Forecasting stock prices with a feature fusion LSTM-CNN model using different representations

[IAS_LLC]

https://en.wikipedia.org/wiki/Hidden_Markov_model
https://en.wikipedia.org/wiki/System_identification

In a nutshell, my goal is to build a probabilistic markovian "state transition model", and than use Bayesian Inference to incorporate market "observations". A transition model can be built using a variety of methods. A popular choice is the EM algorithm.

Correction...my goal is to make money. The above goal is a secondary goal used to achieve the primary.

 

[nooby_mcnoob]

https://en.wikipedia.org/wiki/Hidden_Markov_model
https://en.wikipedia.org/wiki/System_identification

In a nutshell, my goal is to build a probabilistic markovian "state transition model", and than use Bayesian Inference to incorporate market "observations". A transition model can be built using a variety of methods. A popular choice is the EM algorithm

Ah, it's funny how people converge on similar ideas isn't it...

 

[IAS_LLC]

This: https://en.wikipedia.org/wiki/Recursive_Bayesian_estimation

 

[IAS_LLC]

Ah, it's funny how people converge on similar ideas isn't it...

Logical people do logical things.

Your background in engineering or computer science? I'm a Control Systems guy...so state-space modeling and kalman filtering is at the forefront of how I approach problems.

 

[nooby_mcnoob]

Logical people do logical things.

Your background in engineering or computer science? I'm a Control Systems guy...so state-space modeling and kalman filtering is at the forefront of how I approach problems.

Computer science, but I do gravitate towards signal processing. The only problem is I have no formal introduction to it and I don't really know how to become an expert in it. Is there a "signal processing" book that you would recommend?

 

[IAS_LLC]

"Fundamentals of Object Tracking" by Challa et al is what I used to use as a reference for bayesian filtering (Kalman,etc..). It explains the theory, but I primarily used it as a recipe book. Its applications are obviously geared towards aircraft tracking...but a Bayesian Filter is a Bayesian Filter...only the state transition and observations models change.

 

[nooby_mcnoob]

Awesome, will check it out.

 

[geneticien]

From the paper:
"In the case of Naive model, since it is judged that Pt is the same at the time point t + 5, it cannot be determined whether it will rise at the time point t + 5. Therefore, we assumed trading at each time point and as a result we recorded 15.48% revenue."

I don't get it. If P_{t+5}=P_t, i.e., the naive strategy says that the price will not change, then how exactly it can be used to make profit?

"Buy and hold strategy is a simple trading method to buy shares at the time of starting trading and sell shares at the end of the trading period. We confirmed that revenue was 3.09% when we implemented this strategy."

So it seems that the time interval is so small that the market is not really positive in that time window. In other words, even a few years of US equities would lead to a large positive % (except from 2008-2009). Short out-of-sample time series are not really convincing.

Then, their approach shows 17.38% compared to 15.48% of the naive strategy. Can the difference can be attested to overfitting of hyperparameters? They don't show a single figure with evolution of training and validation curves, that is weird.

I hope you find the paper useful but after my quick check it looks like a mess to me. Well, it is PLOS One, that is probably why.

 

[dtrader98]

Just looking at that paper raises several flags, and makes it not really worthy of continuing (other than academic theory)...

Some examples
1) Use 1 year of data??? 2016 that's great for the one year of data. Too bad next year may be nothing like it.
2) RMSE is kind of a useless metric to your trading system. So, you got an RMSE improvement of .34%, so what. How does that translate to profit, or some other suitable metric?
3) Why compare system 1 RMSE to system 2 etc? It's ok, but in many ways, useless. Better to compare all to some kind of baseline (with a more profit oriented objective/loss function).

They may have addressed some of this as I didn't read further.

Markov chains are very useful. Defining the states and optimization criteria, are the tough parts (as is the case with all ML).

 

[nooby_mcnoob]

From the paper:
"In the case of Naive model, since it is judged that Pt is the same at the time point t + 5, it cannot be determined whether it will rise at the time point t + 5. Therefore, we assumed trading at each time point and as a result we recorded 15.48% revenue."

I don't get it. If P_{t+5}=P_t, i.e., the naive strategy says that the price will not change, then how exactly it can be used to make profit?

"Buy and hold strategy is a simple trading method to buy shares at the time of starting trading and sell shares at the end of the trading period. We confirmed that revenue was 3.09% when we implemented this strategy."

So it seems that the time interval is so small that the market is not really positive in that time window. In other words, even a few years of US equities would lead to a large positive % (except from 2008-2009). Short out-of-sample time series are not really convincing.

Then, their approach shows 17.38% compared to 15.48% of the naive strategy. Can the difference can be attested to overfitting of hyperparameters? They don't show a single figure with evolution of training and validation curves, that is weird.

I hope you find the paper useful but after my quick check it looks like a mess to me. Well, it is PLOS One, that is probably why.

Like most papers, they probably have mistakes that prevent the results from being reliable. I usually read for ideas and assume if they don't provide all the code that they don't really believe the results work.