Quote from traderkay:
prophet, how did you find the patterns? just by watching day in and day out or by doing some sort of hardcore datamining exercise?
Both, but by examining historical charts, not watching price in real time since I'm better at coding and testing than watching price in real time.
I take historical T&S, design a simple indicator or buy/sell signal, then overlay that onto a few historical daily charts, see how the indicator responds, see when it works and when it doesn't, check for overtrading or thrashing.
I've concentrated on indicators and signals that perform well (or not too poorly) while paper trading without periods of neutrality, stops or profit targets. The indicator/signal must always indicate long or short (always try to forecast the market), and always reverse using market orders. Ironically, while this is a much harder test for an indicator to pass, it is faster to test numerically because you donât have to brute force test indicators/signals combined with all flavors of stops or secondary rules. I can spend more time mining T&S for undiscovered, yet robust patterns instead of testing countless stop-loss or secondary rule combinations that are only a crutch. Along the same lines, using linear least squares networks (with nonlinear input stages) can be superior to neural nets because they train and converge faster by a few orders of magnitude.
Once I have some promising indicators I do some automated backtesting to characterize the indicators and look for ways to control risk and forecast near-term profitability for each indicator/signal. This is equivalent to trying to determine the appropriate size to trade (or whether to trade at all) for each indicator/signal.
This seems to be a successful approach. I have found about 10 broad types of indicators (not all derived from T&S) and about 7 unique system types that simulate at between 2 to 6.5 points/day on NQ and ES for the last half year of market history. Iâm reaching the point now where I have more indicators and systems testing positive than time and CPU resources to test them. I have 3 dedicated servers for data logging and real-time systems, but only 2 PCs for testing since my code is fairly efficient (Matlab and C++). I need to add more servers, more IB accounts, plus automated testing and signal distribution code. Lots of CPU power isnât the answer.
Things didnât always go so well for me. I spent my first 2 years designing systems and only came up with crap⦠almost quit several times. Iâm no genius at this stuff⦠just persistent and willing to test anything and abandon previous notions when necessary. Reading Elitetrader (current and older posts) has helped tremendously. Thanks everyone!
