I have a time series data with date and revenue columns. That is daily data. I need to forecast revenue for next 365 days. When i applied adf test, my p-value was much higher and hence showed that the data is non stationary. So I applied log and differencing to make data stationary and then applied auto.arima. Is my approach correct? How do I validate my result? Also how can i get back the undifferentiated and unlogged data in forecast?
I have a time series data with date and revenue columns. That is daily data. I need to forecast revenue for next 365 days
$\begingroup$ Forget about assuming pure ARIMA structure ..see stats.stackexchange.com/questions/354726/… for discussion as to incorporate daily effects, weekly effects etc .. into a hybrid model. $\endgroup$– IrishStatJul 11, 2018 at 15:25
There is no need to differentiate before applying auto ARIMA model in R. It will automatically select the order of differentiation. However, I'd recommend to check ACF/PACF of auto ARIMA model just to make sure that there is no significant autocorrelation left in residuals. Also, don't forget to check for conditional heteroskedasticity before forecasting as it might result in inaccurate forecasts.
Regarding validation you could divide your data into two different samples: training (let's say 80% of data) and test (let's say 20% of your data). Use training sample to 'train' the model and then forecast. After that compare your forecasts obtained using train sample with actual values from test sample. This approach would allow you to determine how accurate your forecasts are. If they are quite accurate you could then forecast future values with confidence.