Bitcoin прогноз

1. Доступ к данным Bitcoin из сайта https://coinmarketcap.com/

# load libraries
library(rvest)
library(tidyverse)

url <-read_html("https://coinmarketcap.com/currencies/bitcoin/historical-data/?start=20130428&end=20190328")
data <- url %>% html_table()
df <- data[[1]]
df <- as_tibble(df)
colnames(df) <- c("date", "open", "high", "low", "close", "volume", "market_cap")
df$date <- as.Date(df$date, "%b %d,%Y")
knitr::kable(head(df))

2. Визуализация данных

# Plot using ggplot2
library(ggplot2)
ggplot(data = df, aes(x = date, y = close)) + geom_line(colour = "blue", size = 1) +
  labs(title = "Historical data for Bitcoin",
     subtitle = "Accessing Bitcoin Data for the Forvision project",
     caption = "sources: Bitcoin (BTC); https://coinmarketcap.com",
     x = "Date",
     y = "Close")

# Plot using dygraphs
# Covert data to xts object
df1 <- xts::xts(df$close, order.by=df$date)
names(df1) <- "Close"
library(dygraphs)
dygraph(df1,main = "Historical data for Bitcoin") %>%
        dyRangeSelector(height = 20)

# Last 365 days
df2 <- df[1:365, ]
library(ggplot2)
ggplot(data = df2, aes(x = date, y = close)) + geom_line(colour = "blue", size = 1) +
  labs(title = "Historical data of Bitcoin for last 365 days",
       subtitle = "Accessing Bitcoin Data for the Forvision project",
       caption = "sources: Bitcoin (BTC); https://coinmarketcap.com",
       x = "Date",
       y = "Close")

3. Подготовка данных для прогнозирования

• Точка origin = “2019-01-01”
• C точки origin сделать прогноз на 60 дней

library(dplyr)
train_df <- filter(df, date <= "2019-01-01")
test_df <- filter(df, date > "2019-01-01")
knitr::kable(head(train_df))

test_df <- test_df[(nrow(test_df)-60 + 1):nrow(test_df), ]
knitr::kable(tail(test_df))

# covert data to time series 
train_ts <- ts(data = rev(train_df$close), end=c(2019, 1), frequency = 365)
test_ts <- ts(data = rev(test_df$close), start=c(2019, 2), frequency = 365)
plot(train_ts)

# Seasonal decomposition
fit <- stl(train_ts, s.window="period")
plot(fit)

# Test stationary
library(tseries)
adf.test(train_df$close, alternative="stationary", k=0)

## 
##  Augmented Dickey-Fuller Test
## 
## data:  train_df$close
## Dickey-Fuller = -2.2373, Lag order = 0, p-value = 0.4778
## alternative hypothesis: stationary

big p-value -> data is not stationary

adf.test(diff(log(train_df$close)), alternative="stationary", k=0)

## Warning in adf.test(diff(log(train_df$close)), alternative =
## "stationary", : p-value smaller than printed p-value

## 
##  Augmented Dickey-Fuller Test
## 
## data:  diff(log(train_df$close))
## Dickey-Fuller = -45.264, Lag order = 0, p-value = 0.01
## alternative hypothesis: stationary

plot(diff(log(train_ts)))

Small p-values -> diff(log(data)) is stationary

4 Построение auto.arima модели и прогнозирование

library(forecast)
fit <- auto.arima(log(train_ts), d = 1, ic = "aic", trace = TRUE)

## 
##  Fitting models using approximations to speed things up...
## 
##  ARIMA(2,1,2)(1,0,1)[365] with drift         : Inf
##  ARIMA(0,1,0)             with drift         : -7073.575
##  ARIMA(1,1,0)(1,0,0)[365] with drift         : Inf
##  ARIMA(0,1,1)(0,0,1)[365] with drift         : Inf
##  ARIMA(0,1,0)                                : -7072.756
##  ARIMA(0,1,0)(1,0,0)[365] with drift         : Inf
##  ARIMA(0,1,0)(0,0,1)[365] with drift         : Inf
##  ARIMA(0,1,0)(1,0,1)[365] with drift         : Inf
##  ARIMA(1,1,0)             with drift         : -7073.355
##  ARIMA(0,1,1)             with drift         : -7071.62
##  ARIMA(1,1,1)             with drift         : -7071.534
## 
##  Now re-fitting the best model(s) without approximations...
## 
##  ARIMA(0,1,0)             with drift         : -7079.827
## 
##  Best model: ARIMA(0,1,0)             with drift

# forecast, horizon = 60
forecast_fit <- forecast(fit, h = 60)
forecast <- exp(forecast_fit$mean)

forecast_data <- data.frame(date = rev(test_df$date), value = rev(test_df$close), forecast)
knitr::kable(forecast_data)

accuracy(forecast_data$value, forecast_data$forecast)

##                ME     RMSE      MAE      MPE     MAPE     ACF1 Theil's U
## Test set 330.1088 391.3159 354.2788 8.108618 8.731735 0.858627  60.05321

ggplot() + geom_line(data = forecast_data, aes(x = date, y = value, colour = "value"), size = 1) +
  geom_line(data = forecast_data, aes(date, forecast, colour = "forecast"), size = 1) +
  geom_line(data = train_df[1:100,], aes(date, close, colour = "value"), size = 1) + 
  ggtitle("60-days ahead bitcoin forecast from origin '01/01/2019'")