The RSI indicator has served as the foundation of several popular short-term trading strategies over the years mainly due to its popularity. In the article I compare the results of a basic system based on a two-period RSI to those of a system that is based on a two-day losing streak.

Larry Connors and Cesar Alvarez originally used the RSI(2) in developing trading systems. In this article I show that the basic RSI(2) system has underperformed a simple two-day losing streak indicator during the SPY uptrend since March of 2009 and in GLD with data since inception.

**RSI(2) System**

Buy at the open of next day if RSI(2) < 10

Sell at the open of next day if RSI(2) > 70

**Two-day losing streak system**

Buy at the open of next day if Win Rate (2) < 10

Sell at the open of next day if Win Rate (2) > 70

The win Rate (2) indicator has a value of 100 if there are two consecutive winners and the value becomes 0 for two losing trades in a row. Actually, this indicator assumes the values 0, 50 and 100 when the period is 2. Thus, the second system buys if there are two losing days in a row and sells if there are two winning days in a row.

**Backtest settings**

$100,000 initial capital

Equity is fully invested at each position

$0.01/share commission

**Chart example**

Below is a chart that plots the RSI(2) and Win Rate (2) indicators.

**Backtest results for the period 03/06/2009 – 05/22/2015**

Parameter | Win Rate(2) System | RSI(2) System |
---|---|---|

CAR | 14.98% | 10.58% |

Win rate | 75% | 73.77% |

Max DD | -16.06 | -14.97% |

Trades | 120 | 61 |

Profit factor | 2.54 | 3.45 |

Sharpe ratio | 2.04 | 3.02 |

Payoff ratio | 0.85 | 1.23 |

No. of losing years | 0 | 0 |

It may be seen that although the RSI(2) system has higher Sharpe, profit factor and payoff ratio, its CAR is lower by about 4.5%. The reason for that is that the Win Rate (2) system trades more often and although the expectation of the RSI(2) system is higher, it realizes more profit. This is an aspect of trading systems that new trading system developers usually ignore, i.e. **systems must be able to deliver quantity in addition to quality**.

I wanted to compare these two basic systems in this particular time period because it represents a tough one for traders to match the buy and hold return of SPY before dividends that comes close to 23%.

The conclusion from this first example is that simple systems that obey Occam’s razor usually outperform more complicates systems. In this particular case and when considering only two periods, the RSI(2) is a much more complicated way of trying to gauge mean reversion than a simple indicator based on a 2-day losing streak.

**Modified versions of the systems**

If the exit signal occurs instead when the close is higher than the 5-day moving average, the Win Rate (2) system still outperforms the RSI(2) system by about the same margin.

**Results for GLD with data since inception**

This has been a tough market, especially after the downtrend in gold price that started in 2012. The backtest results with GLD data since inception 50 05/22/2015 are shown on the table below:

Parameter | Win Rate(2) System | RSI(2) System |
---|---|---|

CAR | 9.52% | 5.61% |

Win rate | 63.6% | 66.07% |

Max DD | -33.27% | -22.79% |

Trades | 217 | 112 |

Profit factor | 1.28 | 1.36 |

Sharpe ratio | 0.83 | 0.87 |

Payoff ratio | 0.73 | 0.70 |

Losing years | 2013, 2014 | 2006, 2008, 2014, 2015 |

The Win Rate (2) system CAR outperforms the RSI(2) system by 4%. Also, the RSI(2) system has four losing years as compared to only two for the Win Rate (2) system. However, the latter shows a large loss for 2013.

**Summary**

Most short-term indicators can be effectively approximated by much simpler price patterns. Actually, that was the main idea behind the Price Action Lab software. By adhering to these simpler approximations, we respect Occam’s razor and simplicity over unnecessary complexity. Simplicity was the main reason that the Win Rate (2) indicator outperformed the RSI(2) indicator by such a wide margin. In general, it is true that systems that can deal with the same information but in simpler ways are more robust than those that require higher complexity for achieving the same goal. In essence, this is another reason that machine learning methods, evolutionary algos and neural networks do not usually work and more and more users of such complicated approaches to trading are convinced that the data-mining problems that plague them cannot be easily resolved. Simple and deterministic trading rules lead to less data-mining bias by limiting complexity and randomness and thus keeping the entropy of the trading system development process low.

You can subscribe here to notifications of new posts by email.

Charting program: Amibroker

**Disclaimer**

© 2015 Michael Harris. All Rights Reserved. We grant a revocable permission to create a hyperlink to this blog subject to certain terms and conditions. Any unauthorized copy, reproduction, distribution, publication, display, modification, or transmission of any part of this blog is strictly prohibited without prior written permission.

I like simple systems such as this one. This one definitely appears to be a winner (at least with a quick test). One thing I notice though is that the average G/L using SPY (again, on a quick test) is only about 0.5%. I’m guessing you didn’t include commissions in this? It would take a very high position size and/or low commission size to keep this profitable. For mere mortals such as I, this winning system becomes a losing one when introducing commissions. But it’s still interesting, and perhaps there’s a filter that can be introduced to provide more bang for the buck.

