G
geneticien
You use the term AI inappropriately, better to say network or NN or model or algorithm.
Forecasting stock prices with a feature fusion LSTM-CNN model using different representations
- Thread starter nooby_mcnoob
- Start date
I recall reading that they use bitmap to values of images of the charts, which are simply a matrix of (binary for black and white, 256 for RGB) values presented as input features to the learner. It is common to flatten the image matrix into a 1 dimensional row vector for each input observation. The target output can be the next day price/return/sign or whatever you wish to predict. All this is quite common and old news in the ML world.
If you wanted to know 'which' patterns were profitable you could simply filter your matrix to only return rows for which the next step target variables are above some threshold criteria (on average, or by a suitable metric). You can plot these input sets to get a better sense of what the patterns are. That's a simple version, but there are lots of variations on this.
If you are really interested in following up on the image recognition process, just look up OCR or handwriting recognition examples (of which there are dozens of examples freely available). Keep in mind, the handwriting samples are much more stable processes than financial patterns, so although the idea and process are similar, the results do not simply translate.
I've seen some edge in this, but not the way they present it. And I have not seen a lot of published work on the results. In the literature, you can look up Andrew Lo on kernel regression, which is a similar pattern recognition approach.
Last edited:
G
geneticien
>> What’s wrong with calling a sheep an animal?
Nothing.
However, by calling a simple CNN AI you put yourself a bit farther from people doing research in ML/DL/CS and a bit closer to people doing some AI marketing bullshit and selling or buying snake oil.
Nothing.
However, by calling a simple CNN AI you put yourself a bit farther from people doing research in ML/DL/CS and a bit closer to people doing some AI marketing bullshit and selling or buying snake oil.
Love it. Especially as I love marketing.
Though I’ve also heard a scientist saying that his colleagues should stop inventing new terms and pretending that all that stuff is more than statistics. While I have to use various technologies and would sound more bs if I started naming everything I use.
Anyone can really argue and slice these in any way they like, but arguing over “animal” might be a bit extreme so I won’t anymore.
G
geneticien
Many extremes exist. Some people say "It is AI" about too many things. Other people say that everything is just statistics, e.g., see the scientist that you mentioned. The most useful spot is likely in between. By calling something AI we don't progress much in terms of understanding how some particular mechanism works, instead we get too philosophical and this is rarely useful. By calling the same thing a CNN we are more likely to discuss technical details and learn something, e.g., about how the data was processed or how hyperparameters were tuned.
Personally, I prefer to call things by their technical names. On the other hand, I do believe that the current progress in AI is a real thing and as a scientist in this field I don't see how it will be stopped.
Personally, I prefer to call things by their technical names. On the other hand, I do believe that the current progress in AI is a real thing and as a scientist in this field I don't see how it will be stopped.
what do you think NN on trading? is it possible? ernest chan said no.
G
geneticien
If a human has a trading strategy to make money then a computational approach (stochastic or not) can do the same (unless you believe that god shows you the path in every trade). Whether this approach goes down to a simple NN or has a lot more things to it, it depends on the trading strategy itself. In most scenarios neural networks can become at least a useful component to generate signals. So the answer is yes, it is possible.
On the other hand, I believe that in certain scenarios/markets one can be good with a simple algorithmic approach which a NN could approximate but without providing any additional benefit which would pay off the overkill.
On the other hand, I believe that in certain scenarios/markets one can be good with a simple algorithmic approach which a NN could approximate but without providing any additional benefit which would pay off the overkill.
So basically it comes down to whether an actual holy grail strategy exists. But it doesn’t.
I would only add that it’s possible to replicate results of some top traders, but they won’t be repeatable during longer time periods. The longer the time the more losing streaks you’ll run into and will need to adapt/adjust the strategy (whether through training/nn/dnn or otherwise) and smooth out the risk and the results while sacrificing alpha. Then after adjusting it to work during 10-20 years you end up with very little alpha/profitability, still a risk of loss, and really nothing different than various strategies out there.
I would only add that it’s possible to replicate results of some top traders, but they won’t be repeatable during longer time periods. The longer the time the more losing streaks you’ll run into and will need to adapt/adjust the strategy (whether through training/nn/dnn or otherwise) and smooth out the risk and the results while sacrificing alpha. Then after adjusting it to work during 10-20 years you end up with very little alpha/profitability, still a risk of loss, and really nothing different than various strategies out there.
G
geneticien
This provokes people who have profitable strategies to show up and prove the opposite but I see little to no incentive for such people to do so.
Also, one does not need to beat all the market but to find a submarket where it is possible.