Win Rate: The Most Important Performance Measure

A higher k-ratio indicates a higher positive consistency in trading performance. A ratio higher than or so is generally considered good. As with all Tradervue statistics, you can compare k-ratio between two different trading systems, for example, using the Compare reports.

Other popular styles requite high win rates because the payoff ratio is low. All About Expectancy Part Two.


Balancing Win Rate and Risk-Reward in Day Trading Day traders must strike a balance between win rate and risk-reward. A high win rate means nothing if the risk-reward is very high, and great risk-reward ratio may mean nothing if the win rate is very low.

This new ratio, which he named the k-ratio, was focused on consistency of returns over time and could be applied to daily, weekly or monthly data. The k-ratio provides higher values for systems that produced higher returns with less deviation from a least-squares regression line of the equity curve.

A good overview of the k-ratio is provided by Zephyr Associates who, for incomprehensible reasons, decided that it would provide clarity if they renamed the original version of the k-ratio as the Zephyr k-ratio. Kestner has since modified the k-ratio twice, once in and then again in , so the three versions will be denoted here as k-ratio , k-ratio and k-ratio The figure above shows the first and last 10 entries in a table of rows of entries. The k-ratio is therefore calculated in our Excel example as:.

In , Kestner noted that in mid trader Bob Fuchs brought a small error to his attention regarding the scaling of the k-ratio.

The net effect of this change was that the version of the k-ratio was smaller than the version by the square root of the number of observations. The paper is freely available on SSRN by clicking here. The number of observations in a year is 4 for quarterly data, 12 for monthly data and for daily data. The absolute value of the k-ratio is really of little interest other than we want it to be positive and bigger is better.

However, I am convinced that TF systems with low win rate can never be robust. Trends have existed in markets for centuries, and it would be safe to assume that trends will exist in the future as long as herding behavior and psychology play a part in free markets. Yes, the topic is polarizing.

I just argue that the win rate must be higher than some people think it should be. An example for a simple system can be found in this URL: If you remove some rules from the system, will it perform much poorly? Can you plug the system in to trade new markets without changing the parameters? The ratios are important but the system has to make sense to you and fit your style.

Ed Seykota is a source of inspiration. In when I started developing trading system with System Writer Plus an ancestor of Tradestation I used a triple ma crossover for TF and my partner and I made K in a few months trading currency futures. The system worked fine for one more year. Then it stopped working suddenly in and we lost some of the profits before we abandoned it. It has never worked since. I say this because what worked in the past rarely worked nowadays. Nowadays, my point is, a low win rate can lead to ruin.

Different strategies like MR and TF would complement each other in that they are likely to perform well under different market conditions and my view is that a set of complimentary strategies are easier to trade than each one in isolation. You can justify different win rates due to the nature of each strategy. This adds complexity, but may add to total performance.

Actually they do complement each other. I think I mentioned that in one of my replies. I appreciate you mentioning the risk of over-optimization and selection bias. The sample size can still drop below statistical significance.

In either of these cases, one could conceivably be misled by a very high nominal win rate, no? All performance metrics can be optimized and even liner or non-linear functions of them see earlier comment. Exactly, this is the point. I only say that win rate in the most important even in the case of trend-following.

Of course, I do not imply that the payoff ratio can be anything and not to pay attention to it. It boils down to this: If there is a sufficient representative sample, then the win rate is approximately equal to the probability of success of the next trade. I want this to be high. If it is low, then a streak of losers can ruin my system before the next trend arrives. Thus, I argue that trend-followers with low win rate are knowingly gambling and they just hope that a trend will arrive before a streak of losers or a streak of losers followed by a mediocre trend that will not suffice to cover part of losses.

I must say thanks to everyone for the great comments. I know many people are getting a lot of value from reading them. While win rate is very important I use other methods for judging system performance.

Nothing by itself is representative in my opinion, it has to be used alongside other metrics to arrive at a conclusion. I was going over your Price Action Lab product and found that your approach is very similar to mine.

A lot of what I do is automatic system generation as I feel thats the most hassle-free way of playing the markets. As with any automatic search, the results must be verified. In your blog you give a lot of information about this which is awesome. One approach that I am hypothesizing doing is instead of separating in-sample verses out of sample, I will take the entire dataset and mine for trading systems. Each individual system performance from this search is used as an benchmark.

