Recently I have come across something that I couldn't grasp. In this chat (https://chatwithtraders.com/ep-103-dave-bergstrom/) the speaker talks about parameterless signals to avoid data mining bias and overfitting. However, I am having trouble grasping the concept of parameterless rules... Even a simple "buy if today's open is higher than yesterday's close" rule can be parameterized by adjusting the lookback period for example "buy if today's open is higher than last 2 days' high" , "... last 3 days' high". Can someone elaborate and if possible give an example for such signals/rules?
Parameterless Strategies
- Thread starter kojinakata
- Start date
Parameterless signals could indicate a number of separate signals combined using weightings that minimise the correlation between each individual signal. This allows the overall strategy to become 'parameterless' i.e. there are no individual parameters to tweek since the signal is combined thus avoiding overfitting and data mining bias. Essentially, you are fitting to minimise correlation between the individual signals, rather than fitting for performance.
Thanks for the answer. I am not sure I am able to understand your reply clearly but maybe you can tell whether the following strategy can be made a parameterless one:
There are 3 signals: 1 - MACD cross, 2 - SMA cross and 3 - EMA cross. How do I use weightings to minimise correlation between them? What exactly is my utility function that I am trying to minimise?
As you can see I am very confused, could you give a very basic example of a strategy to clarify?
I once tested that first parameter was coin flip regarding to which signals to take, heads to take and tails to pass, it just showed system made less money over five year time span.
If you using many rules to filter out trades and so many rules for entry and so many rules(filters) for exits, I am doubting whether there is a data mining bias. Really comes down to length of back testing that you are doing and sample size. You test day trading past 9 years, it going to have upward bias period as market has had this bias.
If you using many rules to filter out trades and so many rules for entry and so many rules(filters) for exits, I am doubting whether there is a data mining bias. Really comes down to length of back testing that you are doing and sample size. You test day trading past 9 years, it going to have upward bias period as market has had this bias.
runtrader's example is succint yet accurate.. but in that situation you'd have a backtest where stocks bought at the beginning of the backtest would still in your position 10-20 years later. If you say hold on for 10 years, or start the backtest x years ago , then that "P" word comes up again- kinda hard to avoid. In addition, when you are doing directional trading, Handle123's point is important. Tests after 2009 for long stox would show good results since stocks got beaten down 2008.
Here is an example of a " not quite" parameterless strategy BUT with a non-directional bias as brought up by Handle123. Buy a 20-24 day duration ATM straddle EVERYDAY and hold on to it till expiration in a low volatility environment. Do the same in a mid range volatility environment. Do the same in a high volatility environment. Whatever performance you get under that volatility backdrop is not tweaked, curve fitted,or walk-forwarded since it just buys daily without regard to techincal factors with the straddle ensuring no directional bias.
Here is an example of a " not quite" parameterless strategy BUT with a non-directional bias as brought up by Handle123. Buy a 20-24 day duration ATM straddle EVERYDAY and hold on to it till expiration in a low volatility environment. Do the same in a mid range volatility environment. Do the same in a high volatility environment. Whatever performance you get under that volatility backdrop is not tweaked, curve fitted,or walk-forwarded since it just buys daily without regard to techincal factors with the straddle ensuring no directional bias.
All three use parameters, thus couldn't fit in a "parameterless strategy".
