This is the continuation of a discussion that started in the thread "Price Analysis with Neural Networks".
Please note that Altreva does not use GP in the conventional sense (learn more). With conventional GP it would indeed be easy to achieve very high returns on training data through optimization. However, these returns would be meaningless as the resulting trading rules would be overfitted to the historical training data and are likely to fail on future price data.
A system that does not optimize (overfit) on historical data but that evolves in a time-incremental way (like Altreva) will generally not achieve the spectacular returns on historical data often shown by optimization/overfitting but the returns that it does achieve are much more reliable and significant with respect to future price data.
Do you know any systems that show better returns without using optimization/overfitting on historical price data?
In the future I intend to provide more example models on the website.
Disliked...This also brings me to one more point: I've played around with a free adaptive (genetic) population modeler called Altreva, and it seems to have some features that might be very useful in our development efforts. One key principle it uses to prevent curve-fitting is that its models evolve only in a time-incremental fashion. That is, training and/or population selection is performed in real-time, as the quote series is fed through from beginning to end. Thus, what results is not only a system that performed well, but one that learned well over time. This suggests that such a system would be able to properly adapt to future shifts in market behavior, adapting as circumstances change to maintain profitability. I thus suggest that the learner resulting from this thread only learn in a time-incremental fashion, building knowledge as it progresses from the beginning of time in the market.
Has anyone ever used this type of approach in the application of neural networks to time series?Ignored
DislikedI discovered Altreva a while ago, but I didn't find the models to be all that impressive. The equity curve did look nice, and it appeared that the systems it produced were not curve-fit, but the profits were minuscule. The example model that they show for the S&P does not beat buy and hold.
Do you have any other example models you can share?Ignored
DislikedYou are probaby referring to the return over the last 10 years (which is shown by default in the sample model on the Performance tab) of the sample model that came with version 0.98. The compounded average return over all the 46 years (including the last 10) are close to 20%.
The new sample model that is on the website since the release of version 0.99 (on 12 March) shows a compounded avg annual return of 8.8% over the last 10 years (which is an excess annual return of 6.7%) and a 20.8% compounded avg annual return over the last 47 years. Also notice the smooth equity curve since 1999.Ignored
DislikedAltreva,
Perhaps you can show us a few more example models? I believe GP has the potential to produce models even more profitable than that.Ignored
A system that does not optimize (overfit) on historical data but that evolves in a time-incremental way (like Altreva) will generally not achieve the spectacular returns on historical data often shown by optimization/overfitting but the returns that it does achieve are much more reliable and significant with respect to future price data.
Do you know any systems that show better returns without using optimization/overfitting on historical price data?
In the future I intend to provide more example models on the website.