README.md

Bitcoin closing price prediction

Links to online notebooks (in case unable to load in Github):

• https://nbviewer.jupyter.org/github/yannafaith/PredictingBitcoin/blob/master/BitcoinWithSentiment.ipynb This notebook expands upon the most succesful model from the notebook below, by including bitcoin's total volume and a sentiment indicator (with value indicators ranging from extreme fear to extreme greed).

• https://nbviewer.jupyter.org/github/yannafaith/PredictingBitcoin/blob/master/BitcoinPrediction.ipynb This notebook explores three different (simple and un-optimized) ML models for predicting bitcoin's closing price, given historical closing prices from Jan 01 2018 to Feb 19, 2019.

Summary/Ranking of Models

Multivariate LSTM

• Training set was 365 days, and the test set was the remaining 48 days.
• had 3 input vars/features, looked at the data one timestep back, output a predicted closing price.
• used 200 epochs.
• Had a RMSE of 96.75.

For the rest of the models, the training set was 66% of the dataset.

LSTM (Long Short-Term Memory Model)

• had one input neuron, four neurons in a hidden layer, and one output neuron.
• looked back 3 days
• used 200 epochs and a batch size of five.
• Had a RMSE of 110.88.

AIRMA (AutoRegressive Integrated Moving Average)

• Had a RMSE of 157.71.

Baseline model was a "naive" persistance model.

• At timestep t it outputs (t-1).
• This had a RMSE of about 160.54

MLP (Multi-Layer Perceptron Model)

• had one input neuron, five neurons in a hidden layer, and one output neuron.
• used 50 epochs and a batch size of one.
• Had a RMSE of 160.56.