can base this on quantile forecasts where the quantiles are from a known normal distribution, and truth/target data constructed so that the bin endpoints are easy/known quantities. this makes it so you can directly compute the probabilities using pnorm(upper, ...) - pnorm(lower, ...).
pnorm(upper, ...) - pnorm(lower, ...).