? Is this what you are trying to say? If so, man up and say it. If not, then what is your point?
I think of everything as a percentage return, not a per share return. That is the way to normalize across all price ranges. In fact, the actual performance on a percentage basis is also low at this point in the process so I think your observation is a good one. it's obvious that we're not done with this at this stage and the factor you highlight is a critical one. Think percentage not per share though. Do your system tests as a fixed dollar amount never a fixed share amount.
Mike, thank you. I don't want to write a half assed reply so I will get to this soon. I have some thoughts on volatility I was planning to share anyway and this provides a good framework for those. I will try to do it later tonight but I'm a little short on time... and I want to leave BoWo another little present over in his thread too.
Thanks for the reply Mike. I was more inclined to think in terms of having constant risk rather than constant position size. Which in my opinion are different, even with the 'your position size is your stop approach'. I'm not intending 'constant risk' to mean any kind of fixed stop or specific amount but rather just a function of recent realized volatility and position size. Sorry to be slow about this, but would you mind confirming if I understand where you're coming from through this example. Say we have a MR system that assumes yesterday's close as the "mean" and: 1. Goes long when daily return goes below 1.5 x (20d ATR / Yest Close) 2. Exits at 1 x (20 d ATR / Yest Close) or the close Then we compare trades...one in a high vol environment and one low vol. High Vol: 20d ATR / Yest Close = 4% 1. We enter w/ stock down -6% on the day 2. Exit when stock rallies back to down -4% or the close Low Vol: 20d ATR / Yest Close = 2% 1. We enter w/ stock down -3% 2. Exit when stock rallies back to down -2% or the close In my mind, the stock is approximately equally likely of being down 8% at the close in the high vol scenario as it is to be down 4% in the low vol scenario. In that case with $1000 fixed position sizing we would lose ~ $20 in the high vol and only ~ $10 in the low vol (with equally lopsided rewards). However, if our position size adjusts dynamically with recent vol, say = $40 / (20d ATR / C), we make $20 in both scenarios and keep the risk/reward unchanged. Without yet acknowledging your point on abnormality, this seems ideal since we are not feast or famine across vol regimes and will not be encouraged to bias our system development towards high vol environments. This (finally) leads me to the question: Are you saying that mean reversion is about the absolute amount we are outside of a "normal" range regardless of the recent volatility regime? The (4% - 6%) 2% in the high vol scenario will have a stronger tendency to mean revert than the (2% - 3%) 1% in the low vol scenario, despite the fact that they are both 1.5x outside of "normal" bounds (by recent standards)? Sorry I wrote a lot there just to ask a simple question, which I'm sure you followed the first time, but a clarifying example always helps to get things through my rock head. Thanks again.