panelkit 0.2.3__tar.gz → 0.2.5__tar.gz

{panelkit-0.2.3 → panelkit-0.2.5}/Cargo.lock

@@ -462,7 +462,7 @@ checksum = "d6790f58c7ff633d8771f42965289203411a5e5c68388703c06e14f24770b41e"
 
 [[package]]
 name = "panelkit-estimators"
-version = "0.2.3"
+version = "0.2.5"
 dependencies = [
  "criterion",
  "panelkit-linalg",
@@ -471,7 +471,7 @@ dependencies = [
 
 [[package]]
 name = "panelkit-geo"
-version = "0.2.3"
+version = "0.2.5"
 dependencies = [
  "panelkit-estimators",
  "panelkit-inference",
@@ -482,7 +482,7 @@ dependencies = [
 
 [[package]]
 name = "panelkit-inference"
-version = "0.2.3"
+version = "0.2.5"
 dependencies = [
  "panelkit-estimators",
  "panelkit-linalg",
@@ -491,7 +491,7 @@ dependencies = [
 
 [[package]]
 name = "panelkit-linalg"
-version = "0.2.3"
+version = "0.2.5"
 dependencies = [
  "proptest",
  "rayon",
@@ -623,7 +623,7 @@ dependencies = [
 
 [[package]]
 name = "pypanelkit"
-version = "0.2.3"
+version = "0.2.5"
 dependencies = [
  "numpy",
  "panelkit-estimators",

{panelkit-0.2.3 → panelkit-0.2.5}/Cargo.toml

@@ -3,7 +3,7 @@ resolver = "2"
 members = ["crates/linalg", "crates/estimators", "crates/inference", "crates/geo", "crates/pypanelkit"]
 
 [workspace.package]
-version = "0.2.3"
+version = "0.2.5"
 edition = "2021"
 rust-version = "1.74"
 license = "MIT OR Apache-2.0"

{panelkit-0.2.3 → panelkit-0.2.5}/GUIDE.md

@@ -300,10 +300,16 @@ ev.plot_effect_over_time("effect.png")  # pointwise + cumulative over time, w/ C
 ev.lift, ev.cumulative, ev.significant
 ```
 
-Each estimate gets a confidence interval from a **stationary block bootstrap** of
-its post-period effect path; an **SC in-space placebo** supplies a p-value. The
-ensemble uses the same `weights` choices as `power()` (`"auto"` = inverse-variance
-from each method's bootstrap SE, `"equal"`, or an explicit dict/list). `ev` exposes
+Inference is **in-space placebo** (Abadie): every donor market is refit as if it
+were the treated one, and the spread of *their* post-period effects is the null
+reference — capturing out-of-sample extrapolation error, the real source of
+uncertainty. (A bootstrap of the treated unit's own post-period only sees
+in-sample noise and is wildly anti-conservative — on null data its 90% interval
+falsely flags an effect ~50% of the time; the placebo version sits at/below the
+nominal 10%.) Poorly-fit placebos (pre-period RMSPE > 2× the treated unit's) are
+dropped, per Abadie. The p-value is the placebo rank of the treated effect, and
+`"auto"` ensemble weights are inverse-variance from each method's placebo-null
+spread. `ev` exposes
 `.lift`, `.att`, `.cumulative`, `.significant`, the per-method results in `ev.per`,
 and the ensemble in `ev.ensemble`. Reported numbers: **% lift** (effect ÷
 counterfactual), **per-period ATT**, and **cumulative incremental** over the
@@ -315,13 +321,13 @@ you can see it sits flat (centered on zero) inside the noise band before the tes
 starts (a placebo check) and breaks out after — and the running **cumulative
 incremental**, each as a point estimate with a confidence band. The counterfactual
 is centered on the pre-period, so the gap shows fit quality rather than a level
-offset (SDID matches trends, not levels). The bands come from a **moving-block
-bootstrap** of the pre-period residuals: resampling whole blocks preserves their
-autocorrelation, so the intervals are more conservative than an iid normal
-approximation — the cumulative band in particular widens faster than √k when the
-residuals are positively autocorrelated. Raise `block_len` to capture longer-range
-dependence (wider, more conservative cumulative bands). Pass `exclude=[…]` to drop
-markets from the control pool (e.g. ones you don't trust as donors).
+offset (SDID matches trends, not levels). The bands come from the **in-space
+placebo** distribution: at each horizon, the pointwise band is the spread of the
+donor placebos' per-period effects, and the cumulative band is the spread of their
+cumulative sums (so it fans out with horizon). Placebo inference needs a decent
+donor pool to have power — with only a handful of comparable donors the intervals
+are necessarily wide. Pass `exclude=[…]` to drop markets from the control pool
+(e.g. ones you don't trust as donors).
