I am writing this at the request of a few readers and a few people in Joe Donahue’s chat (which, by the way, you can learn a lot in if you happen to join Joe’s premium service at www.upsidetrader.com (this was a totally spontaneous and unpaid plug for his service on StockTwits.com by the way). Lets talk about models, but not fashion models. I mean trading models. I will likely make this a multi-part series because I am in the middle of two business projects and I want to do some additional back-testing of an hourly forex model. For tonight, we will simply address the numbers.
In later posts I will discuss how I built models using NeuroShell Day Trader Professional Version 6.2. You can find out more about it by going to this link or to this link for more information from Ward Systems Group, Inc.
I for some reason cannot find the original post I did way back in 2004 in which I made comparisons of trading models to models like Jessica Simpson. I think I stated that if trading models were as inconsistent and unreliable as Jessica Simpson that one should basically stop trading. Why is that? Well, it has something to do with statistical expectancy. B.C. Lund wrote brilliantly about such things in his blog post this weekend. Study what he shows in terms of the number of losses one can suffer as reward to risk slips to parity. To quote Mr. Lund at the end of that table, “You can see that if you only take trades that have a 1:3 risk/reward ratio, and you are correct just 50% of the time, you will have a 10R profit on ten trades.” What one must realize is that during that sequence, one will lose on EVERY OTHER TRADE, so that reward/risk ratio will have to be skewed in your favor strongly if you are to end up with profits.
What you see at the top right of the picture (and quit staring at Faith Hill will you?) is a model for Cheesecake Factory (CAKE) I built in August of 2010 (that is right, about 18 months ago). It has three key statistical characteristics that I like in a trading model:
1) Look at the red box a the top right (percent change in price and annual change in price). Those are the numbers for a pure buy and hold from August 25 2010 to February 2, 2012. The box below is is what the return would be from the trades the model produced. The model on its own produced a better return than pure buy and hold (and by a decent margin). That is a primary sign of a good trading model (I will comment on tax implications at a later time, but the point is, the model traded better than if you had simply bought and held the stock.)
2) The win to loss ratio which is $3.69 made for every dollar lost. Typically, I would like to see this ratio at least at 1.60/1 or higher before I would trade it (as is discussed on the Lund post).
3) The percentage of wins to total number of trades is 83.3% (basically 8 out of every ten trades is a winner). My minimum acceptable winning percentage is 60%. Tom Joseph, an expert oil trader who developed Advanced GET a couple of decades ago, used the 1.60/1 and 60% statistical model as his minimum acceptable statistic because he figured you needed at least that kind of edge to cover both for losses and potential trade errors that could be made. I think he was right when he said that.
Models have improved over the years. Jessica Simpson dumped Tony Romo, got married and is now pregnant with her first child. Faith Hill, well, you saw her at the start of Super Bowl XLVI. She hasn’t aged a day and didn’t need to lip sync her intro the way Madonna lip sync-ed her Halftime show. Neuroshell Day Trader Professional has gotten better as well. In 2001, it might take 45 minutes to build one model. Now, it takes as little as 15 seconds given the processing speed of Intel i7 chips and the ability of the software to use multiple cores from all areas of a computer network.
Do I believe that only neural nets can produce trading models with positive expectancies? No, because I have seen and personally traded models that were thoroughly back-tested and consistently make money. Why do I use them then? I use them because they provide me with statistically viable stable models that are basically “buy at open” with stops and targets at the ready. I do not have to use finesse or emotional judgements to make a trade. It eliminates the hesitation to pull the trigger when I have to. I already know that not every trade is a winner, but I have the confidence that the statistics are still on my side when I trade.
In the next installment of this series I will discuss the basic components of what make up a good trading model, and not just a good neural net trading model. There are components of the model that make the good basis for a specific trading plan, which most traders don’t have. Typically, that is the reason most traders fail to make money.
I hope you ladies were not offended by this post, but let’s face it. Some models ARE better than others. The numbers prove it.
Thank you, ladies and gentlemen, for supporting this blog!
{ 2 comments… read them below or add one }
Hey thought you had disappeared at least off twitter. Glad to see your back. Imo, personal I think it is a big misconception to think any trader can achieve 1:3 on 50% of his trades. Just “collecting any proft” consistently on 50% of your trades is challenging much less expecting every other trade to achieve a 3R seems kinda silly. If anyone could do this they would not be writing on a blog but buying up tropical islands w/any kind of position sizing.
One thing Dave and I am just mentioning it is the backtest period covers a very bullish markets phase, they have been moving up since 3/09. You have very good results w/your strategy though and I am only seeing ratio of 2.96 w/mine.
Rob (cebuBound)
But you can rob, and in the case of forex, if you have a long short trading mechanism, you can indeed make money that way as long as you trade every sequence that the model gives you and you do not try to change the entry rules. What most traders fail to realize is that all models do draw down over time. There are folks on the Twitter stream who in fact are very large traders and they do indeed make decent money on 3:1 (reward to risk) ratios even when 50% right. The key there is position sizing and discipline. During drawdowns, one has to reduce the size of the position until things turn around. The key is sticking by your rules and having the discipline to take every trade in the sequence,
A previous model for cake in a not so bullish period did pretty well also, but, as you would imagine, it didn’t trade much. I will get into depth on that in future postings.What I am going to cover in the coming weeks with long only models does in fact deal with how to work with non bullish periods (that is using a screen to skip certain trades, even with a neural net model).
I am now working on a start up and real estate project, so my time on Twitter and Stocktwits is limited. I will do what I can here and on those streams as I have time. Take care Rob!