Example: RNN fit for stock market prediction. The number of inputs, bias, activation function and etc., Suggest the best neural network model to solve the forecasting issues.
Hi, Some sources related to RNN suggest, that it is possible to use RNN, but by my experience. It is better to use Feed forward networks, such as MLP. RNN tends to overfit training dataset. You may use a lot of historical stock data, but todays price movements are not same as they were in 90s and 20s (volatility, volume, momentum frequency distribution). I have quite good results combining a few feedforward networks into one ensemble model. I used 3 kind of price inputs. Raw, percent change and delta change. Raw is normalized close price, percent change is simple normalized (Close[T]-Close[T-1])/Close[T-1] or natural logarhitm LN(close[T]/close[T-1]). Use what you prefer and delta change is normalized close[T]-close[T-1]. For normalization i use MinMax normalization. And i suggest to use 10 neurons on input, 2 neurons with bias neuron as hidden layer, and 1 neuron output layer. Your research is to determine future price or classify future direction movement "Up" or "Down" ?
If the answer would be that simple, I would be a billionair. ;-)
Well, there is no gerneric aproach to your application. It depends on the market and domain you want to model. I was doint ANN based stock market trading for several years, but you can never predict the unforseen event or stock market crisis.
RNN is best for all type of sequential data analysis. As in forecasting data changes with time, and as RNN can learn changes in time domain so it could be better solution for prediction.
Try variant of RNN called LSTM long short term memory concept in your training model hope you find good results.
- Badges
- Science topic
- Similar topics
- Philosophy
- Esthetics
- Interpretation