Regression Type Estimation

#1
Hello,

Attached is a time series data that has a plot of the percentage of breakdowns on a scheduled hours across the days. The objective is to predict the value for the next day.

I tried to use linear regression to fit the data, but got a p-value higher than 0.05 and R square value less than 0.1. I also tried ploynomial regression but the got higher p-values.

Tried ARIMA but the RMSE was high.

Can someone advise me with the different options to fit the data ?
 
Last edited:

noetsi

Fortran must die
#2
You should never do linear regression with time series as it violates the assumption of linear regression normally.

I would try exponential smoothing, a family of time series that includes Holt Winston, Seasonal, Damped trend, Single smoothing and the like. Its arguably the easiest of the true statistical time series models to use. Personally I use MAPE rather than RMSE to assess predictions, but everyone has their own preference :)
 
#4
You should never do linear regression with time series as it violates the assumption of linear regression normally.

I would try exponential smoothing, a family of time series that includes Holt Winston, Seasonal, Damped trend, Single smoothing and the like. Its arguably the easiest of the true statistical time series models to use. Personally I use MAPE rather than RMSE to assess predictions, but everyone has their own preference :)
Thank you for the tip. I will try these methods.
 

hlsmith

Omega Contributor
#5
It probably could break the normality of residuals assumption, but what it does in particular is break the independence of errors assumption, since errors from the same unit (person, thing being observed) are correlated.