I was wondering what others think about parameter insensitive models. These are a set of trading systems that are exploiting the exact same type of anomaly but with varying parameter. For example, trading different versions of RSI2. The idea is to robustly extract profits from an anomaly while avoiding the downside of single parameter set selection. Any thoughts?
Here are a few charts I've put together, but this single experiment may not be very conclusive.
Model1 (rsi1): RSI(2) 50/50
Model2 (rsi2): RSI(2) Buy: <30 Short: >70
Model3 (rsi3): RSI(2) Buy: <30 Sell: >50 Short: >70 Cover: <50
Model4 (no.reb): equal weight but no rebalance
Model5 (reb): equal weight rebalance weekly
Multi-Model
Multi-Model
- Attachments
-
- horizon.png (11.9 KiB) Viewed 4629 times
-
- dd.png (11.55 KiB) Viewed 4629 times
-
- equity.png (10.02 KiB) Viewed 4631 times
-
- Untitled.png (5.03 KiB) Viewed 4631 times
This may be useful: Trading a large number of systems without a large account
FWIW, I thought i'd post a superficial analysis to the idea of voting...
I've taken 3 correlated strategies (mean reverting in nature) (signal correlation attached n=60), and tested the performance of equal weighing each strategy verses voting.
The voting procedure was to take trades if 2 or 3 out of 3 strategies provided long or short signals. Equal weight with 33% allocation to each strategy and rebalancing monthly.
Note: results do not include commission + slippage, simulated using daily return of SPY
Edit: one burning question on the back of my mind is how would one go about formulating strategies given a bunch of entry, exit, and money management components. It's probably confidential, but fall river seem to employ computers to search up trading systems. Anyone have any ideas how this works? I'd love to write an algo and test out some truly uncorrelated models. Thanks
I've taken 3 correlated strategies (mean reverting in nature) (signal correlation attached n=60), and tested the performance of equal weighing each strategy verses voting.
The voting procedure was to take trades if 2 or 3 out of 3 strategies provided long or short signals. Equal weight with 33% allocation to each strategy and rebalancing monthly.
Note: results do not include commission + slippage, simulated using daily return of SPY
Edit: one burning question on the back of my mind is how would one go about formulating strategies given a bunch of entry, exit, and money management components. It's probably confidential, but fall river seem to employ computers to search up trading systems. Anyone have any ideas how this works? I'd love to write an algo and test out some truly uncorrelated models. Thanks
- Attachments
-
- equity1.png (11.86 KiB) Viewed 4174 times
-
- dd1.png (13.78 KiB) Viewed 4174 times
-
- signal-corr.png (34.22 KiB) Viewed 4178 times
-
- rolling250day.png (11.97 KiB) Viewed 4176 times
-
- summary.png (5.24 KiB) Viewed 4176 times