Transcription
You have to write down the rules yourself. You have to program the rules yourself. And you have to test the rules yourself. And you have to test them every single which way you can think of. Try to break them in the nastiest way you can think of. And just keep trying to break them until you really throw your hands up and you say, "I can't take this anymore. I can't break them." That's what you really have to do.
One easy way of trying to break trading rules is to add noise to your inputs. Essentially, let's say that you've got three pieces of data. You're going to take the difference between the Fed funds rate and the 10-year Treasury. You're going to take the current trading volume in the S&P 500. And you're going to take the distance from the 50-day moving average. Just making this up just on the spot. All of those require inputs.
Take the inputs, write a little program that adds noise to each day, random noise from some reasonable distribution. Now, you've got a data series that has noise added to it. Now, run that through your system. What you should find is that your returns with small amounts of noise, it should be unaffected. And with large amounts of noise, it should start to degrade. And what you should see is that there's a curve. It degrades as you add more and more noise to the system. That's what you really want to see.
What you don't want to see, and this is what happens in most systematic [music] systems, is something like this, where some amounts of noise produce a good result and other amounts of noise produce a bad result. That doesn't fly.
>> Why would that happen?
>> From the fact that you your original system is [music] not real. You fitted noise. You didn't fit data.