 
 ### Choosing a specification — `design.recommend(test_lengths, n_geos_options, target_lift, alphas=…)`
 

{panelkit-0.2.3 → panelkit-0.2.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: panelkit
-Version: 0.2.3
+Version: 0.2.5
 Classifier: Programming Language :: Rust
 Classifier: Programming Language :: Python :: 3
 Classifier: Topic :: Scientific/Engineering
@@ -273,8 +273,8 @@ per-cell MDE/confidence/holdout report and a combined figure:
 **Evaluate a test that ran.** `evaluate(...)` is the measurement counterpart to
 the power analysis: fit SC / ASC / SDID on a test that already happened, blend
 them into a weighted-average **ensemble** estimate, and report each one's lift,
-confidence interval (stationary block bootstrap), and cumulative incremental —
-with an SC in-space placebo p-value:
+confidence interval (in-space placebo), and cumulative incremental —
+with an in-space placebo p-value:
 
 ![test evaluation](assets/geo_evaluate.png)
 
@@ -316,7 +316,7 @@ What you get out of the box:
 - **A weighted-average ensemble** of SC + ASC + SDID (combined per placebo window,
   with auto inverse-variance weights) for a steadier estimate than any one method.
 - **Post-test evaluation** — `evaluate()` measures a test that already ran:
-  per-method + ensemble lift, bootstrap CIs, cumulative incremental, and a p-value.
+  per-method + ensemble lift, in-space placebo CIs, cumulative incremental, and a p-value.
 
 See [`examples/geo_demo.py`](examples/geo_demo.py).
 

{panelkit-0.2.3 → panelkit-0.2.5}/README.md

@@ -243,8 +243,8 @@ per-cell MDE/confidence/holdout report and a combined figure:
 **Evaluate a test that ran.** `evaluate(...)` is the measurement counterpart to
 the power analysis: fit SC / ASC / SDID on a test that already happened, blend
 them into a weighted-average **ensemble** estimate, and report each one's lift,
-confidence interval (stationary block bootstrap), and cumulative incremental —
-with an SC in-space placebo p-value:
+confidence interval (in-space placebo), and cumulative incremental —
+with an in-space placebo p-value:
 
 ![test evaluation](assets/geo_evaluate.png)
 
@@ -286,7 +286,7 @@ What you get out of the box:
 - **A weighted-average ensemble** of SC + ASC + SDID (combined per placebo window,
   with auto inverse-variance weights) for a steadier estimate than any one method.
 - **Post-test evaluation** — `evaluate()` measures a test that already ran:
-  per-method + ensemble lift, bootstrap CIs, cumulative incremental, and a p-value.
+  per-method + ensemble lift, in-space placebo CIs, cumulative incremental, and a p-value.
 
 See [`examples/geo_demo.py`](examples/geo_demo.py).
 

{panelkit-0.2.3 → panelkit-0.2.5}/crates/estimators/src/sc/sdid.rs

@@ -87,6 +87,10 @@ pub fn fit_at(panel: &Panel, t0: usize, cfg: SdidConfig) -> ScFit {
     let t = panel.n_periods();
     let t_pre = t0;
     let t_post = t - t0;
+    assert!(
+        t_pre >= 1 && t_post >= 1,
+        "SDID needs at least one pre- and one post-period (t0 in 1..n_periods)"
+    );
     let n_tr = treated.len();
 
     // Treated-average series.

{panelkit-0.2.3 → panelkit-0.2.5}/crates/linalg/src/opt/simplex.rs

@@ -30,17 +30,14 @@ pub fn project_simplex(v: &[f64]) -> Vec<f64> {
     let mut u = v.to_vec();
     u.sort_by(|a, b| b.partial_cmp(a).unwrap()); // descending
     let mut css = 0.0;
-    let mut rho = 0usize;
     let mut theta = 0.0;
     for (j, &uj) in u.iter().enumerate() {
         css += uj;
         let t = (css - 1.0) / (j as f64 + 1.0);
         if uj - t > 0.0 {
-            rho = j + 1;
             theta = t;
         }
     }
-    let _ = rho;
     v.iter().map(|&vi| (vi - theta).max(0.0)).collect()
 }
 

{panelkit-0.2.3 → panelkit-0.2.5}/crates/pypanelkit/src/api_sc.rs

@@ -111,13 +111,15 @@ pub fn fit_sdid(
 /// Fit Matrix-Completion NNM (Athey et al. 2021). `max_rank`, when set, uses a
 /// fast randomized truncated SVD inside SoftImpute (big speedup, low-rank cap).
