Can Machine Reads Like Analysts Do?

Train a CNN to read candlestick graphs, predicting future trend.
Training & testing Dataset from Huge Stock Market Dataset-Full Historical Daily Price + Volume Data For All U.S. Stocks & ETFs.
Pytorch implementation

update:

Related paper: Trading via Image Classification (by J.P. Morgan)

Two Approaches

Approach 1

cnn4matrix.py
Apply convolution on data matrices directly.

Input: (5*n) matrices -> (Open,High,Low,Close,Volume)*(d1,d2,...,dn)

Output: classification result

Approach 2

cnn.py
Generate candlestick graphs first.

Input: candlestick graphs

Output: classification result

Current results

11 layers + residual block