• C++ Programming for Financial Engineering
    Highly recommended by thousands of MFE students. Covers essential C++ topics with applications to financial engineering. Learn more Join!
    Python for Finance with Intro to Data Science
    Gain practical understanding of Python to read, understand, and write professional Python code for your first day on the job. Learn more Join!
    An Intuition-Based Options Primer for FE
    Ideal for entry level positions interviews and graduate studies, specializing in options trading arbitrage and options valuation models. Learn more Join!

Wanna try to beat classical trading algorithms neural networks?

Our company is an early-stage startup working on deep learning for algo trading. We are developing it for over a year now.

DeepCrypto.AI is a tool that mimics the process of the development of trading algorithms but uses neural networks instead of classical algos. NN training replaces algo creation and backtesting phase, and then exhaustive forward testing can be done.

The trained neural network is just an advanced algorithm. It can split incoming data in 100 000 or more features and therefore can find patterns which humans will miss.

We have run over a 1 000 000 back and forward tests and have received promising results.
However, we need professionals to prove that we are wrong if that's possible.

Feel free to do it at: www.deepcrypto.ai
 
Profitable traders USE their tools. Unprofitable ones SELL their tools.

mimics the process of the development of trading algorithms but uses neural networks instead of classical algos. NN training replaces algo creation and backtesting phase, and then exhaustive forward testing can be done.

Buzzwords with zero substance.

However, we need professionals to prove that we are wrong if that's possible.

I actually time travelled back from the year 2050. Prove me wrong if that's possible.
 
Do you really have enough data to make such a claim?

First, you should disclose that you're actually a part of this start-up.

Also that's an absurd premise. Burden of proof, evidence and data is on the person MAKING the positive claim.

I have developed a tool that mimics the process of the development of trading algorithms but uses neural networks instead of classical algos, which enables me to turn $100 into 5 billion in just 20 trading days. Go ahead, prove me wrong. I'll wait.

You guys are just SO generous, right? You've developped a profitable trading tool, and instead of using it yourselves to make money, you decided to sell it to others and spam on QuantNet, yes? Good job!
 
First, you should disclose that you're actually a part of this start-up.

Also that's an absurd premise. Burden of proof, evidence and data is on the person MAKING the positive claim.

I have developed a tool that mimics the process of the development of trading algorithms but uses neural networks instead of classical algos, which enables me to turn $100 into 5 billion in just 20 trading days. Go ahead, prove me wrong. I'll wait.

You guys are just SO generous, right? You've developped a profitable trading tool, and instead of using it yourselves to make money, you decided to sell it to others and spam on QuantNet, yes? Good job!


Why are you so aggressive? Did I offend you somehow?

About being generous... “During the gold rush its a good time to be in the pick and shovel business”.
It might sound counter-intuitive but thats my bet on growing big. The AI hype will force funds to do the research in trading AI area. And they either need to hire own team of data scientists or use our tool.

" We need professionals to prove that we are wrong if that's possible."
That was actually not meant to be cocky. I believe it works, but only evidence I have is forward testing. So maybe I am not taking something into account because of my confirmation bias.
 
Why are you so aggressive? Did I offend you somehow?

About being generous... “During the gold rush its a good time to be in the pick and shovel business”.
It might sound counter-intuitive but thats my bet on growing big. The AI hype will force funds to do the research in trading AI area. And they either need to hire own team of data scientists or use our tool.

" We need professionals to prove that we are wrong if that's possible."
That was actually not meant to be cocky. I believe it works, but only evidence I have is forward testing. So maybe I am not taking something into account because of my confirmation bias.

Your start-up is spamming a paid service on QuantNet, and then you are asking other people to prove a negative. I will point out that I did not once use profanity nor did I use a negative derogatory term to describe either you or your start-up. Nevertheless, any agressiveness that I have displayed is completely justified as a response to spamming of paid services that are likely useless or counter-productive.

Comparing your service to the gold rush is absurd. If only a single person had access to the immense reserve of gold during the gold rush, selling picks and shovels and letting others have access would absolutely be a terrible financial decision.

Any tom, dick or harry can claim his trading services work, and cite favorable, unverified, "forward testing" as evidence. There is no reason or evidence to believe your "service" works at all, and plenty to believe it does not.
 
Your start-up is spamming a paid service on QuantNet, and then you are asking other people to prove a negative. I will point out that I did not once use profanity nor did I use a negative derogatory term to describe either you or your start-up. Nevertheless, any agressiveness that I have displayed is completely justified as a response to spamming of paid services that are likely useless or counter-productive.