 #[pyfunction]
-#[pyo3(signature = (y, treated, treat_time, lambda=None, max_iter=200, tol=1e-5, seed=0, max_rank=None))]
+// `lambda_` (not `lambda`) so it is usable as a Python keyword argument —
+// `lambda` is a reserved word in Python.
+#[pyo3(signature = (y, treated, treat_time, lambda_=None, max_iter=200, tol=1e-5, seed=0, max_rank=None))]
 #[allow(clippy::too_many_arguments)]
 pub fn fit_mcnnm(
     y: PyReadonlyArray2<f64>,
     treated: Vec<usize>,
     treat_time: usize,
-    lambda: Option<f64>,
+    lambda_: Option<f64>,
     max_iter: usize,
     tol: f64,
     seed: u64,
@@ -125,7 +127,7 @@ pub fn fit_mcnnm(
 ) -> PyResult<PyScResult> {
     let panel = Panel::block(mat_from_numpy(&y), &treated, treat_time);
     let cfg = McnnmConfig {
-        lambda,
+        lambda: lambda_,
         max_iter,
         tol,
         seed,

{panelkit-0.2.3 → panelkit-0.2.5}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "maturin"
 
 [project]
 name = "panelkit"
-version = "0.2.3"
+version = "0.2.5"
 description = "Fast, from-scratch causal-inference estimators for panel/geo experiments (SC, ASC, SDID, DiD, MC-NNM)."
 readme = "README.md"
 requires-python = ">=3.9"

{panelkit-0.2.3 → panelkit-0.2.5}/python/panelkit/_panelkit.pyi

@@ -82,7 +82,7 @@ def fit_mcnnm(
     y: npt.NDArray[np.float64],
     treated: Sequence[int],
     treat_time: int,
-    lambda_: Optional[float] = ...,
+    lambda_: Optional[float] = ...,  # NOTE: matches the Rust binding's `lambda_`
     max_iter: int = ...,
     tol: float = ...,
     seed: int = ...,

{panelkit-0.2.3 → panelkit-0.2.5}/python/panelkit/design.py

@@ -42,7 +42,8 @@ def _ensemble_weight_arg(spec):
         raise ValueError(f"unknown ensemble_weights {spec!r} (use 'auto', 'equal', "
                          "a dict, or a 3-list)")
     if isinstance(spec, dict):
-        w = [float(spec.get(m, spec.get(m.lower(), 0.0))) for m in _ENSEMBLE_ORDER]
+        norm = {str(k).upper(): v for k, v in spec.items()}  # case-insensitive keys
+        w = [float(norm.get(m, 0.0)) for m in _ENSEMBLE_ORDER]
     else:
         w = [float(x) for x in spec]
         if len(w) != 3:
@@ -52,26 +53,6 @@ def _ensemble_weight_arg(spec):
     return w
 
 
-def _placebo_paths(pre_gaps, length, block_len, n_reps, seed):
-    """Moving-block bootstrap of the (centered) pre-period residuals into placebo
-    paths of ``length`` periods. Resampling whole blocks preserves the residual
-    autocorrelation, so the resulting CI bands are more conservative than an iid
-    normal approximation. Returns an ``(n_reps, length)`` array (empty if no
-    pre-period or zero length)."""
-    g = np.asarray(pre_gaps, dtype=float)
-    m = len(g)
-    if m == 0 or length <= 0 or n_reps <= 0:
-        return np.empty((0, max(length, 0)))
-    g = g - g.mean()  # null is "no effect" → center the residuals
-    rng = np.random.default_rng(int(seed))
-    bl = max(1, min(int(block_len), m))
-    n_blocks = int(np.ceil(length / bl))
-    starts = rng.integers(0, m, size=(n_reps, n_blocks))
-    idx = (starts[:, :, None] + np.arange(bl)[None, None, :]) % m  # circular blocks
-    paths = g[idx].reshape(n_reps, n_blocks * bl)[:, :length]
-    return paths
-
-
 class _PowerReport:
     """Result of a power analysis across methods, with a report and plots."""
