Week7

I choose to use the gasoline dataset in fpp2.

library(fpp2)

## Loading required package: ggplot2

## Loading required package: forecast

## Registered S3 method overwritten by 'quantmod':
##   method            from
##   as.zoo.data.frame zoo

## Loading required package: fma

## Loading required package: expsmooth

summary(gasoline)

##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##   6.321   8.049   8.713   8.553   9.121   9.815

str(gasoline)

##  Time-Series [1:1355] from 1991 to 2017: 6.62 6.43 6.58 7.22 6.88 ...

autoplot(gasoline)

gas.nn<-nnetar(gasoline,lambda=NULL)
gas.nn

## Series: gasoline 
## Model:  NNAR(11,1,6)[52] 
## Call:   nnetar(y = gasoline, lambda = NULL)
## 
## Average of 20 networks, each of which is
## a 12-6-1 network with 85 weights
## options were - linear output units 
## 
## sigma^2 estimated as 0.05376

gas.fc<-gas.nn%>%forecast(h=24)
autoplot(gas.fc)

?gasoline

Seen as data is weekly, I will use 52 weeks to get a better forecast.

gas.fc<-gas.nn%>%forecast(h=52)
autoplot(gas.fc)

gas.fc<-gas.nn%>%forecast(h=104)
autoplot(gas.fc)

accuracy(gas.fc)

##                        ME      RMSE       MAE         MPE     MAPE      MASE
## Training set 0.0008387389 0.2318627 0.1793097 -0.07735182 2.135956 0.6222515
##                     ACF1
## Training set -0.02809231

gas.arima<-auto.arima(gas)
gas.arima.fc<-forecast(gas.arima,h=104)
autoplot(gas.arima.fc)

accuracy(gas.arima.fc)

##                    ME     RMSE      MAE       MPE     MAPE      MASE
## Training set 27.72258 1579.457 893.4503 0.2722739 3.900232 0.4789312
##                     ACF1
## Training set 0.002271287

autoplot(gas.arima.fc)+
  autolayer(gas.fc)

gas.fc1<-gas.nn%>%forecast(h=26)
autoplot(gas.fc1)

gas.arima1<-auto.arima(gas)
gas.arima.fc1<-forecast(gas.arima1,h=26)
autoplot(gas.arima.fc1)

accuracy(gas.arima.fc1)

##                    ME     RMSE      MAE       MPE     MAPE      MASE
## Training set 27.72258 1579.457 893.4503 0.2722739 3.900232 0.4789312
##                     ACF1
## Training set 0.002271287

accuracy(gas.fc1)

##                        ME      RMSE       MAE         MPE     MAPE      MASE
## Training set 0.0008387389 0.2318627 0.1793097 -0.07735182 2.135956 0.6222515
##                     ACF1
## Training set -0.02809231

autoplot(gasoline)+
autolayer(gas.arima.fc1)+
  autolayer(gas.fc1)

My RMSE is vastly different. WIth the Neural Net producing very small values for the performance metrics. Even though the forecasts seem to fail the eyeball test for the neural nets. I tried 26 weeks and 2 years. In both scenarios the Neural net is better, but that doesn’t make sense.