Deep-learning stock forecasts glued together with Keras and Twitter

What it does Bulbea fetches historical market data through Quandl and pipes it into a configurable Keras RNN for price prediction. It also mixes in Twitter sentiment via Tweepy and TextBlob, returning a polarity score for any ticker you choose. Think of it as a friendly integration layer rather than a standalone modeling breakthrough.

The interesting bit At a time when quant libraries often ignored social noise, Bulbea treated Twitter mood as a parallel signal to OHLCV data. The README even closes the loop with a prediction-versus-actual plot and a single MSE number, which is more scientific rigor than many toy finance repos bother to show.

Key highlights

• One-liner data loading from Quandl’s Yahoo feed via a Share object.
• Built-in normalization and train-test splitting under bulbea.learn.evaluation.
• A thin Keras RNN wrapper that accepts raw layer dimensions for quick experiments.
• sentiment() utility that queries Twitter and returns a TextBlob-derived score.
• Apache 2.0 licensed.

Caveats

• Users still manually reshape NumPy arrays before feeding the model; the abstraction leaks quickly.
• The sentiment feature requires Twitter API credentials and never explains how a raw polarity score translates into a trading decision.
• Installation docs cite python setup.py install and the long-deprecated tensorflow-gpu package, so the project feels frozen in the TF 1.x era.

Verdict A decent sandbox if you want to see how deep-learning price forecasts and social sentiment can share the same repo. If you need a maintained, battle-tested quant stack, look elsewhere.