 
@@ -427,7 +408,7 @@ class GeoDesign:
                              target_power=target_power, recommended=recommended,
                              lookback=lookback, ensemble=ensemble,
                              ensemble_weights=ensemble_weights)
-        idx = self._resolve(treated)
+        idx = list(dict.fromkeys(self._resolve(treated)))  # dedup, preserve order
         names = [self.names[i] for i in idx]
         lifts = list(_DEFAULT_LIFTS if lifts is None else lifts)
         if 0.0 not in lifts:
@@ -463,7 +444,7 @@ class GeoDesign:
             if bad:
                 raise ValueError(f"treated markets were also excluded: {bad}")
             return sub.diagnose(tnames, test_len)
-        idx = self._resolve(treated)
+        idx = list(dict.fromkeys(self._resolve(treated)))  # dedup, preserve order
         names = [self.names[i] for i in idx]
         t0 = self.t - int(test_len)
         diag = _panelkit.geo_diagnostics(self.Y, idx, int(test_len))
@@ -701,8 +682,7 @@ class GeoDesign:
         methods: Sequence[str] = _METHODS,
         weights="auto",
         level: float = 0.90,
-        n_boot: int = 2000,
-        block_len: int = 4,
+        max_placebo: int = 200,
         seed: int = 0,
         exclude=None,
     ) -> "_EvalReport":
@@ -711,9 +691,15 @@ class GeoDesign:
         This is the measurement counterpart to :meth:`power`: given the treated
         markets and the period treatment began (``treat_start``, the first
         post-period column), it fits SC / ASC / SDID, reports each one's effect,
-        and combines them into a weighted-average **ensemble** estimate. Each
-        estimate gets a confidence interval from a stationary block bootstrap of
-        its post-period effect path; an SC in-space placebo supplies a p-value.
+        and combines them into a weighted-average **ensemble** estimate.
+
+        Inference is **in-space placebo** (Abadie): every donor market is refit as
+        if it were the treated one, and the spread of *their* post-period effects
+        is the null reference. This captures out-of-sample extrapolation error —
+        the dominant source of uncertainty — so the intervals are calibrated
+        (unlike a bootstrap of the treated unit's own post-period, which only sees
+        in-sample noise and is far too narrow). Poorly-fit placebos (pre-period
+        RMSPE > 2× the treated unit's) are dropped, per Abadie.
 
         Parameters
         ----------
@@ -725,11 +711,13 @@ class GeoDesign:
             Which estimators to fit and blend.
         weights : "auto" | "equal" | dict
             Ensemble weighting. ``"auto"`` is inverse-variance (precision)
-            weighting from each method's bootstrap standard error.
+            weighting from each method's placebo-null spread.
         level : float
             Confidence level for the intervals (e.g. 0.90).
-        n_boot, block_len, seed :
-            Stationary-bootstrap settings for the effect-path CIs.
+        max_placebo : int
+            Cap on the number of donor placebos used (sampled if exceeded).
+        seed : int
+            Seed for placebo sampling when ``max_placebo`` is exceeded.
 
         Returns
         -------
@@ -745,8 +733,8 @@ class GeoDesign:
             if bad:
                 raise ValueError(f"treated markets were also excluded: {bad}")
             return sub.evaluate(tnames, treat_start, methods=methods, weights=weights,
-                                level=level, n_boot=n_boot, block_len=block_len, seed=seed)
-        idx = self._resolve(treated)
+                                level=level, max_placebo=max_placebo, seed=seed)
+        idx = list(dict.fromkeys(self._resolve(treated)))  # dedup, preserve order
         names = [self.names[i] for i in idx]
         t0 = int(treat_start)
         if not (1 <= t0 < self.t):
@@ -757,27 +745,28 @@ class GeoDesign:
         if unknown:
             raise ValueError(f"unknown methods {unknown}; choose from {_METHODS}")
 
-        fitters = {
-            "SC": lambda: _panelkit.fit_sc(self.Y, idx, t0, 0.0, False, level),
-            "ASC": lambda: _panelkit.fit_asc(self.Y, idx, t0, 0.0, None),
-            "SDID": lambda: _panelkit.fit_sdid(self.Y, idx, t0, 1.0),
-        }
+        def _fit(method, tr):
+            if method == "SC":
+                return _panelkit.fit_sc(self.Y, tr, t0, 0.0, False, level)
+            if method == "ASC":
+                return _panelkit.fit_asc(self.Y, tr, t0, 0.0, None)
+            return _panelkit.fit_sdid(self.Y, tr, t0, 1.0)
+
         treated_series = self.Y[idx].mean(axis=0)
+        post_len = self.t - t0
+        order = methods
+
+        # --- point estimates on the treated set ---
         per = {}
         for m in methods:
-            fit = fitters[m]()
+            fit = _fit(m, idx)
             att_path = np.asarray(fit.att_path, dtype=float)
             cf = np.asarray(fit.counterfactual, dtype=float)
             att = float(fit.att)
             cf_mean = float(np.mean(cf)) if cf.size else float("nan")
-            se, lo, hi = _panelkit.bootstrap_mean(
-                att_path.tolist(), "stationary", int(block_len), int(n_boot),
-                int(seed), float(level))
-            # Full-timeline counterfactual via donor weights (exact for SC; the
-            # dominant term for ASC/SDID). Center on the pre-period so the gap
-            # reflects FIT, not a level offset — SDID is level-agnostic (matches
-            # trends, not levels), so its donor-weighted series sits at a constant
-            # offset that would otherwise look like a non-zero pre-period.
