wanpixiaozi / lstm-time-series-forecasting

Predicting the behavior of $BTC-USD by training a memory-based neural net on historical data

Forecast bitcoin with deep learning

---

What the model does

Unlike previous renditions of this project, this model predicts behavior using the historical data alone. The performance benefit of synchronizing sentiment with stock trends in the multivariate analysis was found to be nominal relative to the cost in data complexity, and far outweighed by the potential for adversarial attack.

Why historical data only?

After inspecting the reddit data used, it became clear that some agent was spamming "I love bitcoin :)" in a way that was clearly inorganic. The signal to noise ratio in the sentiment data is unclear, but there are ways to combat this problem.

Update: Portability

Unlike the previous rendition, this uses pandas_datareader to pull historical data from Yahoo Finance, allowing this project to be modified and trained on any stock/cryptocurrency.

Results

It seems the model is not overfitting, and capturing the general trend quite well.

Reflections

Perhaps a discriminator could be trained to identify posts which are likely not created by authentic human users.