I'm using economic data to forecast banking crises. They occur in about 14% of the total possible samples.
Using basic summary statistics, one feature shows a banking crisis has taken place within 3 years of hitting a threshold 42% of the time (38 times out of 89 times if memory serves).
However, logistic regression only gave three instances where it predicted with greater than 40% probability that a crisis was forthcoming within 3 years. I'm not sure why that would be.
The LR results seem to drastically underprice the the probability.
The feature is debt and when debt is large and growing fast, it suggests money is being spent inefficienty and frivously and will inevitably go bad.
Regardless, any thoughts on why LR would be so low relative to the summary statistics and how to handle this better?