+            # Full-timeline counterfactual via donor weights, centered on the
+            # pre-period so the gap reflects FIT, not a level offset (SDID matches
+            # trends, not levels).
             dids = np.asarray(fit.donor_ids, dtype=int)
             ws = np.asarray(fit.weights, dtype=float)
             if dids.size:
@@ -787,25 +776,40 @@ class GeoDesign:
                 full_cf = np.full(self.t, np.nan)
             per[m] = {
                 "att": att, "att_path": att_path, "counterfactual": cf,
-                "full_cf": full_cf,
-                "cf_mean": cf_mean, "lift": att / cf_mean if cf_mean else float("nan"),
-                "se": se, "att_lo": lo, "att_hi": hi,
-                "lift_lo": lo / cf_mean if cf_mean else float("nan"),
-                "lift_hi": hi / cf_mean if cf_mean else float("nan"),
+                "full_cf": full_cf, "cf_mean": cf_mean,
+                "lift": att / cf_mean if cf_mean else float("nan"),
                 "cumulative": float(att_path.sum()) * n_treated,
                 "pre_rmspe": float(fit.pre_rmspe),
             }
 
-        # Ensemble: weight-average the post-period effect paths, then summarize.
-        order = methods
+        # --- in-space placebo: refit each donor as if it were treated ---
+        treated_set = set(idx)
+        donors = [u for u in range(self.n) if u not in treated_set]
+        if len(donors) > int(max_placebo):
+            rng = np.random.default_rng(int(seed))
+            donors = sorted(int(j) for j in rng.choice(donors, int(max_placebo), replace=False))
+        pb = {m: [] for m in methods}        # per method: list of (att_path, pre_rmspe)
+        for j in donors:
+            for m in methods:
+                fj = _fit(m, [j])
+                pb[m].append((np.asarray(fj.att_path, dtype=float), float(fj.pre_rmspe)))
+
+        # --- ensemble weights ---
+        def _placebo_att_sd(m):
+            if not pb[m]:
+                return 1.0
+            vals = np.array([p.mean() for (p, _) in pb[m]])
+            return float(np.std(vals)) if len(vals) > 1 else 1.0
         if isinstance(weights, str) and weights.lower() == "equal":
             wv = [1.0 / len(order)] * len(order)
         elif isinstance(weights, str) and weights.lower() == "auto":
-            prec = [1.0 / max(per[m]["se"] ** 2, 1e-300) for m in order]
+            # inverse-variance from each method's placebo-null spread (precision)
+            prec = [1.0 / max(_placebo_att_sd(m) ** 2, 1e-300) for m in order]
             s = sum(prec)
             wv = [p / s for p in prec] if s > 0 else [1.0 / len(order)] * len(order)
         elif isinstance(weights, dict):
-            raw = [float(weights.get(m, weights.get(m.lower(), 0.0))) for m in order]
+            norm = {str(k).upper(): v for k, v in weights.items()}  # case-insensitive
+            raw = [float(norm.get(m, 0.0)) for m in order]
             s = sum(raw)
             if s <= 0:
                 raise ValueError("ensemble weights must sum to > 0")
@@ -817,71 +821,103 @@ class GeoDesign:
             s = sum(raw)
             wv = [r / s for r in raw]
         wmap = dict(zip(order, wv))
+        a = (1.0 - float(level)) / 2.0
 
+        def _ci(point, null_samples):
+            """Pivot CI: point estimate ± the placebo null spread (null ≈ 0).
+            Returns NaN when there are too few placebos to form an interval —
+            never a fake zero-width CI."""
+            if len(null_samples) >= 2:
+                return point + float(np.quantile(null_samples, a)), \
+                    point + float(np.quantile(null_samples, 1.0 - a))
+            return float("nan"), float("nan")
+
+        def _kept_att(samples, treated_pre_m):
+            """Placebo att-means after the Abadie 2x pre-fit filter (fallback to
+            all placebos if too few comparable ones survive)."""
