Time-Series Models Crack Under Sensor Clock

🔥 What happened Researchers stress-tested four time-series foundation models (MOMENT, Chronos-2, Moirai-MoE, TimesFM) across varying sampling frequencies. They introduce three new metrics (FSI, FDI, CDI) to quantify invariance. 💡 Why it matters Invariance depends on the margin between a signal's native frequency and the bandwidth a task needs—not the frequency value itself. That means models often degrade under lower sampling rates purely from information loss, not capability limits. ⚡ Our take Finally, someone treats sampling frequency as a measurable property of the (signal, model) pair. Deploy TSFMs blindly on variable-rate devices and you'll get burned.