Does joining an index change what a stock is, or only how it trades? For the STOXX Europe 600 — 289 289 usable scheduled additions (cohort 421), G=46 event_windows.csv usable_c1 scheduled|add · 02_paper_numbers.md section 1 usable scheduled additions and 285 285 always quoted as '285 usable (365 survivors / 406 scheduled)' — never '285 survivors' event_windows.csv usable_c1 scheduled|del · 02_paper_numbers.md section 1 usable surviving demotions, 365 365 survivors among 406 scheduled deletions gate_0_3c_survivorship.csv scheduled|del · 02_paper_numbers.md section 1 survivors of 406 406 scheduled deletion cohort gate_0_3c_survivorship.csv scheduled|del · 02_paper_numbers.md section 1 scheduled deletions, across 46 46 review cycles clustering the headline battery (G=46); the AUM series spans 49 cycles 2014Q1–2026Q1 comovement_results.csv n_cycles · 02_paper_numbers.md section 1 quarterly review cycles — the added-stock beta rise is real at daily frequency and dies under every synchronicity correction, save one window in 2018–21. A companion to the index-rebalancing study.
Correction leg 3 / 5
Dimson ±1
By era · this leg
Middle 2018–2021: 0.08–0.13 median (0.12–0.17 mean) on every leg — marginal at G=16.
Middle 2018–2021 weekly carries two reads: cluster-t is not significant (p 0.086) while the median bootstrap CI [+0.064, +0.406] excludes zero.
| Section | Leg | Era | Median | 95% bootstrap CI | Mean | cluster-t (mean) | n | G | CI status |
|---|---|---|---|---|---|---|---|---|---|
| Pooled | Raw daily | all eras | +0.095 | [+0.032, +0.163] | +0.100 | 3.57 | 289 | 46 | CI excludes zero |
| Pooled | Peer-DiD (cum) | all eras | +0.102 | [+0.033, +0.149] | +0.102 | 4.46 | 289 | 46 | CI excludes zero |
| Pooled | Dimson ±1 | all eras | +0.014 | [−0.073, +0.102] | +0.024 | 0.61 | 288 | 46 | CI includes zero |
| Pooled | Dimson ±2 | all eras | +0.001 | [−0.097, +0.114] | +0.052 | 1.13 | 285 | 46 | CI includes zero |
| Pooled | Weekly (Wed–Wed) | all eras | +0.023 | [−0.062, +0.107] | +0.055 | 1.09 | 289 | 46 | CI includes zero |
| By era | Raw daily | Early 2014–2017 | +0.110 | [+0.038, +0.210] | +0.107 | 1.95 | 87 | 16 | CI excludes zero |
| By era | Dimson ±1 | Early 2014–2017 | −0.001 | [−0.208, +0.178] | −0.032 | −0.37 | 86 | 16 | CI includes zero |
| By era | Dimson ±2 | Early 2014–2017 | −0.045 | [−0.244, +0.212] | +0.060 | 0.58 | 85 | 16 | CI includes zero |
| By era | Weekly | Early 2014–2017 | +0.083 | [−0.067, +0.169] | +0.098 | 1.10 | 87 | 16 | CI includes zero |
| By era | Raw daily | Middle 2018–2021 | +0.133 | [+0.017, +0.270] | +0.139 | 2.33 | 106 | 16 | CI excludes zero |
| By era | Dimson ±1 | Middle 2018–2021 | +0.114 | [−0.046, +0.282] | +0.135 | 1.81 | 106 | 16 | CI includes zero |
| By era | Dimson ±2 | Middle 2018–2021 | +0.078 | [−0.087, +0.251] | +0.119 | 1.34 | 106 | 16 | CI includes zero |
| By era | Weekly | Middle 2018–2021 | +0.130 | [+0.064, +0.406] | +0.172 | 1.84 | 106 | 16 | CI excludes zero |
| By era | Raw daily | Late 2022– | +0.022 | [−0.019, +0.105] | +0.050 | 1.86 | 96 | 14 | CI includes zero |
| By era | Dimson ±1 | Late 2022– | −0.062 | [−0.154, +0.048] | −0.049 | −1.14 | 96 | 14 | CI includes zero |
| By era | Dimson ±2 | Late 2022– | −0.055 | [−0.120, +0.025] | −0.031 | −0.62 | 94 | 14 | CI includes zero |
| By era | Weekly | Late 2022– | −0.122 | [−0.191, −0.074] | −0.112 | −1.47 | 96 | 14 | CI excludes zero |
Pooled index-ex-i beta shift for scheduled STOXX Europe 600 additions across five correction legs, read top-down. Raw daily median +0.095 [+0.032, +0.163] and peer-DiD median +0.102 [+0.033, +0.149] have bootstrap confidence intervals excluding zero. Under Dimson ±1 the median is +0.014 [−0.073, +0.102], under Dimson ±2 +0.001 [−0.097, +0.114], and weekly +0.023 [−0.062, +0.107], each with a confidence interval including zero. By era, a battery-robust positive shift survives in Middle 2018–2021 (raw daily +0.133 [+0.017, +0.270], weekly +0.130 [+0.064, +0.406]); Middle cells are directional and marginal at G=16. Late 2022– weekly is exploratory.