+            keep = [p.mean() for (p, pre) in samples
+                    if treated_pre_m <= 0 or pre <= 2.0 * treated_pre_m]
+            if len(keep) < 5 and samples:
+                keep = [p.mean() for (p, _) in samples]
+            return np.array(keep)
+
+        # --- per-method point CIs from each method's placebo att spread (same
+        #     2x pre-fit filter as the ensemble, for internal consistency) ---
+        for m in order:
+            mp = _kept_att(pb[m], per[m]["pre_rmspe"])
+            lo, hi = _ci(per[m]["att"], mp)
+            cfm = per[m]["cf_mean"]
+            per[m]["att_lo"], per[m]["att_hi"] = lo, hi
+            per[m]["lift_lo"] = lo / cfm if cfm else float("nan")
+            per[m]["lift_hi"] = hi / cfm if cfm else float("nan")
+
+        # --- ensemble estimate + ensemble placebo paths (Abadie pre-fit filter) ---
         ens_path = sum(wmap[m] * per[m]["att_path"] for m in order)
         ens_cf_mean = float(sum(wmap[m] * per[m]["cf_mean"] for m in order))
         ens_att = float(ens_path.mean())
-        se, lo, hi = _panelkit.bootstrap_mean(
-            ens_path.tolist(), "stationary", int(block_len), int(n_boot),
-            int(seed), float(level))
+        treated_pre = sum(wmap[m] * per[m]["pre_rmspe"] for m in order)
+
+        ens_pb = []  # (path, pre_rmspe)
+        for di in range(len(donors)):
+            path = sum(wmap[m] * pb[m][di][0] for m in order)
+            pre = sum(wmap[m] * pb[m][di][1] for m in order)
+            ens_pb.append((path, pre))
+        kept = [p for (p, pre) in ens_pb if treated_pre <= 0 or pre <= 2.0 * treated_pre]
+        if len(kept) < 5:                      # too few comparable placebos → use all
+            kept = [p for (p, _) in ens_pb]
+        pb_mat = np.array(kept) if kept else np.zeros((0, post_len))
+        n_pb = pb_mat.shape[0]
+
+        # pointwise + cumulative + mean CIs, all from the placebo null
+        if n_pb >= 2:
+            point_lo = ens_path + np.quantile(pb_mat, a, axis=0)
+            point_hi = ens_path + np.quantile(pb_mat, 1.0 - a, axis=0)
+            point_hw = float(np.quantile(np.abs(pb_mat), float(level)))
+            cum_pb = np.cumsum(pb_mat, axis=1)
+            run = np.cumsum(ens_path)
+            cum_lo_band = np.quantile(cum_pb, a, axis=0)
+            cum_hi_band = np.quantile(cum_pb, 1.0 - a, axis=0)
+            pb_att = pb_mat.mean(axis=1)
+            p_value = float((1.0 + np.sum(np.abs(pb_att) >= abs(ens_att))) / (1.0 + n_pb))
+        else:
+            # too few comparable placebos → inference undefined (no fake band)
+            run = np.cumsum(ens_path)
+            point_lo = np.full(post_len, np.nan)
+            point_hi = np.full(post_len, np.nan)
+            point_hw = 0.0
+            cum_lo_band = cum_hi_band = np.full(post_len, np.nan)
+            pb_att = np.array([])
+            p_value = None
+        att_lo, att_hi = _ci(ens_att, pb_att)
+
+        cum_curve = run * n_treated
         ensemble = {
-            "att": ens_att, "att_path": ens_path, "se": se,
-            "att_lo": lo, "att_hi": hi,
+            "att": ens_att, "att_path": ens_path,
+            "att_lo": att_lo, "att_hi": att_hi,
             "lift": ens_att / ens_cf_mean if ens_cf_mean else float("nan"),
-            "lift_lo": lo / ens_cf_mean if ens_cf_mean else float("nan"),
-            "lift_hi": hi / ens_cf_mean if ens_cf_mean else float("nan"),
+            "lift_lo": att_lo / ens_cf_mean if ens_cf_mean else float("nan"),
+            "lift_hi": att_hi / ens_cf_mean if ens_cf_mean else float("nan"),
             "cumulative": float(ens_path.sum()) * n_treated,
-            "weights": wmap,
+            "weights": wmap, "n_placebo": n_pb,
+            "low_power": n_pb < 8,   # too few placebos for reliable inference
         }
 
-        # Significance: SC in-space placebo p-value.
-        sc = _panelkit.fit_sc(self.Y, idx, t0, 0.0, True, level)
-        p_value = sc.p_value
-
-        # Full-timeline ensemble counterfactual + gap path (pre-period shows fit,
-        # post-period uses the exact ensemble effect).