I then do the search N times again; each time I will randomly select an interval of data to search. Each of these N search will come up with models that may or may not be in the initial full sample search. Tally up the instances whereby the interval search came up with the same trading system or similar.

The more times a system is repeatedly discovered again, the more robust the system is. A step further would be to do a out of sample test.

Since you randomly selected an interval, there is the other half of the data. Test the models against that other half and compare performance to the full sample search…idea is the closer the performance the better. I agree that one must use other metrics in addition to win rate. One that I use is the profit factor. The data-mining procedure you described is interesting provided that the sampling is done with no replacement.

If replacement occurs, then data-mining bias increases due to data reuse. This is of course if I understood correctly what you are doing. A method I prefer involves cross-validating first the data-mining method using random data.

If the method is sound, then just ignore out-of-sample testing and use the benchmark models. Perform portfolio backtests to increase samples. Here is an example: The method I proposed is essentially a way to judge the consistency of the trading system. If every trade made throughout the entire testing period had the same performance, then it should end up being discovered every time if searched for it in smaller intervals.

Astute observation on sampling without replacement. The maximum N should equate to the size of the interval and the size of the OOS. Are there any procedures you take to filter out some potential candidates?

For example, I understand that you suggested portfolio backtesting. Let say you discovered models. All match your desired performance. You then do a portfolio backtest. Half of that sample become inferior when you test it on a correlated instrument. I feel there is subjectivity involved passed this point.

You are correct and this is the reason. The final system includes all systems and no selection is made based on any test. The portfolio backtest is applied to a a system that consists of all systems and if it fails then all systems are rejected.

I am not following. This is an involved process and maybe the subject of another article. If you apply a data-mining algo to random series and you get many systems that satisfy the metrics, then this could mean that the algo is curve-fitting to noise.

What do you think about trailing stops and chandelier exit stop trading systems? Do you think that exits are more important than entries? I noticed that trailing stops are not one of the exit choices. He told me that he has been successful with it and I am thinking of buying it. I have not had any success with NNs in the past and I have also used genetic programming but results suffer from curve-fitting.

Also, what do you think about maximizing the Sharpe ratio. Thank you in advance. See this article for more details: I also believe that entries and exits are equally important and my experience says that systems with random entries are artifacts of curve-fitting.

However, opinions may vary on this issue. The Sharpe ratio is an important parameter because it is directly related to the statistical significance of the results see formula in earlier comment but focusing on it may cause you to reject some good ideas that you could latter improve. In my opinion this parameter should not be optimized but only be used to evaluate the results.

Otherwise optimization may be to circularity, i. How are they profitable? Because the average win is twice as much as the average loss. Profit expectancy is the most important factor to maximize, which takes into account probabilities of win and loss, AND the accompanying profits per trade.

For more details see: The fact that CTAs have low win rates does not imply that win rate is not the most important parameter. It implies that there are limitation in the way they trade.

Ask any CTA and he will tell you that he would love to have a higher win rate. Other popular styles requite high win rates because the payoff ratio is low.

As you can see from the equation I wrote, the goal of trading is indeed to be right as often as you can because that increases profitability. If you would like to be just profitable, you could manage that with lower win rates. But a higher win rate will increase profitability and this was the main point here. The fact that some people cannot do it it does not mean that others should not try. The point that almost everyone here including the author is missing is that win rate and payoff ratio are inherently linked.

Raising the win rate lowers the payoff… raising the payoff lowers the win rate. This is a mathematical fact. Win rate is no more important than win SIZE when it comes to risk of ruin. Those leaving comments who refer to expectancy as the most important attribute to system design are correct.

There is no arguing this fact… and if you continue to do so you show your absolute ignorance of one of the basic precepts of actually understanding trading. This means that you try to achieve maximum win rate for maximum achievable or desirable payoff ration.

On the contrary, most of those that talk about expectancy are probability fools. Expectancy is nothing more than the average trade.

It is not expected value. Only for sufficient samples it become expected value. That means many many trades. BTW, expectancy is a linear function of win rate. The higher the win rate, the higher the expectancy for given payoff ratio.

Therefore, E is maximum when w, the win rate, is maximum. Since one cannot control avgloss in general, maximizing w maximizes expectancy in general. I recommend that you get up to speed with these simple equations. They say the whole story and debunk your claims that are based on wishful thinking and not on mathematical facts.

Unless you plan on placing only ONE trade in your trading career your 8th grade algebra is useless. It works fine for some of the situations but not all. Which system do you prefer?