The correction battery for scheduled additions. Top: the pooled median Δβin across 46 46 review cycles clustering the headline battery (G=46); the AUM series spans 49 cycles 2014Q1–2026Q1 comovement_results.csv n_cycles · 02_paper_numbers.md section 1 cycles, with review-cycle bootstrap intervals and the mean as an open marker — only the raw daily leg and the peer-DiD clear zero. Bottom: the same legs by era. In 2018–21 the magnitude holds on every leg while the two sanctioned reads disagree: the weekly median interval clears zero ([+0.064, +0.406] [+0.064, +0.406] 95% CI [+0.064, +0.406] directional, marginal at G=16; never “significant”. Weekly Middle carries two reads: cluster-t n.s. (p 0.086) vs median boot CI excl 0 — report both and the disagreement csw_results.csv full|middle|scheduled|add|uni_w|dbeta_in · 02_paper_numbers.md 3.1 ) while the cluster-t on the weekly mean does not reach conventional levels — sixteen cycles keep the window directional. The late-era weekly cell is exploratory.
The Sample
The Anatomy
At daily frequency the added-stock beta rise is a fact. The median event-level shift is +0.095 +0.095 95% CI [+0.032, +0.163] raw daily shift, scheduled adds comovement_results.csv full|scheduled|add|uni_d|dbeta_in · 02_paper_numbers.md 2.1 , with a review-cycle bootstrap interval of [+0.032, +0.163] [+0.032, +0.163] 95% CI [+0.032, +0.163] cycle-block bootstrap CI of the median comovement_results.csv full|scheduled|add|uni_d|dbeta_in · 02_paper_numbers.md 2.1 that clears zero, and the mean is +0.100 +0.100 raw daily shift, scheduled adds; mean and median are separate estimands (B-1); cluster-t is a mean statistic comovement_results.csv full|scheduled|add|uni_d|dbeta_in · 02_paper_numbers.md 2.1 at a cluster-t of 3.57 3.57 raw daily shift fires; cluster-t (mean), G=46 comovement_results.csv full|scheduled|add|uni_d|dbeta_in · 02_paper_numbers.md 2.1 . It also sits above all 2,000 cycle-conditional placebo draws (rank 1.000 1.000 rank evidence — cycle-conditional band, not cluster-robust robustness_matrix.csv 5_peer_placebo scheduled|add placebo_vs_unid · 02_paper_numbers.md 2.5 ) — no peer-pool artifact produces it. A matched-peer difference-in-differences leaves it intact — median +0.102 +0.102 95% CI [+0.033, +0.149] DiD rules out market-wide beta drift — never 'confirms the effect' did_results.csv full|scheduled|add|cum|did · 02_paper_numbers.md 2.1 , interval [+0.033, +0.149] [+0.033, +0.149] 95% CI [+0.033, +0.149] CI excludes 0 did_results.csv full|scheduled|add|cum|did · 02_paper_numbers.md 2.1 , cluster-t 4.46 4.46 peer-DiD (cum); cluster-t (mean), G=46 did_results.csv full|scheduled|add|cum|did · 02_paper_numbers.md 2.1 on the mean — which rules out market-wide beta drift, and nothing more.
None of that survives the corrections. Under Dimson lead–lag betas the median falls to +0.014 +0.014 CI incl. 0 95% CI [−0.073, +0.102] the raw rise dies under Dimson ±1 csw_results.csv full|all|scheduled|add|dimson1|dbeta_in · 02_paper_numbers.md 2.1 , its interval spanning zero ([−0.073, +0.102] [−0.073, +0.102] CI incl. 0 95% CI [−0.073, +0.102] CI includes 0 csw_results.csv full|all|scheduled|add|dimson1|dbeta_in · 02_paper_numbers.md 2.1 ); at weekly frequency it is +0.023 +0.023 CI incl. 0 95% CI [−0.062, +0.107] the raw rise dies at weekly frequency csw_results.csv full|all|scheduled|add|uni_w|dbeta_in · 02_paper_numbers.md 2.1 , interval [−0.062, +0.107] [−0.062, +0.107] CI incl. 0 95% CI [−0.062, +0.107] CI includes 0 csw_results.csv full|all|scheduled|add|uni_w|dbeta_in · 02_paper_numbers.md 2.1 . The anatomy says why. On the matched 288 288 matched-sample event count (events usable under both raw uni_d and Dimson ±1, usable scheduled adds, merged 1:1 on event_id) comovement_panel.csv + csw_panel.csv matched usable scheduled adds · 02_paper_numbers.md 2.3 -event sample the correction lifts the pre-inclusion beta from 0.950 0.950 matched-sample (n=288) mean pre-event β, raw daily comovement_panel.csv + csw_panel.csv matched usable scheduled adds · 02_paper_numbers.md 2.3 to 1.028 1.028 Dimson lifts the PRE-event β to ~1.03 (+0.0781): the pre-window was non-synchronous comovement_panel.csv + csw_panel.csv matched usable scheduled adds · 02_paper_numbers.md 2.3 and leaves the post-inclusion beta essentially unchanged (1.049 1.049 matched-sample (n=288) mean post-event β, raw daily comovement_panel.csv + csw_panel.csv matched usable scheduled adds · 02_paper_numbers.md 2.3 to 1.052 1.052 Dimson leaves the POST-event β unchanged (+0.0028) comovement_panel.csv + csw_panel.csv matched usable scheduled adds · 02_paper_numbers.md 2.3 ): the raw rise lives in the pre-window measurement, and the post-event beta never moved. The mechanism corroborates it — additions gain roughly +26.2% +26.2% exp(mean Δlog turnover)−1, printed as a MEAN shift; genuine (price-free) turnover rise on additions; n=278 liquidity_results.csv full|scheduled|add|dlog_turnover (recomputed exp()-1, asserted vs manifest section 5.1) · 02_paper_numbers.md 5.1 in genuine share turnover at the mean, so a stock that trades faster stops lagging the index by a day, lifting the daily beta while touching nothing at weekly frequency. This is the Chen–Singal–Whitelaw (2016) result on a two-sided European sample, with two legs that critique lacked: the matched-peer control and the turnover mechanism.
