Yes, although I don't use neural networks but a similar, trainable binary logic network. If you're worried about lag don't use moving averages as they will always lag -- use a band pass filter of some sort. I previously used linear regression and am now changing to inverse fourier transforms as a band pass filter. I can select the wavelengths that I want to pass through and leave any noise behind.
BACKTESTING: only 25% winners on MA crossovers?
- Thread starter clambill
- Start date
Yes, although I don't use neural networks but a similar, trainable binary logic network. If you're worried about lag don't use moving averages as they will always lag -- use a band pass filter of some sort. I previously used linear regression and am now changing to inverse fourier transforms as a band pass filter. I can select the wavelengths that I want to pass through and leave any noise behind.
Quote from neye_eve:
It was unfortunately sobering :-( I wanted a free lunch with little effort
In any case, I'm going to keep doing backtesting on various systems to try to keep eliminating them so as to prove to myself that there's no magic bullet.
Next up will be EMA's instead of SMA's, and then I'll take a look at some of the more popular candlestick patterns.
The drawback to all of this testing is that it doesn't include volume, which is likely pretty darn important. As I go through this, I'll hopefully figure out a good way to incorporate that as well.
neye_eve, please do share your experience. Thanks!
If you use a basic MA cross with average-to-poor rules, you deserve to lose your trading capital in the short time it will take.
There are robust high frequency forex/FUT index systems, however, that utilize a weighted EMA cross for either entry/exit or both. The key is to properly incorporate market tempo filters, what snapshot to use (x time, x vol, x tick, etc), and constantly be adjusting for tempo. The naked cross itself pales in importance compared to the filters and execution rules that encapsulate it. I would strongly wager that most, including experienced traders, never dig deep enough and end up pre-maturely signing off CO development as a "newb" route. Some of the perceptions in this thread are indicative of that.
Whoever stated fading a 25% win rate cross equates to 75% win rate is an absolute idiot. Perhaps that was sarcasm I didn't pick up on?
There are robust high frequency forex/FUT index systems, however, that utilize a weighted EMA cross for either entry/exit or both. The key is to properly incorporate market tempo filters, what snapshot to use (x time, x vol, x tick, etc), and constantly be adjusting for tempo. The naked cross itself pales in importance compared to the filters and execution rules that encapsulate it. I would strongly wager that most, including experienced traders, never dig deep enough and end up pre-maturely signing off CO development as a "newb" route. Some of the perceptions in this thread are indicative of that.
Whoever stated fading a 25% win rate cross equates to 75% win rate is an absolute idiot. Perhaps that was sarcasm I didn't pick up on?
Try and test different modifiers to the strategy and see what it does to performance. For example, and entry modifier would be a certain minimum volume threshold. An exit modifier would be price not confirming a profitable crossover.
Quote from Gyles:
Thanks bigdavediode, for the answer. Is it possible to give some examples? Thanks!
Here's an Excel spreadsheet that can perform band pass filtering on any kind of data, and inverse fourier transforms to reproduce a close approximation to the original series.
Change some of the 1's to zeros to see the filtered output:
http://www.ecr.unimelb.edu.au/~xlr8tr/techs/Fourier.xls
Nice to see someone share something aside from cryptic circumlocution for once. Did you write it?
Although, there are a couple of problems.
For one, the dc (mean) is missing from the
filtered reconstructed data set.
You can just add a constant of the average of the input time series to
the Reconstructed data row to take care of this.
2) If you put in a real time series, it
doesn't work too well. I don't have time to troubleshoot at the moment (it's likely that the freq sample resolution is too limited) but just try any stock or real series as the input column and look at the filtered output.
Lastly, using fourier decomposition and inverse reconstruction typically has wraparound problems at the ends as well as limited bin resolution on financial time series data.
It's not a good way to filter financial
time series for a lot of other reasons.
Better to use a real bandpass filter for what you are attempting.