Hacker News Reader

Comment by fredliu | Hacker News Reader

Total noob question in this space (algotrading with DL algorithms).

What DL is good at, is to automate tasks that are easy for humans (e.g. telling cat from dog, understand a sentence from sound wave, translation, etc) but hard for machines without DL.

Now, if a task is even hard/borderline impossible for humans to achieve (e.g. predicting stock market with higher accuracy and consistency), why would we believe DL could do a better job than humans?

View on HN

DL isn’t just good at automating human tasks (see AlphaZero and AlphaGo).

It can be used to recognize patterns and train to solve problems better than humans. Though in this case since the market is people making predictions about a prediction I’m not sure how much it’d help.

There’s probably something about humans using a rough estimate of what the stock has been historically to inform where it could go though so it’s value is probably non-zero.

By DL, i meant generic DL techniques like CNN, LSTM (used in this article), which relies on large amount of labeled data to train, and predict on similar data, thus my comment about automation. Alpha Go (Zero) is very specialized for the game of go, not sure how much of its specialized algorithm could transfer to other generic use cases.

>Alpha Go (Zero) is very specialized for the game of go, not sure how much of its specialized algorithm could transfer to other generic use cases.

AlphaZero is a generalized successor, and it does just that:

https://en.m.wikipedia.org/wiki/AlphaZero

Good to know it generalized, but seems it's only generalized on board games like problems where you have a problem space to search through, is that understanding correct? If so, it probably won't help in the use cases we are talking about here, right?

Yeah it's not an AGI, but I think the response was more that DL isn't necessarily limited to automating tasks that are easy for humans but hard for machines without DL (which I think was your original statement).

DL could potentially do a better job by recognizing patterns in the dataset that lead towards winning (making more money) that humans might miss. Like how AlphaZero can recognize moves in Go or Chess that humans don't understand are the best moves to make.

It's not obvious to me how this would be done, but I think it's plausible that some clever implementation could help.

I think relating DL with human skills is confusing you. DL already beats humans in recognizing dog breeds. DL could always do a better job, because it looks at more patterns than humanly possible.

The reason I relate human skill with DL, is because for tasks where there are real patterns that humans can consistently recognize, DL is applicable, because it relies on human labeled inputs. For patterns, even humans can't consistently recognize (or if there are patterns at all), there's no way for DL to be better than humans, simply because the inputs from human can't be trusted. The example of DL does better job at e.g. recognizing dog breeds, recognizing cancers, whatever, is because those jobs are possible for humans to begin with. The only difference there would be DL could work on more human labeled data than any real human, so that they can recognize more patterns. But if those human labeled data/patterns dont exist in the first place, why we would believe DL could do a better job?

I think you are still confused. Replace DL with software.

Because there's no requirement that deep learning must only excel at tasks humans are good at.

That said, the pitfalls of using ML in trading systems are numerous.

I didn't mean it's a requirement, but a fact that what ever DL could do is based on its input (labeled data), and that input has to come from human, or the task underneath is possible for human.

Reply on news.ycombinator.com