Sure, I meant the (non-)stationarity of the return distribution. http://en.wikipedia.org/wiki/Stationary_process Ninna
frost, I'm still curious about this. I attached a pic showing what I'm talking about. I'm curious about this because it's a bias I've come across when using traditional backtesting software to do the type of analysis you are doing. I haven't tested using NinjaTrader, though. The first chart shows 5 trades with 20% winners. The second shows 2 trades with 50% winners. Most software will only kick out case 2, and you might put RSI(3) < 30 into a set of building blocks thinking that the condition generates 50% wins when it actually has a win rate of 20%. In any case, I'm enjoying following this work that you're doing.
The problem with RSI, is the same as the problem with stochastics. both these indicators compress as they approach their maximums. Both indicators can oscillate between 80 and 100 for quite some time while the trend surges on. So selling at overbought and buying at oversold is a dangerous practice. Ranging markets is what these oscillators do best.
RSI surely works better in range markets. What I meant however is that RSI performs better on a gaussian process than on market prices. The challenge isn't on optimizing RSI for specific market conditions but on normalizing market prices. Ninna
Actually one of the better ways I have found to trade stochastics (and most oscillators) is to buy on the first crossing into overbought ( > 80 stochastics, > 65 RSI ) and expect a trend move to carry 1-3 bars into the future. This works well for momentum moves and is usually only a breakeven trade or small loss on failures. The standard (ie idiot) way to trade these is to sell/short immediately upon the first reaching of an overbought reading and wait for a downtrend to set in. Any strong trend will decimate you using that approach.
By normalizing market prices, do you mean something like detrending the price? For example creating a series that is difference of the price and a moving average? I have heard of the trend adjusted stochastic which I think incorporates this into the indicator. I never covered these topics when I took statistics, or perhaps I forgot, but in looking up the definition of a gaussian process, it appears that it essentially describes something with a normal distribution. Since stock market returns do not have a normal distribution I don't know how you would convert them into something that does. I don't think simply detrending would do this. Perhaps that is one of your trade secrets, in which case I understand if you don't want to elaborate for those of us that are statistically challenged.
1. A 10 Day RSI with a two months of historical data will be exactly the same as a 10 day RSI with 20 years of historical data. 2. The ability to translate price to a 0-100 scale and have momentum comparisons not identical to price movements lies the true value of RSI. When are they different? How is that important? Good Luck. 3. OP is on a good path. Stay the course. You need to keep digging deeper. When you find it, you won't be publishing it here.