Most “which filter?” articles end in a single champion. Locked GSRF characterization does the opposite: job first, then tool.
One table chooser
| Job | Prefer | Locked reason |
|---|---|---|
| Mid-band ring / chatter calm | GSRF Practical | ~+35% osc31; ~−70% sweet-band power on phase6.2 packs |
| Peak chase toward a normal | GSRF Practical | ~38% peak-dev mean; 49/52 wins |
| Minimum MAD tracking of raw | EMA / Kalman | GSRF 0/52 MAD wins when spring on |
| Delayed plant as command feedback | Not balanced GSRF | +59% / +97% / +106% MAD worse at 3/6/10 min |
| Fixed-FA residual event detection | EMA residual / detector | Recall ~0.44 GSRF vs ~0.83 EMA |
| Probabilistic state + noise model | Kalman (or similar) | Different objective entirely |
| Permanent level shifts | Adaptive x* GSRF or EMA | Frozen x* ~9.7 → adaptive ~1.5 error probe |
Questions that decide
- Is the costly failure ring around a known normal, or lagging the truth?
- Is your “noise” mid-band (period ~50–120 samples) or slow drift you must track?
- Will the filtered stream drive a delayed closed loop?
- Is your KPI MAD-to-raw, peak excursion, or false-alarm rate?
Trap: “just replace EMA”
Dropping balanced GSRF into an EMA slot without checking delay, FA, and MAD will look worse on the metrics EMA was hired for. Read When GSRF lost first.
Verdict
If ring/spikes around a normal dominate → try GSRF (/try). If tracking/delay/detection dominate → EMA/Kalman/detector. Full table: /compare.
Sources: Evidence · compare · excellence hunt v2 · characterization v4 delay · probes v5 alarm