Comparing your service to the gold rush is absurd. If only a single person had access to the immense reserve of gold during the gold rush, selling picks and shovels and letting others have access would absolutely be a terrible financial decision.

Any tom, dick or harry can claim his trading services work, and cite favorable, unverified, "forward testing" as evidence. There is no reason or evidence to believe your "service" works at all, and plenty to believe it does not.

It is not a paid service. We are still doing research. That's why I wanted some feedback from quants, as I perceive them as highly professional people. But I was not expecting that some people here are that good that can provide fundamental feedback even without taking a look at the platform.

"There is no reason or evidence to believe your "service" works at all, and plenty to believe it does not."
This is typical Round-trip Fallacy. You have no evidence at all that our product does or does not work. And your lack of successful experience with deep learning is not relevant to this. Neither are useless products you encountered before.
 
Comparing your service to the gold rush is absurd. If only a single person had access to the immense reserve of gold during the gold rush, selling picks and shovels and letting others have access would absolutely be a terrible financial decision.

We are not talking about AI as Skynet. Neural network - it's just one algorithm. Is it possible that a plenty of profitable algorithms are working in same period of time? And also amount of trading capital matters a lot. So it's not that we have "access to the immense reserve of gold".
 
It is not a paid service. We are still doing research. That's why I wanted some feedback from quants, as I perceive them as highly professional people. But I was not expecting that some people here are that good that can provide fundamental feedback even without taking a look at the platform.

So you're not looking to monetize your services? Not now OR planning to? If you're willing to go on record and write that the purpose of the start-up is purely altruistic, and that you are not seeking a financial return by charging any sort of fee to your users in the future, I'll retract some of what I wrote. Are you willing to do so?

You have no evidence at all that our product does or does not work.

Profitable trading tools aren't the ones that are shared with anyone and everyone. If I find a way to print money, I wouldn't post on QuantNet "Hey guys look at what I found!" The notion that YOU would do so stretches credulity.

We are not talking about AI as Skynet. Neural network - it's just one algorithm. Is it possible that a plenty of profitable algorithms are working in same period of time? And also amount of trading capital matters a lot. So it's not that we have "access to the immense reserve of gold".

Efficacy of a profitable trading tool is reduced dramatically (as in to zero) if many people know about it. Even if you hide behind the notion that it's "different algorithms", trading is still a zero sum game (before commissions & fees). The fact that you don't appear to grasp this makes me believe you are disingenuous.
 
So you're not looking to monetize your services? Not now OR planning to? If you're willing to go on record and write that the purpose of the start-up is purely altruistic, and that you are not seeking a financial return by charging any sort of fee to your users in the future, I'll retract some of what I wrote. Are you willing to do so?

Profitable trading tools aren't the ones that are shared with anyone and everyone. If I find a way to print money, I wouldn't post on QuantNet "Hey guys look at what I found!" The notion that YOU would do so stretches credulity.

Efficacy of a profitable trading tool is reduced dramatically (as in to zero) if many people know about it. Even if you hide behind the notion that it's "different algorithms", trading is still a zero sum game (before commissions & fees). The fact that you don't appear to grasp this makes me believe you are disingenuous.

We build RESEARCH SOFTWARE. The math field of possible trading algorithms is vast. NN is not using 5 trading indicators, it splits the data into 100 000 or more indicators. For each run. 10 000 different models an hour. There is a room for everyone. 100s of trading indicators that can be used at once or in any combinations of those100s.

"So you're not looking to monetize your services?".
There is no need to attack me. I did not say that.

"Efficacy of a profitable trading tool is reduced dramatically (as in to zero) if many people know about it. Even if you hide behind the notion that it's "different algorithms", trading is still a zero sum game (before commissions & fees). The fact that you don't appear to grasp this makes me believe you are disingenuous."
Deep learning for trading is same size as algo trading as whole, or even bigger. It works somehow that there are more people making money with algo trading...
 
We build RESEARCH SOFTWARE. The math field of possible trading algorithms is vast. NN is not using 5 trading indicators, it splits the data into 100 000 or more indicators. For each run. 10 000 different models an hour. There is a room for everyone. 100s of trading indicators that can be used at once or in any combinations of those100s.

Which is a counter productive tool that gives the illusion of statistical significance. If there are hundreds of millions ways to design a trading algorithm, inevitably a few will be 5 or 6 standard deviations more profitable than the mean, even if the underlying indicators have ZERO correlation with price movement. Do you not understand that?

"So you're not looking to monetize your services?".
There is no need to attack me. I did not say that.

Asking a question is attacking you? You asserted that it's not a paid service, I asked if you EVER have plans to monetize your userbase, yet you complain that I'm "attacking" you. Wow.

