Team Name: VPSS.csv

Code Link:

github.com/Sushmitha-KK/Datathon-CSE

 

Case Team Mentor:

• Prof. Dr. Subhabaha Pal ([email protected])

 Case Team:

• Varun Ramaprasad([email protected])
• Priyanka Mohekar([email protected])
• Suraj Shiwal ([email protected]@datascience.edu.net)
• Sushmitha K  ([email protected]@datascience.edu.net)

Team Toolset:

• R: dplyr, ggplot2, forecast, lubridate, caret
• Excel

Business Understanding

• Given a dataset consisting of Crypto-currency Market, the objective of this project is to design a model that predicts the cryptocurrency prices.
• The model designed will be useful for the cryptocurrency investors and investment firms to take informed decision on their investments.

Data Understanding

• The data set  consists of the following variables:
  • time – this includes the date and time corresponding to the bitcoin prices.
  • refID_coin – the reference_ID of each bitcoin.
  • price – the market value of the bitcoin at a given time.
  • marketCap – total value of a particular bitcoin share in the crypto currency market.
  • Circulatingsupply – Number of coins traded and existing at a moment.
  • Volume24h – How much of bitcoins have been traded in a day.
  • Movement1h – The variation of the price in an hour.
  • Movement24h – The variation of the price in 24 hours.

 Data Preparation

• The following cryptocurrencies were subsetted from the data : Bitcoin, Bitcoin Cash, Bitcoin Gold, Cardeno, Dash, Dogecoin, Eos, Ethereum, Ethereum Class, Iota, Lisk, Litrcoin, Monero, NEMcoin, Neo, Ripple, Stellar, Tether, Tron, Zcash.

datathon_bit <- datathon$refID_coin[datathon$refID_coin %in% c(1442,1445,1456,1446,1453,1477,1452,1443,1457,1451,1460,1448,1454,1447,1449,1444,1450,1474,1455,146)]
price_coin <- datathon %>% filter(refID_coin %in% c(1442,1445,1456,1446,1453,1477,1452,1443,1457,1451,1460,1448,1454,1447,1449,1444,1450,1474,1455,146)) %>% select(time,refID_coin,price)
View(price_coin)
crypt_currency <- split(price_coin,price_coin$refID_coin)
View(crypt_full[[1]])

•  The data has missing timestamps. To treat that, we use a sample dataframe which has all the dates between a given interval.

a<- “17/1/2018 11:25:00”
b<-“21/1/2018 02:50:00″
a<-as.POSIXct(a,format=”%d/%m/%Y %H:%M:%S”)
b<- as.POSIXct(b,format=”%d/%m/%Y %H:%M:%S”)
z <- seq.POSIXt(a, b, by = “5 min”)
as.data.frame(z)

• The above columns were then matched with the parent data frame and the values were imputed in price column using na.approx and in the price column using na.locf.

missing <- function(coin){
df <- merge(date_framed,coin,all.x =T)
df$refID_coin <- na.locf(df$refID_coin)
df$price <- na.approx(df$price)
return(df)
}

Modeling

• The data has been converted to time series using ts, which has properties start and frequency. The frequency is 288. The calucaltion for the same is as follows:

In an hour, there are (60/5) = 12 times 5 minutes occuring and in a day, there are 24 hours. Hence, (12*24) = 288, which is the frequency.

crypt_currency_ts <- lapply(crypt_full,function(df){ts(df$price,start = c(1,1),frequency=288)})

• The data was passed into the functions for ARIMA, Neural Networks and Exponential Smoothing

analysis <- function(price){
pred_ar = c()
price1=price
for (i in 1:288) {
price1=price[i:length(price1)]
arim_mod <- auto.arima(price1)
arima_forecast <- forecast(arim_mod,h=1)
pred_ar <- c(pred_ar,arima_forecast$mean)
price1 <- c(price1,arima_forecast$mean)
}
ets_mod <- stlf(price)
ets_forecast <- forecast(ets_mod,h=1)
pred_et <- c(pred_et,ets_forecast$mean)
neural_mod <- nnetar(price)
neural_forecast <- forecast(arim_mod,h=1)
pred_nn <- c(pred_nn,neural_forecast$mean)
return(list(arima=pred_ar,ets=ets_forecast,neural=neural_forecast))
}

forecasted <- lapply(crypt_currency_ts, analysis)

• The above code runs for 288 times as the model is designed to forecast the values for one day over 5 minutes interval, which is the frequency.
• We use auto.arima function to get stationarised time series data, i.e. time and variance are made constant with respect to time.
• We forecast the auto ARIMA model for the next 5 minutes and repeat for 288 times to get values for the whole day.
• To calculate the estimated value, we used mean function.
• The exponential smoothing is done by function stlf and and its forecasted value is calculated by mean of the stlf output and same is repeated for neural network using nnetar and its mean was calculated to estimate the value.
• The model gives point data for all the models used. The model finds scope for further enhancement for predicting the data more accurately.