This is a summary of my machine-learning model that predicts the price of Bitcoin.

This is an ML model project in TensorFlow that utilizes transfer learning, convolutional neural networks, and relational neural networks.

Table of Contents

• Motivation
• Summary of approach
• Results
• Insights
• What I learned
• How to use this repository

Motivation

Upon starting my ML journey, I wanted to create something that could predict stuff in the future like house prices or Bitcoin prices.

Summary of approach

I started with a basic sequential CNN with 2 layers with a portion of data and then I scaled them up. The evaluation results are here:

Model1:

Model2:

Model3:

The results weren't what I wanted so I decided to create a Conv1d model. The evaluation results are here:

{'mae': 574.57104,

'mse': 1195531.0,

'rmse': 1093.4034,

'mape': 2.5650182,

'mase': 1.0093622}

I decided to use the LSTM for the time series data. The evaluation results are here:

{'mae': 579.0948,

'mse': 1206146.1,

'rmse': 1098.2468,

'mape': 2.613475,

'mase': 1.0173092}

It looks like our model is performing worse. I decided to use multivariate time series data for the next model. The evaluation results are here:

{'mae': 568.42395,

'mse': 1179631.1,

'rmse': 1086.1083,

'mape': 2.541055,

'mase': 0.99856347}

I then replicated the layers presented in the N-Beats model from "NEURAL BASIS EXPANSION ANALYSIS FOR INTERPRETABLE TIME SERIES FORECASTING" paper. Here is what the model looks like:

The evaluation results are here:

{'mae': 590.5516,

'mse': 1231043.1,

'rmse': 1109.5238,

'mape': 2.6545231,

'mase': 1.0374355}

To end it off I created an ensemble and staked the models on top of each other. Results are shown in the results section

Results

{'mae': 566.77386,

'mse': 1151240.8,

'rmse': 1072.9589,

'mape': 2.5520275,

'mase': 0.9956647}

Prediction results:

Comparing Models:

Insights

Trying to predict future results does give us a model that is successful in doing so however, especially when it comes to cryptocurrency everything is volatile and relies on a series of factors that can never fully be implemented into our model. It is like predicting the stock market. We can get close but never fully accurate. The N-beats models and others do get close but they can never predict the future fully.

What I learned

• How to scale data into models using Windows.
• Implementing RNNs and how to create your layers in Tensorflow.
• How to vizualize timeseries data.
• Can never fully predict the future, especially with cryptocurrency.

How to use this repository

Open the ipynb file in Google Colab or Jupyter Notebook and run. Make sure there is a GPU with high RAM.