If you have read my previous posts, there won’t be a lot of new info in this post. I wrote this as a personal record for myself, in case my approach change again in the future.
A little background story first, since I don’t have written record for my previous approach. I started this algo trading journey on June 2021. As with most of other traders, I started by playing around with indicators. It felt great when your backtest result shows 100,000% over a year. And to have such result is (relatively) easy. We are playing with known data. All we need is to find the “best” parameters.
Long story short, such approach won’t produce a strategy that can survive long-term while having (relatively) big profit, which I summed up as the triangle of trading. Freqai (machine learning) isn’t really helpful as well, since from my experience, it will result in lesser entries and premature exits (which mostly were losses that would be winners if the strategy didn’t exit prematurely).
Someone else who have spent 4 years working on RL (Reinforcement learning) strategy summed up his experience in one sentence
If you know what factor/feature drives the market, you don’t need any machine learning algo
That’s the main issue with my previous approach. My strategies couldn’t accurately guessed what drive the market, even after 3 years. My personal theory of why it won’t ever be able to do so can be read here. In short, all my strategies tried to do were testing their luck and see which one survived longer before being crushed by the randomness of market. After 3 years playing the backtest games, finally I burned out and stopped doing it. But then which approach should I use to replace my previous approach?
The next chapter of my approach came after someone mentioned “Fooled by Randomness” book written by Nassim Nicholas Taleb. Please do read them if you have the time. In short, most of the time, strategies made profits due to luck, but we thought that the strategy is the real deal. There are a lot that I got from that book, but one of the points that is relevant for this post is how I should understand the pair(s) that the strategy gonna trade. It aligned with what that user said above, which is I need to know what drives the market.
While thinking of what the approach will be, I opened reddit after a long time for some inspiration, and I found the one post that inspired me. The light bulb moment came at that time. Instead of having complicated strategy with many indicators, the current approach focus more on pairs selection. So here are the steps I do to develop a strategy
- Pick which side do I want, long or short. Choose only one.
- Come up with the pairs. It can be static or dynamic, but mostly I do dynamic. If it’s dynamic, needs to figure out how to get the best pairs (in theory) that are best suited for the choice you chose at #1. The pairs can either be generated using FT’s existing pairlists, or by using external script, or combine them.
- Start a dry run with a “blank” strategy that always enter trades. Since I believe the pairs will go in the direction I assumed (point #1), I don’t use stoploss and I have adjust position logic.
- Analyze the trades, especially the losing trades. See which indicators (if any) might helps. Doesn’t need to be indicator derived from candle data. You can use any other external data that you can get, and they don’t need to be backtest-able, for example current ticker funding rate, long/short ratio, or others.
- Reset the dry run with modified strategy. Analyze it again, reset dry run again. Rinse and repeat.
- If after several attempts bring no good result, then maybe the pair or the side (point #1) is wrong. Start over.
The approach is very fluid and there is no real boundary on it, which makes it exponentially harder than my previous approach, so the number of strategies produced will be a lot less. Previously it’s easy to create one working strategy per week. With current approach, it’s considered good if I can produce 1 working strategy per month.
So far it’s fun to come up with scripts to get specific pairs. We will see when will I change this current approach with a new one.
[…] 20260426 – Current strategy approach […]