Time series Algorithm suggestion I have a univariate timeseries data which is always cumulative. So the trend is always upwards. I want to build timeseries forecast for future forecast. Currently i have 5 different models running: Auto-Arima, Holtswinter, Holtsmethod, FBprophet and LSTM. I want to replace LSTM with a simple timeseries model. Can you suggest me some? I tried searching and tried XGboost and random forest, but output was giving flat forecast
 A: I agree with the comment, Xgboost and random forests are NOT anywhere near my definition of 'simple'. And besides, tree models cannot forecast outside of the known data so the known presence of a trend means trees won't work plug-and-play.
For simple time series I would look at naive methods like the last value or if you have a seasonal pulse of 12 then just repeat the last 12 values. Additionally, you can use the mean or median. Although you may want to test with detrending since these are 'flat' forecasts.
I would say the most 'advanced' simple model in my eyes would be early time series decomposition methods such as the classical method (seasonal_decompose in statsmodels) which is just a centered moving average and average seasonal indices.
These methods will probably produce results that minimize your error a surprising amount of times but may not be 'useful' forecasts.
A: As you have a time series which is non-stationary, your most important step will be to stationarise the time series as much as possible. It's difficult to recommend precisely without seeing your data but potentially:

*

*Difference the data to make it non-cumulative.

*Subtract mean and divide by standard deviation.

Given a more appropriate time series target (and that your problem is univariate), I think the most simple methods would be any autoregressive-based method (e.g. ARIMA).