Deep learning for trading is same size as algo trading as whole, or even bigger. It works somehow that there are more people making money with algo trading...

Are you aware of the many, MANY differences between trading in stocks/futures vs trading in crypto? And that a lot of the things which make algo trading so lucrative for SOME cannot possibly be provided with your service?

Funny how articles, appeared to be written by you, on medium.com, always mention "looking into the future of markets". Have you invented time travel? That is snake oil, not a real service.
 
Which is a counter productive tool that gives the illusion of statistical significance. If there are hundreds of millions ways to design a trading algorithm, inevitably a few will be 5 or 6 standard deviations more profitable than the mean, even if the underlying indicators have ZERO correlation with price movement. Do you not understand that?

I am not sure if you fully understand how neural network work. They try to figure out correlations from the data they get. So if "deviation algorithm" makes sustainable profit on 100 datasets is it possible that it really works?

Are you aware of the many, MANY differences between trading in stocks/futures vs trading in crypto? And that a lot of the things which make algo trading so lucrative for SOME cannot possibly be provided with your service?
Can you be more specific?

Funny how articles, appeared to be written by you, on medium.com, always mention "looking into the future of markets". Have you invented time travel? That is snake oil, not a real service.
"Looking into the future of markets" is just for marketing and maybe its not good. But once again you cannot make conclusion from a poor slogan that service is "snake oil". You have already decided that our service is bad and you are not interested in any fair evaluation of it. Thats confirmation bias.
 
I am not sure if you fully understand how neural network work. They try to figure out correlations from the data they get. So if "deviation algorithm" makes sustainable profit on 100 datasets is it possible that it really works?

So the answer is: no, apparently you DON'T understand the statistical concept of overfitting, how taking the best performing algorithm out of millions of possible ones inevitably give the illusion of statistical significance.

Let me ask you this: If I randomly generate 100,000 time series variables, all independent of each other, and look at each variable's correlation to S&P 500 returns, do you think there would be at least one which shows strong correlation?

Let's list the differences between crypto trading and trading US stocks/futures: liquidity, trading volume, lower transaction costs in bid-ask spreads/commissions/withdrawal & deposit fees. The regulatory environment, protection offered to speculators and investors, susceptibility to manipulation, counterparty risk etc. is completely different. The notion of "if algo trading works for SOME people here, it can also be consistently profitable in crypto" is absurdly childish.


"Looking into the future of markets" is just for marketing and maybe its not good. But once again you cannot make conclusion from a poor slogan that service is "snake oil". You have already decided that our service is bad and you are not interested in any fair evaluation of it. Thats confirmation bias.

"Of course that claim is outlandish but it's just marketing, so lying is okay." ....right. That does wonders for your credibility by the way.

Who would be more invested in whether this service of yours works or not? Me, an uninterested or observer, or you, someone who is part of the start-up? But I'm sure I'm the one more susceptible to confirmation bias, right?
 

Daniel Duffy

C++ author, trainer
We build RESEARCH SOFTWARE. The math field of possible trading algorithms is vast. NN is not using 5 trading indicators, it splits the data into 100 000 or more indicators. For each run. 10 000 different models an hour. There is a room for everyone. 100s of trading indicators that can be used at once or in any combinations of those100s.

"So you're not looking to monetize your services?".
There is no need to attack me. I did not say that.

"Efficacy of a profitable trading tool is reduced dramatically (as in to zero) if many people know about it. Even if you hide behind the notion that it's "different algorithms", trading is still a zero sum game (before commissions & fees). The fact that you don't appear to grasp this makes me believe you are disingenuous."
Deep learning for trading is same size as algo trading as whole, or even bigger. It works somehow that there are more people making money with algo trading...
This is real drivel. Write up or shut up.
 
So the answer is: no, apparently you DON'T understand the statistical concept of overfitting, how taking the best performing algorithm out of millions of possible ones inevitably give the illusion of statistical significance.

Let me ask you this: If I randomly generate 100,000 time series variables, all independent of each other, and look at each variable's correlation to S&P 500 returns, do you think there would be at least one which shows strong correlation?

Let's list the differences between crypto trading and trading US stocks/futures: liquidity, trading volume, lower transaction costs in bid-ask spreads/commissions/withdrawal & deposit fees. The regulatory environment, protection offered to speculators and investors, susceptibility to manipulation, counterparty risk etc. is completely different. The notion of "if algo trading works for SOME people here, it can also be consistently profitable in crypto" is absurdly childish.

"Of course that claim is outlandish but it's just marketing, so lying is okay." ....right. That does wonders for your credibility by the way.

