# How to restrict forecasts from a VAR model to be nonnegative?

I have two variables, assets and operating expenses year-ends for roughly 20 years. I would like to fit a VAR model with these two in order to forecast a couple years ahead.

1) The trouble is that operating expenses have been falling in general year on year. This leads to forecasts at t+2 turning negative which is obviously nonsensical. How can I restrict forecasts of those endogenous variables? In the ARMA framework, I would log-transform, but I am not sure if this is feasible or even advisable for a VAR model.

2) In general, I am concerned about the forecastability of those variables and I have not been very successful with ARMA and VAR forecasts. I have not tried Markov Chains, but coming from a frequentist background, I am not certain that if I devote a lot of time, I would yield results.

For reference only, I have used the vars package in R. Happy to try other R libraries. Please comment if any additional information is needed.

2. Since you have a small sample (just 20 data points), you could try some very simple models or models that do some regularization. If it is a VAR, go for low lag order, e.g. VAR(1), or for a Bayesian VAR if you have some idea about suitable priors; alternatively, try exponential smooting or auto.arima (functions ets and auto.arima from "forecast" package in R).