ML task, anyone can help?

[rounder twentytwo]

Hi, im new so im not sure if this is the proper category to post but here it goes my question:

I applied for an intership at an AM firm and they gave me a machine learning assignment to solve. My dutie was to distinguish real returns time series from fake ones (generated by them). I have tried my best. I have tried to test normality, volatility persistance, stationarity... but im only on the 70% accuracy band and I want to improve this result. Any ideas?

 

[pingu]

isn't 70% good enough? aren't you afraid of overfitting?

 

[rounder twentytwo]

I am asking about other relevant features or maybe another approach (a neural network?) to improve this result. However, i dont know what to use or what to do. But i think that the predictive power of the model could be more than 70% without incurring in overfitting

 

[rajanS]

you would have to plot the test MSE to identify if overfitting occurs.

Did you try to check the ACF of the two series? If it's close to 1, it clearly can't be returns.

Plot of Return squared would help as well. After running a t-test to see if mean is 0, if not, subtract the mean and plot the return squared. But I might be wrong here, lack of volatility clustering might not indicate the series is fake. It just says the conditional variance is constant.

Box Ljung test to test of serial autocorrelation? If there is heteroscadacity, try breusch-Godfrey test

Plot the kde of each series ?

 

[rajanS]

Can you try generative adversarial networks using long short term memory (lstm)?