Who would be more invested in whether this service of yours works or not? Me, an uninterested or observer, or you, someone who is part of the start-up? But I'm sure I'm the one more susceptible to confirmation bias, right?

If model overfits one dataset it performs poorly on previously unseen data. If model is performing good on a plenty of unseen datasets it is not overfitting, it is a good model. That's why testing on other datasets is a part of our research pipeline.

"The notion of "if algo trading works for SOME people here, it can also be consistently profitable in crypto" is absurdly childish."
Sure, but I did not claim that.
I claimed that deep learning for trading can produce a plenty of different trading algorithms, and some of them can be profitable. Crypto market data uses same ohlcv candles and therefore for a neural network will be no difference in between btcusd or aapl dataset.
I claimed that deep learning for trading is a huge research area and we provide a research tool, not a magic bullet.

"Of course that claim is outlandish but it's just marketing, so lying is okay." ....right. That does wonders for your credibility by the way.
Its not lying, its storytelling. And marketing is also a research process. And your claims about us selling time travel are absurd.

Confirmation bias is human thing, everyone is susceptible to it.
 
This is real drivel. Write up or shut up.

Neural network consists of input layer, hidden layers and output layer.
Input layer.
Input layer in our case will be always historical market data and/or trading indicators.
Testing indicators predictability in data science is called feature engineering. A lot of different indicators can be combined together in a various ways to improve model predictability. Finding out which trading indicators combinations are most predictable is a big research area.
Hidden layers.
Each hidden layer has N amount of neurons each of them contains a number. For example first neuron of first hidden layer might be: (open * its weight + close * its weight + ... + bollinger upper band * its weight + ...) and non-linearity function applied on this. And for each neuron similar values are calculated with unique weights. Those weights are learned on each epoch to find a combination that has biggest profitability. We can say that each neuron contains unique feature that has some correlation with end result. If you have 10 neurons in input layer, 100 neurons in 1 and 2nd hidden layer its a 100,000 features already. And you can have more layers with more neurons (or less).
The amount of hidden layers and neurons is a big research area.
Output layer.
Output layer gives you a prediction "buy", "sell" or "wait".
In a trained network the weights are stored and once you put new candles in it will apply them to received data and give you a prediction - "buy", "sell" or "wait".

The research process.
1. Training of neural networks.
First you select market data and trading indicators for training.
Than you select amount of neurons and hidden layers.
Than you press "Run" and will do all the training for you. After that you will receive a report containing profit of each model, its maximum drawdown, amount of trading operations, % of profitable trading operations etc. Most of the models will be trash, so at this point you select most promising ones for forward testing.
2. Forwardtesting
At this stage you can test you promising models on a plenty of other datasets to see if they show consistent profitability.

This is very rough description, but its hard to put it in one page.
What do you think?
 

Daniel Duffy

C++ author, trainer
Neural network consists of input layer, hidden layers and output layer.
Input layer.
Input layer in our case will be always historical market data and/or trading indicators.
Testing indicators predictability in data science is called feature engineering. A lot of different indicators can be combined together in a various ways to improve model predictability. Finding out which trading indicators combinations are most predictable is a big research area.
Hidden layers.
Each hidden layer has N amount of neurons each of them contains a number. For example first neuron of first hidden layer might be: (open * its weight + close * its weight + ... + bollinger upper band * its weight + ...) and non-linearity function applied on this. And for each neuron similar values are calculated with unique weights. Those weights are learned on each epoch to find a combination that has biggest profitability. We can say that each neuron contains unique feature that has some correlation with end result. If you have 10 neurons in input layer, 100 neurons in 1 and 2nd hidden layer its a 100,000 features already. And you can have more layers with more neurons (or less).
The amount of hidden layers and neurons is a big research area.
Output layer.
Output layer gives you a prediction "buy", "sell" or "wait".
In a trained network the weights are stored and once you put new candles in it will apply them to received data and give you a prediction - "buy", "sell" or "wait".

The research process.
1. Training of neural networks.
First you select market data and trading indicators for training.
Than you select amount of neurons and hidden layers.
Than you press "Run" and will do all the training for you. After that you will receive a report containing profit of each model, its maximum drawdown, amount of trading operations, % of profitable trading operations etc. Most of the models will be trash, so at this point you select most promising ones for forward testing.
2. Forwardtesting
At this stage you can test you promising models on a plenty of other datasets to see if they show consistent profitability.

This is very rough description, but its hard to put it in one page.
What do you think?
What I think? I am not impressed. It's bla bla and hot air. Give it up, you are wasting our time.
What is the precise algorithm?

Have a look at this report
https://www.datasim.nl/application/files/8115/7045/4929/1423101.pdf
 
Top