On checking correlogram at level we find distinct AR(1) signature
We conduct Unit root test and find that we cannot reject null hypothesis that there is unit root. Thus the series is non stationary.
We check the correlogram at first difference and find that the series becomes white noise process. Hence we cannot de trend and proceed in this way.
1 st method we proceed with AR(1) process. We find that coefficient of AR(1) is below 1 and significant with p- value of t-statistics (0.000) . We also find the p-value of f-statistics is 0.000, hence the model as a whole is significant and R^2 is a respectable 0.78
We check the residual diagnostic -> correlogram Q-statistics and find that the process has been converted to a white noise process. We also note that the Prob values are > 0.05
We run static forecast to forecast for last 5 days. 8/9/2008 8/14/2008. The MAPE comes out to be 1.54%.
We run dynamic forecast and get Mean Absolutute Percentage Error at 1.48%.
The residual graph comes out to be
Method 2
Deseasonalize the data.
Series ds=d(sen,0,5)
Check correlogram, here we see AR(1) signature and possible MA(1) signature
On conducting unit root test we find that the series has turned into non stationary series, p-value of t-statistics is 0.0005
We run the command
ls d(sen,0,5) AR(1) MA(1)
We find that AR(1) is significant with probability of t-statistics at 0.003, while MA(1) is insignificant.
We check the correlogram for residual diagnostics and find that we are getting SMA(5) signature.
We run the equation including SMA(5) and find both AR(1) and SMA(5) are significant. Coefficient of AR(1) and SMA(5) are < 1.
ls d(sen,0,5) AR(1) SMA(5)
We check the residual correlogram and find that the process has been converted to white noise process, all probability values are > 0.05
Next we run static forecast and find that MAPE is 0.83%
We run dynamic forecast and find that MAPE is 0.91%
We further check RSIDs and note the following plot
Method 3
Both deseasonalize and detrend the data.
series dst=d(sen,1,5)
We find distinct SMA(5) signature. We run unit root test and find that the p-value is < 0.05, thus we can reject the null hypothesis that there is a unit root. Hence the data has now become stationary.
We run the following command
ls d(sen,1,5) SMA(5)
We find that SMA(5) is significant with p-value at 0.000
We further check residual diagnostics -> correlogram q-statistics and find that it has been converted to white noise process, moreover p values are > 0.05
We run static forecast for the same dates and find that the MAPE is 1.03%
We run dynamic forecast and find that the MAPE is 1.21%
We check Resids and check the Actual, fitted and residual plot.
Conclusion
We find that method 2, de seasonalizing the data was most suitable in this case as we got the least Mean Absolute percentage error in that case and highest R^2 of 81%.