EmpiricalCV

EmpiricalCV

CLI: --model cv_empirical. Class: loqculate.models.cv_empirical.EmpiricalCV.

No regression. LOD is always $\infty$ (supports_lod() is false). Weights are ignored.

Fit

Observations are grouped by unique concentration. Per group, CV is Bessel-corrected std / mean (vectorized_cv_stats). Groups with fewer than 2 observations get nan CV. If every group has fewer than 2 observations, fit() raises ValueError.

min_replicates default 3: concentrations with fewer replicates emit UserWarning and are still used.

params_ is {}.

LOQ

Positive concentrations only (blanks at $x=0$ are dropped).

find_loq_threshold on the sorted unique positive calibration levels, not on an interpolated grid. Default sliding_window=3, cv_thresh=0.20, non-strict <=. The window is consecutive levels.

Predict

predict(x_new) linearly interpolates group means. Outside the calibration range it holds the nearest edge mean (np.interp).

CLI

--sliding_window is not forwarded to this class. A CLI cv_empirical run uses constructor sliding_window=3. --cv_thresh is used, because loq(cv_thresh) reads the argument.

compute_loqs_bulk exists on the class for a vectorized many-peptide pass. The CLI fit path does not call it; it constructs one EmpiricalCV per peptide.

What the notebook actually showed

comparison_report.ipynb compared EmpiricalCV and PiecewiseWLS (not PiecewiseCF). Findings recorded there, not a universal ranking:

• EmpiricalCV FDR on null curves depended strongly on replicate count. At cv_thresh=0.20, window 3, FDR was high at $n=2$ (the notebook prints 49% in that panel) and reached 0% at $n \ge 10$ in that experiment.
• PiecewiseWLS with window 3 already had FDR $< 1%$ at $n=2$ in the same notebook.
• Original CV (window 1) was described there as 64-99% FDR across concentration-grid densities.

I treat EmpiricalCV as a model-free screen. I do not treat those FDR figures as a property of PiecewiseCF; CF was not in that notebook.