Bitcoins predictions - A Time Series Presentation

About the project:

The project is a presentation of a time series analysis.

Goal

The project aims to accurately predict the value of bitcoin for the following 10 days using the appropriate time series analysis (using ARIMA models).

Data

The dataset is the daily closing price of bitcoin from the 27th of April 2013 to the 24th of February 2019 (Source: coinmarketcap.com)

Presentation

The presentation covers three main sections:

1. General observation of a time series.
2. Model specifications.
3. Predictions using the most appropriate model.

The project demonstrates the various steps in a time series analysis:

1. Trend and Seasonality analysis.
2. Time series transformations.
3. Models fitting.
4. Model Diagnostics

The project is presented in the following video

The analysis report includes descriptive analysis, time series visualisation, model specification, model fitting and selection, and diagnostic checking.

You can downloaded the R markdown report here.

What I’ve learned from this project

1. Overfitting: The project is an example of the debate between having an exact fit models versus accurate forecasts. The best fitted model does not have the lowest forecasting MASE, while the model with the best forecasts (MASE) capacity does not provide the best fit.

The project demonstrates the ability to work with R libraries including fGarch,forecast,fUnitRoots, lmtest, rugarch, TSA, x12 etc

Note: The presentation used MAE instead of MASE, this was a last minute change