# Are boosted regression trees good for time series?

I am trying to use gbm to model a 21 year time series (n=1800 over the whole dataset) to understand an increase in my response variable over time. I have ~ 35 predictor variables which are made up of environmental variables (that also change with time; such as temperature, nitrogen, etc) and site (or dataset) specific variables (such as species, latitude, location, etc).

Initially, I was modeling the response variable against time (in years) along with all the other variables. However, this just tells me how much the response variable is changing with all the other variables - for example, we might expect a higher response at higher temperatures. As you might expects, time had the highest importance. Additionally, variables that don't change over time, such as latitude, are making it high up on the important predictor variables.

I am instead looking for a temporal explanation - what is causing this increase over time based off my predictors variables.
I know gbm can handle time series data but is it possible to run some sort of time series using gbm for inference instead of prediction?

• Why are you using gbm, vs say an ARIMA(p,d,q) model? – Jon Sep 21 '16 at 21:20
• @Jon I am used gbm to explain my response variable. I want to keep it consistent when trying to explain the temporal component. – Danib90 Sep 22 '16 at 14:18