In 2022, optimizing with MT5 was trending

In 2023, the boom of optimization seems to have faded as over-optimization became a barrier

I am one of those who were inevitably knocked down by the wall of over-optimization

I believe EA has two parts: the logic part and the system part

There is a rule: do not optimize the logic part, but optimization is OK for the system part

The system is not about things like martingale multipliers or averaging down rules, it’s about the entry point

Well, strictly speaking... it can cause curve fitting, you know lol

Putting optimization aside for now,

these days I’m focusing on learning

Overview of the learning curve

Optimization and learning look similar at a glance, but there is a big difference

The following image shows a certain deep learning model’sloss curve(left) andaccuracy curve(right)

The loss curve has training (blue) and validation (orange) curves

The vertical axis is loss (like points deducted in a test), and the horizontal axis is the number of learning iterations (like class sessions)

As learning iterations increase, loss decreases, but validation stops decreasing midway

No matter how many times you attend classes, you cannot get 100 points on the test

In EA, optimization is like training, and forward pass is like validation

The accuracy curve similarly has training (blue) and validation (orange) curves

Instead of loss being a score, accuracy feels like the probability of passing

Loss is quantitative, accuracy is qualitative

Similarities between optimization and learning

If you continue learning, loss can even increase

This is called overfitting

Similarly, if you fine-tune parameters too much during optimization, you may fail to fit future data

Therefore, besides the optimization period, a forward test section is prepared

If you overfit the forward test section, you start fitting to the forward test data instead

The concept of over-optimization and over-learning is shown below

The plot carries noise; this is market noise

Noise means price movements that are not reproducible

By capturing this noise, you gradually deviate from the proper curve

It’s necessary to reduce the number of plots and create a moderately accurate curve

Differences between optimization and learning

From the above, optimization makes it hard to determine the actual curve location and evaluate it, but

learning can plot the loss curve and stop at an appropriate point

By avoiding improvement from noise, it becomes easier to handle unseen data

The ability to handle unseen data is called generalization performance

Often people talk about the EA expiration date, but if you can evaluate this generalization performance

you can lengthen the EA’s shelf life

Moreover, the more items to optimize, the more likely curve fitting becomes, but

in learning, more features allow learning finer market nuances

Concrete learning example

When creating a typical EA, indicators are usually about 10 types at most

Also, shifts of 1, 2, or 3 are common

Indicator 10 types × shift 3 = 30 items

Even at most this much, optimizing everything would quickly lead to over-optimization

Right now, I have about 300 types of indicators as features, and shifts from 1 to 200...

Using around 60,000 data points to predict the market

Collecting data is challenging

I aim to solve optimization challenges with machine learning while building a strong EA