Whoa... sounds like Wilmott. Let me simplify - if you really want intraday RTM, you'd better be trading highly correlated spreads. Why not just use the Beta scatterplot functionality on the Bloomberg and be done with it. Jeez... you guys are making this way more complicated than it needs to be. FedEx vs. UPS, Conoco-Philips vs. Chevron, WalMart vs. Target, ad nauseum. Just steer clear of earning season. Nice phD thesis, though. (and I'm an Engineer by training)
Pairs trading is a clever way to transform an otherwise non-stationary process into say weak-stationary but it too is exposed to volatility shocks that happen extremely quickly. For example, financial stocks made great pairs trade for a number of years. And they still do offer good opportunities, but it became exposed to tremendous risk very quickly. Just ask those trading BSC and LEH pairs (before Oct. 08 very good candidates for pairs) Unfortunately, and this is my opinion, the tests that you mention require data to perform, and by the time there is sufficient sample size, your already in a major DD. At the end of the day your stuck with some sort of understanding of human psychology/behavior theory and how markets operate and the participants so you can try to side-step these inevitable periods. So no holy-grail in time series analysis yet! I have read some interesting ideas on neural nets for forecasting regime shifts....I haven't really researched enough to make any conclusions. Perhaps you have some thoughts? As per your suggestions for the chart, I'm feeling a little lazy but I'll get around to it since I'm interested in what you have to say. In the meantime I'll wax philosophical in the NYC thread.
How are you handling situations where price doesn't go in your favor? Do you have a set stoploss or are you averaging down?
What, I think, you are all missing is the Galton's principle of derivative reversions demonstrated in his QUINCUNX device. It is the most important principle in constricting a usable RTM method. http://www.bun.kyoto-u.ac.jp/~suchii/quinc.html
For those interested in a very good and intuitive read on many of the questions posted; such as, what is the optimal coefficient for out of sample exponential smoothing constant prediction, why are splines better than moving averages, etc... PM me. Warning, it's more of an academic approach on par with some of the thinking that has surfaced here, not your typical TA ...
This article makes one wrong assumption. It considers constant mean to determine the deviation as an input for their predictor. Flexibility of the mean impacts prediction significantly. Any practical RTM method should be based on a stable relationship between the mean deviation and the price deviation to that mean. If the majority or runs ensure that the mean deviation is smaller than the price-to-mean deviation than the algorithm has positive expectations. if it is not than the method is under the danger of a significant ruin.
I like this thread. I don't trade at a frequency where much, if anything, is normal. I'm trading sub-minute strategies in very liquid instruments and sub-second liquidity providing strategies. This is a plot of the distribution of some normalized[1] data from that time-frame. [Attached] If you go into R and run the shapiro.test() on the 5k most recent values, you get: Shapiro-Wilk normality test data: [filtered] W = 0.8124, p-value < 2.2e-16 I'm using non-parametric procedures to work on results from this particular distribution. In this particular strategy, I'm looking to collect a few cents per trade -- removing liquidity on entry. Where I used to work, they used to move out of these time-frames just to get the normalized result. They were trading with a latency of around 40-80ms, with worse at higher rates. [1] Normalized in a generic sense, not referring to a normal distribution.