Course 9

Don’t build an algo that fits the past

What you will learn:

  • Don't spend too much time curve fitting your algo

  • Past market movement will rarely happen again

  • Make your algo transportable instead

How can I be sure that my algo will work in the future?

Well, I can’t.

Why? Simply because I can’t control the market. That’s why I can’t promise you anything with my algos. Past performance does not predict future results.

The only thing we can control are our actions. So, stop spending time back testing all the time, adjusting every single variable to make a back test result marvelous with high profits. Instead, start trading live with your algos to see how they perform on the real market.

When you back test and you utilize let’s say 2 variables, you should get very similar results even if you change slightly these variables. Why? Because maybe 20% of your back tested profit was produced by 1 trade, which happened randomly, and that may never happen again.

If you change each variable exactly to get the highest profit, I can insure you that you won’t get the same profit in the future, just because things are going to be different.

I used the example of 2 variables because this is to me the most variables you should have. If you start having too many variables, then your system is all over the place. It becomes not portable, meaning if you want to use it on another coin, you can’t.

A great algo is an algo that works on most assets, it has 1 or 2 variables, with a set of fixed rules, which do not change. I don’t mean to brag but my #1 algo does have rules, they never change, it has only 2 variables, and one of them actually does not change when I use the algo on different coins. This shows me that the theory behind it works.

I can backtest the algo, but I don’t sleep on it. I back test to make sure the overall result is up only, then I double check which value for the variable had the most luck on all periods, and I launch it. Easy.

Conclusion: do not have more than 2 variables on your algo and do not fit your algo to the past. Instead, look for value that provides similar/constant results.