+        # full-timeline counterfactual + gap path (pre shows fit; post = effect)
         ens_full_cf = sum(wmap[m] * per[m]["full_cf"] for m in order)
         full_gap = treated_series - ens_full_cf
-        full_gap[t0:] = ens_path                       # exact ensemble post effect
-        counterfactual = treated_series - full_gap     # consistent everywhere
-        pre_gaps = full_gap[:t0]
-        sigma_pre = float(np.std(pre_gaps, ddof=1)) if t0 > 1 else float(np.std(pre_gaps))
-
-        # CI bands from a MOVING-BLOCK BOOTSTRAP of the pre-period residuals.
-        # Blocks preserve autocorrelation, so the bands are more conservative than
-        # an iid normal approximation — especially the cumulative band, whose
-        # spread grows faster than sqrt(k) under positive autocorrelation.
-        post_len = self.t - t0
-        a = (1.0 - float(level)) / 2.0
-        paths = _placebo_paths(pre_gaps, post_len, int(block_len), int(n_boot), int(seed))
-        if paths.size:
-            point_lo = np.quantile(paths, a, axis=0)
-            point_hi = np.quantile(paths, 1.0 - a, axis=0)
-            point_hw = float(np.quantile(np.abs(paths), float(level)))  # symmetric, full-timeline
-            cum_paths = np.cumsum(paths, axis=1)
-            cum_band_lo = np.quantile(cum_paths, a, axis=0)
-            cum_band_hi = np.quantile(cum_paths, 1.0 - a, axis=0)
-        else:
-            point_lo = point_hi = np.zeros(post_len)
-            point_hw = 0.0
-            cum_band_lo = cum_band_hi = np.zeros(post_len)
-
-        ens_post = ens_path
-        run = np.cumsum(ens_post)
-        cum_curve = run * n_treated
-        cum_lo_curve = (run + cum_band_lo) * n_treated
-        cum_hi_curve = (run + cum_band_hi) * n_treated
-
-        ensemble["sigma_pre"] = sigma_pre
+        full_gap[t0:] = ens_path
+        counterfactual = treated_series - full_gap
         ensemble["full_gap"] = full_gap
-        ensemble["point_hw"] = point_hw                       # constant pointwise half-width
-        ensemble["point_lo"] = ens_post + point_lo            # per-period CI on the effect
-        ensemble["point_hi"] = ens_post + point_hi
-        ensemble["cum_curve"] = cum_curve                     # cumulative incremental path
-        ensemble["cum_lo_curve"] = cum_lo_curve
-        ensemble["cum_hi_curve"] = cum_hi_curve
-        ensemble["cum_lo"] = float(cum_lo_curve[-1]) if post_len else float("nan")
-        ensemble["cum_hi"] = float(cum_hi_curve[-1]) if post_len else float("nan")
+        ensemble["sigma_pre"] = (float(np.std(full_gap[:t0], ddof=1)) if t0 > 1
+                                 else float(np.std(full_gap[:t0])))
+        ensemble["point_hw"] = point_hw
+        ensemble["point_lo"] = point_lo
+        ensemble["point_hi"] = point_hi
+        ensemble["cum_curve"] = cum_curve
+        ensemble["cum_lo_curve"] = (run + cum_lo_band) * n_treated
+        ensemble["cum_hi_curve"] = (run + cum_hi_band) * n_treated
+        ensemble["cum_lo"] = float(ensemble["cum_lo_curve"][-1]) if post_len else float("nan")
+        ensemble["cum_hi"] = float(ensemble["cum_hi_curve"][-1]) if post_len else float("nan")
 
         return _EvalReport(names, t0, n_treated, per, ensemble, p_value, level,
                            treated_series, counterfactual)
@@ -981,11 +1017,14 @@ class _MultiCellReport:
                      f"({', '.join(map(str, self.cells))})")
         lines.append(f"Test duration     : {self.test_len} periods")
         lines.append(f"Shared donor pool : {len(self.donor_names)} markets")
-        lines.append(f"Combined holdout  : {100*self.pooled_holdout:.1f}% of total volume")
+        lines.append(f"Combined holdout  : {100*self.pooled_holdout:.1f}% of total volume "
+                     f"(all cells together)")
         lines.append(f"Powered at {int(100*self.target_power)}% power, "
                      f"{int(100*(1-self.alpha))}% confidence "
                      f"(each cell vs. the shared pool).")
         lines.append("")
+        # Per-cell 'Holdout' is that cell's share of its OWN sub-panel (cell +
+        # shared donors); the Combined holdout above is over the full panel.
         lines.append(f"{'Cell':<14}{'Markets':<28}{'MDE':>8}{'Conf':>7}{'Holdout':>9}")
         lines.append("-" * 64)
         for label, rep in self.cells.items():
@@ -1050,8 +1089,11 @@ class _EvalReport:
 
     @property
     def significant(self):
-        """True if the ensemble CI excludes zero (effect detected)."""
