why did you take the difference of the decomposition data when it is already stationary to begin with? it says that p-value < alpha, therefore, its stationary as it is.. it also shows that acf was decreasing drastically to zero. did you differentiate it further so that you get lag order 7 for ACF? doesnt make sense
Really appreciate your content, but I had a question regarding the Dickey-Fuller test: In the Dickey-Fuller test you went from a p-value of 0.016 in PART 2 to a p-value of 0.01 in PART 3. If in both cases it was significant, could you not have left it at the previous step? why did you choose to further differentiate? Thanks in advance!
I am using your ARIMA forecasting method to forecast hourly electricity prices using hourly historical data. I realized that i have to use mstl() as my data has multiple seasonality. Hence i used the following code: count_ma = ts(na.omit(data_price$cnt_ma24), frequency=24) #where cnt_ma24 is moving average of every 24 hours, signifying a daily moving average and i am converting all those points to TS decomp = mstl(count_ma) deseasonal_cnt
What difference are you using ? And what was the auto arima model value? Try using a higher difference to get a lower P-value. You want the P-Value less then 0.5 or even lower. Then take the auto arima value and add in the lag value for your run with a higher difference and lower P-value. This is one of my favorite processes for a campaign with no control (99.9% of all campaign and forecasting requests I get never have a control). :)
why did you take the difference of the decomposition data when it is already stationary to begin with? it says that p-value < alpha, therefore, its stationary as it is.. it also shows that acf was decreasing drastically to zero. did you differentiate it further so that you get lag order 7 for ACF? doesnt make sense
Really appreciate your content, but I had a question regarding the Dickey-Fuller test:
In the Dickey-Fuller test you went from a p-value of 0.016 in PART 2 to a p-value of 0.01 in PART 3. If in both cases it was significant, could you not have left it at the previous step? why did you choose to further differentiate?
Thanks in advance!
I am using your ARIMA forecasting method to forecast hourly electricity prices using hourly historical data. I realized that i have to use mstl() as my data has multiple seasonality. Hence i used the following code:
count_ma = ts(na.omit(data_price$cnt_ma24), frequency=24) #where cnt_ma24 is moving average of every 24 hours, signifying a daily moving average and i am converting all those points to TS
decomp = mstl(count_ma)
deseasonal_cnt
What difference are you using ? And what was the auto arima model value? Try using a higher difference to get a lower P-value. You want the P-Value less then 0.5 or even lower. Then take the auto arima value and add in the lag value for your run with a higher difference and lower P-value. This is one of my favorite processes for a campaign with no control (99.9% of all campaign and forecasting requests I get never have a control). :)
Can you say how to do square root transformation for time series data in R
We could do that, but then it would be less accurate then the custom arima models....