Exactly. But it doesn't have to be anything more than 1. noting relevant metrics/criterion that exist when you make your human call, into a data structure. 2. And also noting that call you made, into that same data structure. That was training. The more examples you train, the better. Now, when it's time for the algo to make a call, it should compare the current set of metrics/criterion to those stored in the data structure, and pick the call that's matched with the closest matching metrics. You can even add weighting to each criteria. Then it becomes a simple, one node, neural network ... a perceptron. (And once you get the understanding, you could add more nodes ... just don't cause it to overfit ) Just some thoughts of mine. Maybe they're useful to you now, or later. Maybe not.
If a part of your method involves fundamental or judgement, it is very hard to code those. Often the challenge is how to quantify/assign value to fundamentals and judgements and convert them into tradable signals.
That's where machine learning comes in. Suppose, for example, a trading savant trades exceptionally well, but doesn't know why he "feels" he should buy or sell when he does. He simply looks at the chart and knows. So, you give a ML algo all of the charts where Hero said buy; and all of the charts where he said sell. Then you let the algo find out what all the buy charts have in common; and what all the sell charts have in common, via training. The more training--the better. Now, the algo can emulate the savant. How well this occurs ... depends. So yeah, what you post is exactly one of the problems ML is there to fix.