Taking Neural Networks to the next level - page 15

 

Chris70:

Fun fact: any discretionary trader uses neural networks - they're also called 'brains' ;-)   ... they just don't seem to work on autopilot if we're not at the screens

The difference between Artificial Neural Networks and the average Human Brain, is that a human Brain has roughly 100 billion neurons and each neuron in the Brain is like a small independent computer. Unless you have huge computing power at your disposal, I think your comparison doesn't really work. I'd also like to believe that there's a huge difference between a live conscious mind and a machine that's trying to mimic some basic functions of a human brain. I might be wrong, though, and we are all just machines that were programmed to "think" that they are alive and/or have "free will".

 

Chris70:

... given identical input data, a neural network will always beat any technical indicator. It has to! It's by definition forced to do so! Because it will self-improve until it does.

And there's your problem. Artificial Neural Networks have to be trained to give you the expected result, which means that you have to know the result before you can train your Neural Network. But since you don't have the answer to "what will happen next", you can NOT train your ANN to tell you. Even though ANNs are good at finding similarities and categorizing large amounts of unfiltered data, they can't see into the future. Because of this, there is absolutely no way of telling whether the answer you've just received from your ANN is any better than a simple dice roll.

 

"I think your comparison doesn't really work."

Of course it was a joke. But if we stay in the picture: there are many task where no computer in the near future could possibly beat a human brain. But there are also many examples for the opposite. Even the little calculator program in my smartwatch can take better square roots than me.

You're not exactly correct saying that any biological neuron is a computer by itself. I should mention that my learned profession is being a physician in internal medicine, so I guess I have some understanding of brain and cell physiology. If we only look at specific neuron functions and not other cell complexity like any 'stupid' cell like e.g. a fibrocyte (connective tissue cell) has: the similarities with artificial neural network cells, escpecially memory cells, are astonishing. The real power of a brain is not in the neurons, but in the connections (axons / dendrites / synapses). The individual neuron is rather "stupid": it's just about a membrane potential that depolarizes once it has been excited beyond a certain threshold. That's all the "intelligence" a neuron has. The excitation and depolarization of a heart muscle cell by the way works based on exactly the same principles. The reason why the heart muscle - compared to a brain - can't solve equations or win the Nobel Prize is just that it lacks the extremly complex connection network (it has only cell-to-cell connections with direct neighbors).

Furthermore, brain complexity and brain size don't automatically translate to great analytical power, as show many examples from the animal kingdom that would be rather bad traders. Most parts of the brain serve motor functions and processing of sensory inputs (eyes, ears, equilibrium...). If we substract all that and and only keep whats necessary for analytical thinking, there's not much left of those ~100 billion neurons: mainly just a tiny fraction of what's called association cortex, so there's good reason to be humble about our 100 billion neurons.

Still, the human brain is more fascinating than any computer will ever be...

...and it was a joke.

------------

"since you don't have the answer to "what will happen next" ...

True, but this is no argument for or against neural networks. Technical indicators also don't KNOW what will happen next, and still they make the assumption (just like a neural network does), or at least we arbitrarily interpret the indicator in a way to make this assumption.

Training based on future time-steps that obviously aren't available in live trading is just the method for FINDING the "formula" and no longer a requirement once a good weight matrix is found: once having completed training, they don't need secret information about future timesteps no more than any technical indicator. This doesn't mean that the formula itself is automatically(!) any better or worse than a technical indicator. But I gave some arguments to believe that they indeed give the superior formula.

Infact, it totally makes sense to validate any formula against what it is trying to tell, in this case a prediction into the near future.

With indicators, you make just the same predictions. The difference is that an indicator is just the formula (however correct it is) without the validation part.

 

Chris70:

With indicators, you make just the same predictions. The difference is that an indicator is just the formula (however correct it is) without the validation part.

That is why my most important instrument when trading is my Brain. The indicators are there to give me a bigger picture of what is/was happening on the market, but it's just visual aid, nothing more. The difference between an indicator with a fixed formula and a Neural Network that constantly tries to find the "best" formula, is that I can rely on the output of an "indicator" with a fixed formula (given the same input, it will always give me the same output), but I can't rely on the output of something where the formula keeps changing.

 

"Even though ANNs are good at finding similarities and categorizing large amounts of unfiltered data, they can't see into the future."

Incorrect.

What about examples as simple as a network that predicts the next values of a sine wave, with extreme accuracy, after only having been trained based on future timesteps (and never having been told how a sine function is calculated)?

What about weather forecasting?

What about natural language processing? When you enter a text (WhatsApp, Google Search...) and receive suggestions for the next word, this is often based on machine learning and not just taken from some lookup table about statistical occurences.

These are just a few "proof of concept" examples that it is indeed possible to project into the future, only based on supervised learning using future timesteps during the training phase.

But if so, then why are we so much worse with predicting the markets then predicting if it will rain today?

