Just to recap this thread before it got derailed.
Earlier on there was a request by Nick Radge for Howard to look at TechTrader, seconded by tech/a. There has been a lot of talk over the past months about understanding 'why' a system works. I have felt there has been a lot of repetition of this message without any actual answers and lots of delivery of this paraphrased message from tech/a saying, "it's not that hard people, look at me, I'm just a builder and I managed to become a successful system designer, here's some bank details so you can all see how successful I am".
No-one actually critiqued TT with the seemingly elusive answers to this , WHY question?
So a little while ago, before this thread got onto the track that its now on, I asked the question: in this market ('96 -> '06), how lucky or unlucky can a trader actually be? The summary of the conclusion is:
not that unlucky...and providing you create some basic parameters for your trading, the probabilities of being unlucky drastically decrease.
My initial study created a system with the following parameters:
# Universe of shares, the current All Ords lists minus listed trusts, > 3 letter tickers etc. (I've since added about 100 delisted and non-XAO shares to try to counter what everyone keeps telling me is start-date bias, so I can't reproduce this study with my current list...
tech/a, would love to do it with the BT300 list!)
# 10 years of test data from 1/1/96 until 31/12/06.
# Each test was conducted with 2500 runs to raise statistical significance. More would have been better, but system execution is slow due to the way the Amibroker code had to be implemented and in any case relatively smooth distributions were still apparent with this small number.
# A randomised start date was generated for each run of between 21 and 60 days delay from the 1st of January 1996. This was done so as to reduce start date bias.
# Subsequent new trades were delayed by a random number of bars between 1 and 20. This was also to reduce the effects of having the system load up too rapidly, which would introduce another kind of start date bias.
# Each position was held for a random number of days between 20 and 120 bars (1 to 6 months, approximately).
# Maximum concurrent open positions were only limited by available equity.
# Position size was randomised to be between 5% and 50% of total equity. This was to simulate a punter who did not practice diversification.
# Commissions were included.
# No margin was used.
# The TT > $500k moneyflow filter was used.
# From memory the < $10 price filter was in there too.
The original question was: can a random entry/exit system outperform TT? The answer is, yes! There is an overlap in CAGR in the Monte Carlo distribution of TT and this system. But given the above set of parameters the overlap represents a relative handful of runs...regardless, there are certainly paths through the data which beat it (Coopers Pale is a top drop, so yes, gladly
).
Next phases of the testing added these components:
# Position size restricted to 10% of equity.
# As above, with a 10% stoploss.
# As above, with a ROC filter.
Position sizing increased CAGR, decreased Max DD and bunched both distributions. A 10% stoploss decreased CAGR, massively improved Max DD and substantially bunched the Max DD distribution. Adding the ROC filter increased CAGR again, reduced Max DD and bunched both even further.
This is a really, really simple system. Distribution histograms can be seen here:
http://theasxgorilla.blogspot.com/2007/08/2500.html
It can be argued that these distributions are more representative of the market conditions than a TradeSim Monte Carlo run on TT for the reasons Curtis Faith describes when talking about the limitations of Monte Carlo runs that breakup the sequence of big market events like the Asian crisis and 9/11.
So, to try to come full circle back to where this got derailed and answer the question WHY does TT work exactly as it has during the single run documented?
* Start date bias
* Leveraged exposure to a strong and protracted bull market
* Non-cap-weighted exposure to the market leading to index out-performance
* Price filter encourages taking positions in what are likely to be lower liquidity issues with inherently greater growth potential
* The entry trigger has an edge that improves entry efficiency (or whatever you want to call it). So fewer breakouts ought to be false breakouts...reducing the number of initial stops being hit.
* Discretionary entry trigger over-ride may have added some performance edge
Those points in bold are validated by the random entry/exit study, IMO. I haven't included a point about the exits here, for a few of reasons. One, based on the study I tend to agree with what Curtis Faith discovered when he tested a
time-based exit: simple exits are over-rated. Two, a 180-day EMA is not a particularly sophisticated exit, so I don't believe it provides much of an edge. Three, tech/a, you turned the system off yourself before trailing stops were enacted using this exit, which suggests to me that you don't even trust this part of the system.
Nizar, thanks for your wise question about statistical relevance. My advice, don't just ask this question when token data is used to disprove something, also ask when it's being used to prove something. Suckers buy black box systems this way too.
There is a lot of one sided thought around this subject on this forum for obvious reasons: those making the "for" points have a tendency to talk the most. I don't mind playing devils advocate and representing a more balanced view...even though in the background I could well be creating TechTrader3, aka. Revenge of the Long Term Long Only Stock Trend Following System.
In the meantime, it's hole-digging time...I wasn't joking you know:
esok:
ASX.G