Bitcoin Price Forecasting
Naveen Venkataraman
June 4th, 2015
Goal: Forecast Net Profit For May 2015
• Collect BTC Data
• Analyze Price Volatility
• Isolate Stationary Series
• accounting for stationarity, seasonality and trends
• Comparison of Models, Forecasts and Metrics
• Potential Model Improvements / Follow-up Analysis
STEP1: FETCHBTC PRICEDATA
STEP2: CONDUCTEXPLORATORYANALYSIS
summary(btc.data)
OHLC, Volume, Date Ranges
tsdisplay(btc.data)
adf.test(btc.data)
plot(decompose(ts(data,frequency=365),type=type))
Separating out trend and seasonality
STEP3: ISOLATE 2015PRICES
(TOCREATETRAIN/TESTDATA)
2015DataisStationary
Train:Jan–Apr2015
Test:May2015(28days)
JAN–APRdata:PossiblyanMA(2)process
STEP4: MODELFITTINGAND FORECASTING
HoltWinters1
Exponential smoothing
With trend
Without seasonal component
HoltWinters2
Exponential smoothing
Without trend
Without seasonal component
ARIMA
ARFIMA
fractional differencing
ForecastingMetricsIndicateARFIMAToBeTheBestModel
Holt Winters (without trend and without seasonality) is next best
Follow-on Analysis
• Use Cross Validation for tuning parameters
• Factor in Volume information in relation to price
• Estimate discontinuous / missing information on certain
trading days
• Use Complete dataset (2011 – 2015)
• Weigh recent information more
• Smooth Extreme Volatility
• Evaluate Co-pair trading strategy with other asset classes
• Gold
• Currencies
APPENDIX
Using Quandl
• Setup an account (FREE)
• Get an API KEY (FREE)
• Install Quandl library (based on platform)
• Choose data exchange format
• Supported platforms: https://www.quandl.com/tools/full-list
Learning About Bitcoin
• “Mastering Bitcoin”
by Andreas
Antonopoulos
• MOOC:
http://digitalcurrency
.unic.ac.cy/free-
introductory-mooc
• MIT
• Bitcoin Club
• Media Lab

Bitcoin Price Forecasting