Well, because they don't know how to properly back test data. ACD is much more quantitative then technical. It's less about looking at pretty charts and more about looking at data.
So do the algos that worked last year get dusted off and work again where risk is USD.JPY and if so what happens at the 125 level.
With that in mind I am just about to go all quantitative with my data In my last post I assumed the data to be stationary - making assumptions in this game is a flawed concept. I have a feeling that the data might not be stationary but I am going to test it using a few splits over the coming week. I then plan to go down the normal route of testing independence, randomness, serial/auto correlation etc etc. It is in these kind of studies where the power of "R" really shines through, there really is nothing better. Clifford Sherry´s "The Mathematics of Technical Analysis" goes over these concepts an more. I also managed to find this paper that's worth a read for anyone who is interested.
Number lines are for the most part stationary but like all time series data, the lookback period is important. Over very short time periods, the data may not be stationary but in the longer term number lines always swing back and forth above and below which makes them very robust for time series forecasting. And yes, R is a great tool for ACD data. I'm trying to utilize it more and more.
Mav have you tried just testing the raw day scores for stationarity? I have heard of many people using differencing on time series to de-trend before testing for stationarity. The actually scoring of days should de-trend the data for us and then I would like to test just the raw scores for being stationary or not.
I haven't run a Dickey-Fuller root test yet on the data but will. Just from visual inspection I can see the data go back and forth above and below zero as it should which shows it exhibits some form of stationarity. Numberlines are by defintion bounded by time right, they can get to extreme values but they can only go so far because they start dropping off so there is a natural bound built in which means you can use a stochastic process to model it. But I'm definitely interested in hearing what you find as I have not run the tests yet. So don't let me stop you!
Honestly, I didnt find anything in that article useful. He's basically saying "the rules work until they dont." It sounds to me like he was trying to get more published/famed than alluding to anything useful.
Agreed its really just a dumbed down explanation of what stationarity is. I linked it just in case anyone was interested in what I was waffling about @ Mav I believe that the URCA package in R has a Dickey-Fuller root test. When I get some time I will run the raw daily scores through it and see what we get.
Since it was so simple to do, see the attached .txt files one for raw scores and one for 30 day. Start dates were 23/10/2015 for 30 day and 14/09/2015 for the raw. The results seem to agree with what you were saying Mav. I will run more tests when I am less tired
I didn't know you could do so many statistical tests on just the NL...makes me feel like i'm really not up to everyone's level yet.