What it does

Feeds sliding windows of stock price data—open, high, low, close, volume—into a 1D convolutional neural network. The model outputs two confidence scores: bullish or bearish. The author trained a proprietary model but released the architecture for anyone willing to supply their own historical data and GPU time.

The interesting bit

The core conceit is genuinely clever: technical analysis is already about pattern recognition on charts, so why not use a tool built for pattern recognition? The project flattens the problem into 1D convolutions ([batch_size, 128, 5] tensors) rather than pretending price data is a 2D image. It’s an earnest, slightly naive translation of computer vision intuition to financial time series.

Key highlights

• Pure TensorFlow 1.x implementation with manual tf.variable_scope and truncated_normal_initializer—a clear fossil from ~2016
• 1D convolutions with width-9 kernels and stride-2, stacked through batch norm to a [batch_size, 2, 1024] representation before softmax → sigmoid
• Five input channels: Open, High, Low, Close, Volume
• Moving window of 128 timesteps per prediction
• Author explicitly states the trained weights are proprietary; you bring your own data and compute

Caveats

• No validation metrics, backtest results, or performance numbers anywhere in the repository
• “You have to tinker with the hyper parameters, archeteture… if you want to achieve good results”—the README’s own admission that this is a starting kit, not a working system
• TensorFlow 1.x and Python 3.5 dependency chain; significant archaeology required to run today

Verdict

Worth a look if you’re teaching or learning about unconventional CNN applications, or if you want a clean example of 1D convolutions in raw TensorFlow. Skip it if you need reproducible trading performance, modern frameworks, or any evidence that this actually beats a coin flip in practice.