Techniques to improve accuracy of time series use case I am working on forecasting weekly revenue of 10,000 sectors. Applied basic time series models and average RMSE(in thousands) on hold-out set (last 32 weeks) as below. 
In my view, ma12 and ETS(ANN) models seems reasonable. 
1. Please let me know techniques to reduce RMSE. 
2. Which other time series models can be tried ?

 ma4       ma8       ma12      stlf      ses       holt      ets-ANN   ets-ZZN   ets-ZZN12 linear
 18.79894  18.65662  18.78054  21.34638  19.66270  19.27242  18.17565  18.52025  20.08258  25.14404 


Box Plot.

Violin plot.


Update (Plot for two sample sectors)


 A: 
In my view, ma12 and ETS(ANN) models seems reasonable.

Judging by the plots all your models have equal quality, statistically identical.


  
*Which other time series models can be tried ?
  

Not sure if you have tried ARIMA. It's also not clear how exactly you applied the linear model.
After looking at raw data charts:
I think you could try more steps on top of what you have done:


*

*Replace values that show a clear pattern of being out of the line
with the rest of the time series with mean/median of neighbours. This
may (bot not must) improve a model quality.

*Combine all you data together (take average all
over the sales points for each week). Apply very detailed auto.arima to this one
meta series, or, even better, build a model manually. It can be that
one model (one set of ARIMA params) will serve well for all of your
time series.


In order to deal with different scales of sales for different points, I would do this: for each time series take x_diff == diff(log(X)), than take mean(x_diff ) across all series, and then restore original scale.


*

*Make a better decision of whether your linear model is additive or
multiplicative. It may be that the variance in different years is
strongly not constant, thus additive model (lm(y ~ t + s)) may work
poorly.

*Maybe you want to smooth the time series by applying a running mean
window in order to deal with outliers.

