deep-learning-bitcoin

Why we included this project

Bitcoin's price history is notoriously noisy, which makes this project an interesting experiment rather than a shortcut to profits. The core idea is to render candlestick charts as images and let a small AlexNet-style convolutional network learn to read them, after first resampling raw Coinbase tick data into labeled five-minute OHLC bars. The author reports about 70 percent accuracy on a 20,000-sample training set and marks the whole thing as work in progress, so nobody should expect a working trading system. What it does offer is a complete, reproducible pipeline, from downloading market data to training a classifier with Caffe and Docker, which makes it a useful reference for anyone who wants to try deep learning on financial time series.