ForumSections

Machine learning in trading: theory, models, practice and algo-trading - page 3749

New comment
Ivan Butko #:

That’s not what I’m talking about: a forward is a form of exploratory peeking. You compare it once – that’s it. That’s already the dirty phase.

I’ve also given up on the forward strategy. It’s better to include it in a train and select a good combination of train and forward. Well, in reality, it all depends on how you’ve chosen it (taking into account underfitting and overfitting) and how lucky you are with market volatility...
Ivan Butko #:
A forward is a form of exploratory peeking. You compare it once – that’s it.

A failed forward test invalidates the entire process, not just a set of parameters.

In other words, if the forward test fails, absolutely everything goes in the bin – the cycle and the results of the optimisations/runs, the expert advisor, the trading strategy and the signals. EVERYTHING goes in the bin

If you don’t chuck it all out in one go, you end up with feedback looping back through yourself

Aleksei Kuznetsov #:
I’ve also decided against using a forward. It’s better to include it in the training set and select a good model using the training set plus the forward. Well, in reality, it all depends on the choice made (taking into account underfitting and overfitting) and how lucky you are with market volatility…
The market will be different; I completely agree. But will the patterns be different too? If you take a random forest, the error stabilises after 50 trees, which are the patterns. In other words, the patterns are stable, but it turns out that the relationship between the predictors and the target is unstable. The patterns remain, but their meaning is different. That is why we say ‘the market is different’. In machine learning, the meaning of these terms begins to be broken down into layers and ultimately boils down to the probability of a pattern existing in the future. At each step, we obtain a quantitative assessment that indicates the likelihood of achieving an acceptable final result.
СанСаныч Фоменко #:
In other words, the patterns are stable
A cube also has only 6 faces. The stability of the patterns is in no way related to the order in which they appear.
Ivan Butko #:
A cube also has just 6 faces. The stability of the patterns has nothing to do with the order in which they appear.
Most likely not. The question should be phrased as follows: how long will the patterns last?
СанСаныч Фоменко #:
The question should be phrased as follows: how long will these patterns last?

Or rather, how quickly can you identify when they cease to exist?

Unless you investigate why they exist in the first place, it is impossible to understand why they might cease to exist; consequently, it is impossible to predict or determine in advance the approximate time of their demise or the clear signs of it, and therefore it is impossible to halt trading in a well-founded manner in good time.

СанСаныч Фоменко #:
Most likely not. The question should be phrased as follows: how long will these patterns last?
There are two positions:

1) A pattern is unshakeable, and specific behaviour is expected following it

2) The predictive function of the pattern is revised with every tick / candle / mini-pattern / other candle aggregations

In theory, once a putative pattern has been identified, the situation changes with every tick. In other words, its predictive power is, hypothetically, subject to revision. This means that training the system to recognise patterns based solely on historical data is insufficient. A more complex mechanism is required than that proposed by the articles here, which focus on static patterns.

A concept closely related to this is RL (Reinforcement Learning) and agents. But as a rule, the policy and other aspects there aren’t quite what’s needed. They still focus on the static nature of patterns.
Ivan Butko #:

That’s not what I’m talking about: a forward is a form of exploratory peeking. You compare it once – that’s it. That’s already the dirty phase.



Forwards aren’t great for currencies, but they’re fine for metals and shares. There’s no volatility in currencies; it’s purely a return to the mean. It’s not that simple in the market – gold is top

1...37423743374437453746374737483749
New comment