Hello DL4J community,
I have been experimenting with a recurrent LSTM network for time series prediction for some time.
I use one-hour candlestick data. Each time step has 7 inputs: OHLC prices, trading buy and sell volumes and number of trades. Outputs are the same.
The training data has approx. 32000 data points, the subsequent evaluation data 2400. I have divided the data points sequentially into individual files with 50 entries each in order to be able to properly work with the help of SequenceRecordReader and DataSetIterator.
I let the LSTM predict 1 data point into the future (=1 hour) at a time.
Everything works properly and without any errors.
However, a recurring pattern emerges in the chart as soon as I plot real and predicted close price data (This behaviour occurs with both training and evaluation data).
The pattern looks like this:
It seems like for every 50 time steps, which as mentioned represent 1 file, the visual pattern restarts from a lower entry point and starts to rise up from there.
Does anyone have a guess about this pattern or how to get rid of it?
As a first step, I would like to approximate the forecast data to the real data as well as possible before forecasting further into the future.