I am working on this project to get a better understanding of neural nets. The goal is to create images from historical forex data and see if a net can find trends and make accurate predictions.
- Launched an AWS instance to created the images and vectors.
- Successfully implemented parallel processing to cut down on creation time (still took about 4 hours on a 40 core machine)
- Created over 272,000 images and their corresponding vectors; each containing 5 ticks on an hourly scale from 2001 to 2015.
- Transferred the vectors to a GPU AWS instance to begin training the net.
- Altered the net framework from the cifar10_cnn Keras Net (this example utilizes RGB color)
- Tried running the net as a classifier; predicting whether the market will go up, down, or stay the same in the next tick.
- I am hoping that this will give me an indiction of whether this idea has any potential.
- Using 50x50px images; 5 ticks; close, high, and low data.
- Using fill_between to add more content to each image.
- Using a unique background color for each currency.
- Best validation accuracy is about %50.01...
- Using 50x50px images; 5 ticks; close, high, and low data.
- Using fill_between to add more content to each image.
- Best validation accuracy is about %50.0...
- Using 50x50px images; 5 ticks; close, high, and low data.
- Using fill_between to add more content to each image.
- Best validation accuracy is about %48.9...
- Using 40x40px images; 5 ticks; open, close, high, and low data.
- Best validation accuracy is about %48...
- Using 50x100px images; 10 ticks; open, close, high, and low data.
- Best validation accuracy is about %48... Not much difference.
- Using 50x50px images; 5 ticks; close data only.
- Best validation accuracy is about %48... Again, not much difference...
- Only include images from prime trading hours.
- Only graph closing ticks, or some other variety.
- Change number of ticks being graphed.
- Allow for overlap in ticks on each graph. Currently no graph data overlaps any other graph.