I have been working on a LSTM model these days, however the progress is not going very well. There are several problems I still need to figure out.

lstm (Graph source: Towards Data Science,TDS, by Michael Phi)

  1. The model I built seems fit on stocks with more ’linear’ price fluctuation, but doesn’t work well on those with ’exponential’ price fluctuation (no matter upside or downside).

  2. Input parameters are not enough. Right now, I’m using time series of past close price for the forecasting. However, there are many more technical indicators I can use, I need more time to test each of them and select those with high predictive power.

  3. Fine tuning of hyper parameters. LSTM algorithm is extremely powerful, especially when working on data in sequence with short gaps. However, it still requires many works to be done on setting the hyper parameters. I suppose one fine tuned LSTM model can only work well of a specific stock. This actually makes sense because if you want to profit from trade, it is quite essential that you are familiar with the stock you trade. You have to know some characteristics of it, such as its Beta, drawdown pattern, MA, etc,.

  4. I’ve tried other models as well, including SVM, Randomforests, KNN and the powerful XGBOOST. But their results are not very encouraging. Or, maybe I can try to convert the predicting targets to a binary target, and use the features from previous day to forecast the outcome of present day. Haven’t tried this idea yet, but it seems worth a shot.

I will write a post to introduce the model when it passes my test and shows real ability to help me forecasting the direction of a certain stock.

visitors