Hello.

I am trying to predict the closing price of a forex market for the present day. Specifically, for testing I am currently trying to predict the EURUSD. The lowest relative absolute error which I was able to get was 2.0518% which is quite good. I wasn't even able to get this accuracy with an MLP (Multi-Layer Perceptron). The problem is that I have one variable which is dependent on the closing of this current day. This variable is the typical price and is calculated as (High+Low+Close/3). The idea is that since this is the typical price, it will not change a lot during one day. However, with some volatility in the market, it will drastically change and hence arrive at incorrect predictions.

My current data set contains the following variables:

Yesterday's volume
Yesterday's typical price
Today's typical price
Yesterday's low
Yesterday's high
Yesterday''s open
Yesterday's close
Yesterday's weighted price
Today's open
Today's Close

The last one, Today's close is of the course the one which is predicted. The mentioned low relative absolute error was achieved by doing a 10 fold cross validation. However, the today's typical price was the actual price on closing, and nof for example, during the middle of the day. So this is why the error was so low.

My question is, what can I improve in my dataset to predict Today's closing with at least 65-70% accuracy?

P.S. my historic dataset contains 2000 instances which contain these variables for every day's closing.

Any help would be much appreciated.

Thanks and Regards,
Chris