Keep it simple. The more maths you add, the lower the signal to noise ratio will be. NNs for example fit equations to noise. GPs are even worse and many poor souls are using these stupid tools because of the hype and lack of understanding of data mining bias. Any system found by such tools has less than 1 in a billion chances of being a true edge. Those who do not get that deserve to lose their money to other traders who can use it for a better cause. This is the purpose that trading serves ultimately.
I agree with NN and most non linear ML methods, but am currently trading live a system from a gp result. Doing quite well too, but offcourse I validated on 3x as much of OOS result from IS and verified the pattern again in R.
It's a fitted system. No question about it and any claim to the contrary is ludicrous. It's doing well because market conditions haven't changed. When they do the equity curve will go straight down. So I hope you know that and you'll be careful with your money.
Well, its backtested with over 5000 trades in 6 years, trade at least 10 rt per day now. It would be hard to curve fit that no ? backtest results show that market conditions have existed for 6 years at least, so im at least expecting another year or two before it degrades.
Pls no BS. You are obviously trading an intraday system because there are only 1500 days in 6 years so you are doing at least 3 trades per day and at that rate unless you are an MM commission and slippage will erode your capital fast. Only audited statements will provide evidence to the contrary. So, pls no BS.
yeah, comission is a big cost of my average trade profit, but backtest shows me profitable, last 2 mths live proves it. Surely other ppl here on et have had a similar system ? I was thinking if itz still stable for at least another 3 months i can go prop ? I can't rent out my system like dom or host on C2 because the latency requirements. Not LLT but at least under 300ms otherwise there will significant tick entry miss.
A little bit on curve fit in backtesting. Since a lot of people here shouts curve fit on every computer generated strategy. For any given financial time series containing OHLC data, If the main input to your optimization program is finding the best lookback period to use for Moving Averages, Boilinger or whatever indicators This will ofcourse will yield a solution set for the backtesting period where the NetProfit > 0 which will try to optimize the following equation MA(Short Period) > MA(Long Period) Then Buy Or Ma(Short Period) < MA Long Period) Then Sell Then Solve for Short Period and Long Period where NetProfit > 0 over a ts<t<te Then every time series will have surely a nonzero solution space. And adding a non linear function to the above, such as the sigmoid function used by NNs will probably allow you to quintiple your net profit. If you are trying to optimize or generate strat this way your doing it wrong. Which is admittenly what most off the shelf optimization and gp strat generation tool available for retail trader do. The correct way is to use input indicators that represents certain market conditions that is not dependent on the lookback period. But the positive correlation of the indicator to Net profit function is representative of certain market characteristics. Thats what I'm using my custom GP program for, to discover market characteristics that exist in differing asset classes.
We hear you. Please reply with just one "input indicator that represents certain market conditions that is not dependent on the lookback period" I would like to see that because from the other stuff you wrote there appears to be some serious mixup of concepts. For example you wrote The above is NOT the equation that is optimized. The optimization is as follows: max{Netprofit} (or whatever else you want like -DD or Sharpe ratio or a combination of metrics) wrt short period, long period subject to the constraint that signals are generated by {MA(Short Period) > MA(Long Period) Then Buy} (plus any other conditions) This is the correct optimization statement.