DeMark ML (MIMIC model) 2.0

The Problem of Model Overfitting​

or Robots don't understand us..​

​

What happens when machine learning ends up in inexperienced hands?​

It leads to garbage generation.
It results in hundreds of “cool” expert advisors that show perfect backtests — but quickly turn into scams in real trading.

Creating an overfitted scam model — very easy — 3 min
Creating a stable model — very hard — ?

There are many issues along this path. One of the biggest, in my opinion, is that we delegate too much responsibility to the model itself.

We tell the model:

Here's a bunch of data, labels, features, weights — go ahead and decide for me when to buy and when to sell.

Click to expand...

The model obediently takes all of it and starts searching for patterns. Sounds logical, right?
But how it finds them — remains a mystery.
We lose control over the process, and end up with a beautiful backtest and a total failure in live trading.

Of course, hidden patterns in the market do exist.
But if attempts to let the model discover them consistently fail, while simple linear strategies next door keep working (not always profitable, but stable), then a question arises:

Why not teach the model to understand our indicators the way we do?
Why not interpret features not the way it's “convenient” for the machine, but the way a human would?

Click to expand...

P.S.
If you think that simply showing a model two moving averages — MA(10) and MA(100) — will make it trade on their crossover...
You're mistaken.

​

MIMIC model​

In a specially controlled environment, the model is trained on ideal, pre-structured data. Over and over, it is shown slightly varied patterns that clearly reveal the core logic of the feature being used for training. The model is allowed to interpret the feature — but only within strict boundaries.

The goal of such a model is to mimic the behavior of the feature and become an expert in understanding it — not the market as a whole, but one specific feature (or a combination of them).

Important:

This type of model is not a standalone trading unit — it simply answers the question:

What is the probability of SELL or BUY within the context of its assigned feature(s)?

Click to expand...

Since the model never sees the actual market time series, it is not subject to classical overfitting.

If the feature (indicator) is too simple and the training conditions are overly basic, the model will usually just replicate the indicator’s behavior without adding any real “intelligence.”
This can be either a strength or a limitation — depending on what you’re aiming to achieve.