The problem is that it is impossible to make predictions if there's absolutely no correlation between past and future data. It's about the DATA! Not some general flaw of the method itself. Bad data lead to bad results however perfect the applied formula is.

But this is true for indicators just as well. If past data are 100% irrelevant, we shouldn't use indicators, too - with the same arguments as for ANN's.

So the question is if the input data are bad (for both ANNs and indicators), not if neural networks are bad. There's just no way to predict anything based on something that contains no relevant information. This is true for ANY method.

But if nevertheless we want to find the scarce relevant information within bad data (because they're all we have), there are good logical arguments why a neural network is better for the job anytime.

I see you don't believe me. That's okay!

 
Chris70:

The problem is that it is impossible to make predictions if there's absolutely no correlation between past and future data. It's about the DATA! Not some general flaw of the method itself. Bad data lead to bad results however perfect the applied formula is.

Things where the pattern keeps repeating are predictable. But there is no repeating pattern on the markets. Take a 100-bar snapshot from any point in the History and try to find the exact match in the past or the future. You won't find it, because - what happened in the past isn't going to happen the exact same way in the future. And that is why everyone will tell you that the results from the past should NOT be taken as an indication for the future. Because the past is the past and the future hasn't happened yet. If you believe that you can predict the future, then good luck. If you succeed, you will be the first.
 
NELODI:
Things where the pattern keeps repeating are predictable. But there is no repeating pattern on the markets. Take a 100-bar snapshot from any point in the History and try to find the exact match in the past or the future. You won't find it, because - what happened in the past isn't going to happen the exact same way in the future. And that is why everyone will tell you that the results from the past should NOT be taken as an indication for the future. Because the past is the past and the future hasn't happened yet. If you believe that you can predict the future, then good luck. If you succeed, you will be the first.

I interrupt you right away, NN applied to trading has already proved their efficiency, numbers of traders use such systems including me.

 
Icham Aidibe:

I interrupt you right away, NN applied to trading has already proved their efficiency, numbers of traders use such systems including me.

Every system works - until it doesn't. Even a very simple strategy can be profitable, if the market conditions are right. What I highly doubt, is that a Neural Network will be any better on the long run than carefully selected indicators based on your own observations of the market. I'd be happy to be proven wrong, but I simply do not trust a black box that just happens to produce correct results 64% of the time.

In fact, I've tested a very simple strategy (using an EA in the Strategy Tester), which made ALL of its trading decisions ONLY based on the current open volume and profit/loss of open positions. No complex math and no other indicators. And it actually worked on EURUSD for 14 months straight, generating huge profits. But then, it blew up the entire account in a single day, losing all the profits and the initial account balance.

What you call "an edge" could just as well be pure coincidence.

 
NELODI:
Things where the pattern keeps repeating are predictable.

This was true for a long time. Many time series analysis models, like ARIMA and FFT indeed depend on periodically repeating patterns. But the landscape has changed since we have networks with memory cells. They are able to store information over long time distances and with irregular spacing of their appearance. This can be put to the extreme by introducing skip connections.

NELODI:
Take a 100-bar snapshot from any point in the History and try to find the exact match in the past or the future.

We don't need excact matches. Neural networks are very well capable of producing similarities and "associations".

May I ask: if you say that the past is no indication for the future, then why do you make trading decisions by applying indicators to the past?

I'm not questioning right or wrong of your statement. It's just that saying this and using indicators doesn't appear logical.


Personally, I would put it like this: the past is an extremely poor (yet not entirely useless) indication for the future (and by far too weak to justify the use of arbitrarily chosen unvalidated indicators, just cause we have decided to believe they work, or based on backtests also with PAST data).

 
Chris70:
May I ask: if you say that the past is no indication for the future, then why do you make trading decisions by applying indicators to the past?

I don't make trading decisions ONLY based on indicators. But even after I do place a trade, I keep watching it, until I see where it is going. And since I am scalping, I try to close all my trades within 10-20 minutes, either with a profit or a loss. What I should point out, is that I don't look at indicators to tell me what will be happening on the market next. Instead, I look at them to see what the market has been doing in the past and what is happening now. And I look at the charts as a whole. All the charts, not just the one. If I see an opportunity, I start by entering a trade with the absolute minimum volume I can open. If the trade doesn't work out, I close it with a loss. And if my assumptions were right, I add more volume once I see where things are going. When I think the move is over, I close all my positions. For this kind of strategy to work, you need a lot of Screen-time, but that is one thing you get automatically, if you just keep trading manually. And for me, it doesn't matter whether I trade DEMO or a REAL account, because it's all about the Screen-Time. If I know that I won't be able to focus entirely on trading (like now, for example), or that I might have to step out while a trade is still open, I only trade DEMO. It doesn't make me any money, but it does add to my Screen-Time, which improves my profitability when I do trade on a REAL account.

So ... my real "edge" is my Screen-Time and my Brain, not the indicators I use.

Can I turn my strategy into an EA? No, I can't. Because I can't copy my Brain into a computer.