- **`SyntheticControl` cross-validation + inverse-variance `V`-selection (ADH 2015 §; Abadie 2021 §3.2(a), Eq. 9).** Two new `v_method` values complete the ADH-2015/Abadie-2021 `V`-selection menu (joining `"nested"` / `"custom"`), each threaded through the in-space / leave-one-out / in-time placebo refits so a diagnostic uses the **same** estimator as the headline fit. **`v_method="cv"`** selects the diagonal predictor-importance `V` by out-of-sample cross-validation: the pre-period is split positionally at `v_cv_t0` (new constructor param; default `len(pre)//2`, Abadie 2021's `t0 = T0/2`) into a training and a validation window, `V` is chosen to minimize the validation-window outcome MSPE of the training-fit weights (`mspe_v` now reports this validation MSPE under cv), and the final reported weights are re-estimated on the validation-window predictors (ADH 2015 step 4). Each predictor spec is **re-aggregated** over each window (its mean/sum/identity recomputed over only the periods that fall in that window — a separate `dataprep` per window, exactly as ADH 2015's CV does, since R `Synth` has no built-in CV function), so the V-search is genuinely out-of-sample for every predictor type and the same `V*` drives both fits with no zeroed coordinate (`v_weights` reproduce `donor_weights` on the validation-window predictors, and `predictor_balance` is reported on that validation-window basis). **Fully-spanning precondition (fail-closed):** re-aggregating a predictor on each window requires it to be observed in **both** windows, so `cv` **requires every predictor to span both the training and validation windows** and raises `ValueError` otherwise — satisfied by ADH 2015's shared covariate / multi-period `special_predictors` (which span the windows) but NOT by the default per-period outcome lags (each is single-period and lives in one window only), so `cv` with the bare default predictors is rejected with guidance to pass spanning predictors. In-time-placebo truncation that breaks the fully-spanning precondition (a kept spec stops spanning both windows at the truncated split) marks that date `infeasible`. A second fail-closed gate covers windows that span but carry **no cross-donor variation** (every re-aggregated predictor constant across the donors, so `X0·W` is constant in `W` → a flat, unidentified weight solve that would otherwise return arbitrary "converged" weights — even when the treated unit differs, since donor distinguishability, not treated-vs-donor variation, identifies `W`): the headline fit raises `ValueError`, in-space placebo refits whose donor pool is indistinguishable in a window are dropped from the reference set, and such in-time-truncated dates are marked `infeasible`. Abadie 2021 footnote 7's CV non-uniqueness is handled by a **deterministic tie-break** (prefer the `V` closest to uniform among ties), making the selected `V*` among equally-good optima independent of the multistart evaluation order. The cv fit is reproducible for a fixed `seed` (like `nested`) but is not seed-independent — the multistart fills any slots beyond the distinct heuristic starts with seed-dependent random Dirichlet draws, so the tie-break removes start-order dependence among ties, not seed dependence. The tie-break is convergence-aware (a non-converged optimizer candidate cannot displace a converged incumbent on an objective tie). If the training-window solve that defines `mspe_v` truncates (e.g. `inner_max_iter` too small), the fit fails closed — `mspe_v=NaN` and the fit is marked non-converged — rather than reporting an invalid Eq. 9 criterion. **`v_method="inverse_variance"`** uses the closed form `v_h = 1/Var(X_h)` (variance over donors+treated on the unstandardized predictors), applied to the **raw** predictors so the effective objective is the unit-variance-rescaled `Σ_h diff_h²/Var_h` (Abadie 2021 §3.2(a)); the `standardize` pre-scaling is intentionally bypassed on this branch (inverse-variance weighting *is* the unit-variance rescaling — applying it on already-standardized rows would double-rescale to `Σ_h diff_h²/Var_h²`), so it is equivalent to uniform `V` on standardized predictors. No search (`mspe_v=None`); a zero-variance row gets 0 weight and an all-zero-variance panel falls back to uniform `V` with a warning. `custom_v` is rejected (fail-closed) for both methods and `v_cv_t0` is rejected unless `v_method="cv"`. On the degenerate **single-donor** path (`J=1` forces `w=[1]`) `V` is unidentified — every `V` yields the same synthetic — so `v_weights` is **uniform** and `mspe_v=None` for ALL `v_method`s (cv / inverse_variance included; their selected / closed-form `V` would be inert), with a `UserWarning`; the donor weights / gap / ATT are unaffected. An explicitly pinned `v_cv_t0` that no longer fits the truncated pre-fake window is nulled to the `//2` default for the placebo refit (a pinned value that still fits the truncated window is kept). **Validation:** R `Synth` has no built-in CV function (ADH 2015's CV is a manual `dataprep`+`synth` re-run), so cv is anchored by deterministic equivalence to the R-anchored `custom_v` path (the step-3 validation MSPE of the training-window fit and the step-4 validation-window weights each match a `custom_v=V*` fit on the correspondingly re-aggregated predictors) plus cv self-consistency (`in_time_placebo` under cv == a fresh cv fit on the backdated panel to 1e-7); inverse-variance is anchored bit-for-bit to a `custom_v=1/Var(X)` fit. Documented in `docs/methodology/REGISTRY.md` §SyntheticControl (new `**Note:**` labels for the per-window re-aggregation convention, the flat-MSPE tie-break, and inverse-variance), `docs/api/synthetic_control.rst`, the LLM guides, and `README.md`. The remaining ADH-2015 items (`W^reg` extrapolation diagnostic, sparse-SC subset search) stay tracked in `TODO.md`.
0 commit comments