I am working on a time series forecasting project, and I have monthly data with a seasonality of 12 months. The goal is to forecast the next 24 months. I applied SARIMA and Facebook Prophet, yet the results aren’t really satisfactory.

I added a step of feature engineering: I added lagged variables and moving window statistics variables, and the results on the test are impressive. However, my problem now is how to forecast the next 24 months when I need these engineered features. I will need to use the forecasted values of N+1 to forecast N+2, N+3, etc. I am afraid the error will grow exponentially! Can anyone help me with this issue?