+        """True if the ensemble CI is well-defined and excludes zero. Returns
+        False when inference is undefined (too few placebos → NaN interval)."""
         lo, hi = self.ensemble["att_lo"], self.ensemble["att_hi"]
+        if not (np.isfinite(lo) and np.isfinite(hi)):
+            return False
         return (lo > 0) or (hi < 0)
 
     def summary(self) -> str:
@@ -1073,17 +1115,25 @@ class _EvalReport:
         lines.append(f"   ensemble weights: {wstr}")
         lines.append("")
         if self.p_value is not None:
-            lines.append(f"SC in-space placebo p-value : {self.p_value:.3f}")
-        verdict = ("✓ Significant lift — the ensemble interval excludes zero."
-                   if self.significant else
-                   "~ Not distinguishable from zero at this level — the ensemble "
-                   "interval includes zero.")
+            lines.append(f"In-space placebo p-value    : {self.p_value:.3f}  "
+                         f"(ensemble, {e.get('n_placebo', 0)} donors)")
+        if e.get("low_power"):
+            lines.append("⚠ Few comparable donors — inference is low-powered; treat "
+                         "intervals/p-value with caution.")
+        if self.significant:
+            verdict = "✓ Significant lift — the ensemble interval excludes zero."
+        elif not (np.isfinite(e["att_lo"]) and np.isfinite(e["att_hi"])):
+            verdict = ("? Inference undefined — too few comparable donor placebos "
+                       "to form an interval.")
+        else:
+            verdict = ("~ Not distinguishable from zero at this level — the ensemble "
+                       "interval includes zero.")
         lines.append(f"Headline (ensemble)         : {100*e['lift']:+.2f}% lift, "
                      f"{e['cumulative']:,.0f} cumulative incremental")
         if "cum_lo" in e:
             lines.append(f"Cumulative {cl}% CI          : "
                          f"[{e['cum_lo']:,.0f}, {e['cum_hi']:,.0f}]  "
-                         f"(moving-block bootstrap, block_len-aware)")
+                         f"(in-space placebo, {e.get('n_placebo', 0)} donors)")
         lines.append(verdict)
         lines.append("=" * 66)
         return "\n".join(lines)
@@ -1569,7 +1619,7 @@ def _plot_eval(rep: "_EvalReport", path):
     axc.set_title("Lift by method", fontweight="bold")
     axc.grid(True, axis="x", alpha=0.25)
 
-    pv = f"  ·  SC placebo p={rep.p_value:.3f}" if rep.p_value is not None else ""
+    pv = f"  ·  placebo p={rep.p_value:.3f}" if rep.p_value is not None else ""
     verdict = "significant" if rep.significant else "not significant"
     fig.suptitle(f"panelkit · test evaluation — ensemble lift "
                  f"{100*rep.ensemble['lift']:+.2f}% ({verdict}){pv}",
@@ -1582,10 +1632,10 @@ def _plot_eval(rep: "_EvalReport", path):
 def _plot_eval_timeline(rep: "_EvalReport", path):
     """Pointwise + cumulative effect over the full timeline, with CI bands.
 
-    Bands come from a moving-block bootstrap of the pre-period residuals (so they
-    capture autocorrelation): the pointwise band is the per-period placebo spread
-    around the estimate; the cumulative band grows with horizon as the bootstrap
-    placebo cumulative-sums spread out."""
+    Bands come from the in-space placebo distribution (every donor refit as if
+    treated): the pointwise band is the per-period placebo spread around the
+    estimate; the cumulative band grows with horizon as the placebo
+    cumulative-sums spread out."""
     _, plt = _require_mpl()
     import numpy as _np
     from matplotlib.gridspec import GridSpec
@@ -1632,7 +1682,7 @@ def _plot_eval_timeline(rep: "_EvalReport", path):
     cum = e["cum_curve"]
     axc.axvspan(-0.5, t0 - 0.5, color="#f3f4f6", alpha=0.8)
     axc.fill_between(seg, e["cum_lo_curve"], e["cum_hi_curve"], color=_PK_GREEN,
-                     alpha=0.15, label=f"{cl}% band (block bootstrap)")
+                     alpha=0.15, label=f"{cl}% band (in-space placebo)")
     axc.plot(seg, cum, color=_PK_GREEN, lw=2.4, label="cumulative incremental")
     axc.axhline(0, color="#111827", lw=1.0)
     axc.axvline(t0 - 0.5, color="#374151", lw=1.2, ls=":")