Once you have found your strategy, you will find that tuning the parameters of the strategy is not obvious. How do you choose the parameters so that the strategy is optimized?
There are two camps: Data Mining and Diversification. The earlier misgivings about data mining also holds here.
The general method for tuning a strategy for both data mining and diversification is like this:
You select a strategy
You select different parameter combinations for your strategy
You run the strategies with different parameter combinations on hundreds to thousands of simulated price data.
You select the strategy parameters with certain metrics. The usual choice is the best sharpe ratio and the smallest correlation with each other.
In the case of the diversifiation camp, you select multiple parameters instead of just one.
The difference between the two camps lies in the simulated price data.
The Data Mining camp might use a modeled price process for their simulated price data.
Whereas the diversification camp will not use a fitted process, instead they generated data from a simple process - for example they might use a sawtooth pattern generateor to indicate uptrends and downtrends.
For more info on this, see: https://www.youtube.com/watch?v=-aT55uRJI8Q
This is all pretty abstract, so let's give an example.
Suppose you want to test an EMA strategy which takes two timeframe parameters - one for lagging and one for leading. But you dont know which parameters to choose.
You limit your choices to (1, 2), (2,4), (4,8), (16,32), (32,64), (64,256).
After the fitting process, you conclude that the ones that correlates the least are (2,4) and (64,256) so you choose those parameters.
Quantifying Forecasts from Strategy
Congratulations on making your strategy! Now, to make the system work as a whole, we need to change our strategy's output into a standardized forecast value.
The reason for having a standardized forecast value is so that our system works for any combination of strategies with different instruments of different prices.
Below are characteristics of a forecast value:
A forecast should be within a reasonable range. Robert Carver suggests that we use the value -20 to +20. Thus, the output of our strategy needs to be multiplied by some scalar so that the expected absolute value of the strategy forecast is 10.
Forecasts should also be Volatility Standardized. This is so that our forecast values do not depend on volatility, and we can control the volatility of our portofolio later on in Volatility Targeting. This means that our forecast should be proportional to: .
Again, those are all too abstract. So we'll look at an example:
Suppose we have a bollinger band mean reverting strategy
We can convert the location of the price relative to the bollinger bands into Forecasts.
For example, +1 std can mean a forecast of 10, and +2 std can mean a forecast of +20. While -1 std can mean a forecast of -10, etc.
Dont worry if this forecast value seems arbitrary, it will make more sense in the subsequent chapters.