I am using a probit model to predict US recessions one month ahead. I am using monthly US Treasury Spread, Employment Rate, and S&P500 annualized returns from Jan-1966 to Jan-2020 as independent variables, using a US recession indicator as my binary dependent variable.

When deciding which lags of each variable to use, I am unsure of the most appropriate method to use. Currently, I have chosen to use Bayesian Information Criterion to select my lag lengths before running. I then run a probit regression using each lag up until the lag specified by BIC for each of my independent variables. i.e. if BIC suggests 5 lags for Employment, I use each lag up to the fifth in my regression.

Much of the literature, such as Estrella and Hardouvelis (1991) and Estrella and Mishkin (1995) choose lag length based upon their forecasting period rather than through information criterion. For instance, a lag length of three months is used to predict recession three months ahead. Additionally, these papers include only the lag specified, rather than each lag up to the specified lag. Contrarily, Kauppi and Saikkonen (2008) use 'statistical methods' to choose lag length, but do not specify which methods.

I am therefore asking for advice on the most appropriate way to choose lag length for my model.

Any help would be much appreciated.