Quote from bwolinsky:
This is done by curve fitting. Not the same as having a working method for every symbol. Optimization for methods that are consistent won't produce out of sample results different or significantly different than the results you backtested on.
10x better? Hardly, maybe if you start with a 0.5% apr and make it a 5% APR. Even then, still with 20-50% drawdowns, so those programs don't work. I've seen the neural networks look at candlestick bars and all they are doing is curve fitting.
If you have a strategy that can be applied the same way every time, your optimization is not a curve fit, but simply by optimizing does not make it a curve fit if you have enough evidence that it has statistically significant results and produces robust results.
A 1 year backtest that makes 80% and drawsdown 10% is a curve fit due to its short term data. Adding 1 year will likely negate whatever edge it might have found, but working with at least 2 years should solve that problem. I'm not opposed to using all of the data. I think that just limits how robust the system is, and if you find your system needs to be optimized regularly then optimize regularly. There's nothing wrong with that. You always want to have the best edge possible, but certainly if you take it too far you'll get results on a walk forward that aren't what you'd expected.
People with no experience writing their first system have to sort a lot of ideas and once you've been able to classify system types eventually you come to a realization about what works and what doesn't, but it's a lot of coding and backtesting.
20 lines? Lol. Every good system I've seen with 12 point font is at least 20 pages long. That's around 500 lines of code.
