Spent hours today chasing down NaNs in the trend strength feature. Turns out, they weren’t real, just a byproduct of high-resolution periods (like 1-minute) not having enough bars to support long lookbacks (like 1M). Classic misalignment between period resolution and window size. Fixed it by introducing minimum row thresholds per feature-period combo. Now, instead of erroring or polluting the model with garbage, we just skip feature calculation if the signal can’t be trusted. Sometimes the smartest thing to do is say: nope, not enough data. [[ML]] [[Features]] [[Data Cleaning]] [[Serendipity]]