Hello Matt,

I used $0.01/share as noted in the article. In SPY, from inception to 05/22/2015 I get an average trade of 0.72% and average win to average loss of 0.81 for the RSI(2) system. Note that the performance of this system was basically flat in 2014 and this could mean that its potential gains are being arbitraged out. In comparison, the win rate (2) system returned 10.2% in 2014. I expect both systems to become unprofitable soon if serial correlation returns to the markets. Thanks and best regards to you.

Michael,

I too prefer simple rules. Looking at your original stats, I wanted to know the exposure % for each method. To me, it is unclear which is better without this. Everyone has their own statistics that they like to look at when comparing strategies and this one is important to me.

I ran your original rules and got very similar results. The exposure for RSI2 is 18% while it is 40% for WR2. That is a big difference to me.

I love down days and use it in the mean reversion strategies that I trade. But one problem with it is that you cannot test finer values. WR2 only has 3 values. With RSI2 you can test many different values.

I ran in optimization on different entry and exit levels on RSI2. What I was looking for was values that gave an exposure near 40%. That to me would be a fairer comparison, for me. I found that the values of RSI2 < 20 and exit over 80 gave an 40% exposure. This gives a CAR of 15.68, MDD of 14.04, trades of 119. Only slightly better than WR2.

But looking for the highest CAR combination, results in using an entry of 30 and exit of 90. This gives a CAR of 20.95 and MDD of 15.34. Then again I would not trade this since the test period is basically a very strong bull market.

Again this is not to say that RSI2 is better than WR2. They are different indicators and one needs to understand the pros and cons of each.

Thank you for your good posts,

Cesar

www.AlvarezQuantTrading.com

Hello Cesar,

I briefly talked about the different exposure of the two systems in the article and this is also evident from the number of trades in a way. I see exposure as a parameter of a system and not something to match necessarily across different systems because in many cases that may not make any sense. Probably in this case you are right and it may make sense. I will have to think about that.

Regardless, the purpose of the post was to show that there are simpler indicators than can match the behavior of more complex indicators. Actually, I would not trade any of those systems because my analysis says that they have no intelligence but instead rely on a mean-reverting tendency of the market that may suddenly disappear. I will have a post on that next week I hope.

Thanks for your comments. Your input is appreciated and provides food for thought.

I agree that simpler indicators can match more complex ones. Once I adjusted for exposure the results were similar.

Cesar

Hi Michael,

Thank you for the article. I look forward to reading everything you publish.

Kind Regards,

James

Hello James,

Thanks. Good luck with your new software. I see it's going along well.

Here is the website URL for those interested in machine learning: www.adaptivetradingsystems.com/

Michael

This is a very ցood tip particularly to thoѕe fresh

to the blogosphere. Вrief but verү accurate info… Appreciate your sharing thijs оne.

A must rеad post!

Very good article Michael. I love your work and I my trading depends on the signals generated by your PAL software. Looking forward to articles about using robustness and portfolio backtest in validating price action strategies. Michael

Hi Michael,

I wanted to ask. The RSI(2) only uses 2 periods, my understanding is that this isn't a large enough sample size to produce reliable statistical results. Of course you pointed out that the system makes money.

My question is: Is there a reason to justify the 2 periods? Is there a specific reason 2 periods works or is more reliable over the others?

Kind regards

Jacques Joubert

Hi Jacques,

Usually the sample size refers to the number of trades and not to the number of periods used by an indicator. As far as your question, I think it's a good one but I would like to point out again that I 'm not the inventor of the RSI(2) system although, as far as I know, I was the first to talk about the WR2 system (win rate (2)). However, it's possible that someone else has already done that in the past.

I would speculate that the answer to your question is that 2 periods works fine because of the persistent mean-reverting nature of the market after serial correlation was arbitraged out by algos in the 1990s and beyond.

Thank you Michael for sharing your work.

In your above reply to Cesar Alvarez, you note: "I would not trade any of those systems because my analysis says that they have no intelligence but instead rely on a mean-reverting tendency of the market that may suddenly disappear."

Question: Do you have an SP500 L/S strategy that you do regard as viable? For instance a strategy that you regard as being more promising than those discussed in this blog post which you could share.

I am an end of day only SP500 L/S trader, focused on very next day or two direction.

Most end of day strategies that I have looked at are some variation of Larry Connors' mean reversion strategies. Just wondering if you know of something different, perhaps more promising.

Thank you.

Hello Jim,

The mean-reversion strategies based on RSI are actually a small subset of the strategies that are available. Most strategies are based on momentum and trends but they have a longer duration than what you are looking for.

Most quants that focus on the next day or two use NNs. However, success has been limited. My p-indicator was designed with that timeframe in mind and relies exclusively on parameter-less price patterns. It calculates directional probabilities and their significance. I must tell you however that the amount of noise in the last couple of years has been high in the S&P 500 and related ETFs but elsewhere things were better. Thanks.

Two day losing streak > RSI = Antifragile

Hi Tim,

Possibly, I'll have to think about that.

Michael