Dimson ±1 correction. Pre-event β 1.028 (+0.0781 vs raw); post-event β 1.052 (+0.0028 vs raw).
Hover or focus a row for its observed median and placebo band.
Dimson lifts the pre-event beta from ~0.95 to ~1.03 while leaving the post-event beta unchanged — the raw daily-beta rise is a pre-window non-synchronicity artifact.
Panel means (+0.100 raw, n=289 / +0.024 Dimson ±1, n=288) are a distinct per-event-average statistic from the matched-sample deltas — close in magnitude while resting on a different computation. The placebo band is a rank-based, cycle-conditional randomization band — rank evidence read for position rather than cluster-robust significance.
| Series | Estimator | β | Lift vs raw |
|---|---|---|---|
| β_pre | Raw OLS | 0.950 | — |
| β_pre | Dimson ±1 | 1.028 | +0.0781 |
| β_post | Raw OLS | 1.049 | — |
| β_post | Dimson ±1 | 1.052 | +0.0028 |
| Row | n | Observed median | Placebo band | Rank evidence |
|---|---|---|---|---|
| Add — raw daily (uni_d) | 289 | +0.095 | [−0.039, +0.044] | rank q = 1.000 |
| Add — Dimson ±1 | 289 | +0.014 | [−0.039, +0.044] | rank q = 0.669 |
| Del — raw daily (uni_d) | 285 | −0.040 | [−0.041, +0.048] | inside band |
Anatomy of the additions daily-beta rise. Left: a matched-sample slope chart (n=288). Under raw OLS the pre-event β is 0.950 and the post-event β is 1.049; the Dimson ±1 correction lifts the pre-event β to 1.028 (+0.0781) while the post-event β holds at 1.052 (+0.0028). Right: a placebo strip of three rows. Additions raw daily observed median +0.095 sits outside its cycle-conditional band [−0.039, +0.044] (rank q = 1.000); additions Dimson ±1 observed median +0.014 sits inside the same band (rank q = 0.669); deletions raw daily observed median −0.040 sits inside its band [−0.041, +0.048]. The placebo band is cycle-conditional rank evidence read for position rather than cluster-robust significance. Panel means (+0.100 raw, n=289 / +0.024 Dimson ±1, n=288) are a distinct per-event-average statistic from the matched-sample deltas — close in magnitude while resting on a different computation. The placebo band is a rank-based, cycle-conditional randomization band — rank evidence read for position rather than cluster-robust significance.
Left: the Dimson lead–lag correction lifts the pre-inclusion beta from 0.950 0.950 matched-sample (n=288) mean pre-event β, raw daily comovement_panel.csv + csw_panel.csv matched usable scheduled adds · 02_paper_numbers.md 2.3 to 1.028 1.028 Dimson lifts the PRE-event β to ~1.03 (+0.0781): the pre-window was non-synchronous comovement_panel.csv + csw_panel.csv matched usable scheduled adds · 02_paper_numbers.md 2.3 while the post-inclusion beta is essentially unchanged (1.049 1.049 matched-sample (n=288) mean post-event β, raw daily comovement_panel.csv + csw_panel.csv matched usable scheduled adds · 02_paper_numbers.md 2.3 to 1.052 1.052 Dimson leaves the POST-event β unchanged (+0.0028) comovement_panel.csv + csw_panel.csv matched usable scheduled adds · 02_paper_numbers.md 2.3 ) — the raw daily-beta rise is a pre-window non-synchronicity artifact. Panel betas are a distinct statistic from the matched-sample Δβ. Right: the observed median Δβin against the peer pseudo-event placebo band (2,000 draws, peers redrawn within fixed cycles); the raw addition median sits above every draw (rank 1.000 1.000 rank evidence — cycle-conditional band, not cluster-robust robustness_matrix.csv 5_peer_placebo scheduled|add placebo_vs_unid · 02_paper_numbers.md 2.5 ), the Dimson-corrected median mid-band (rank 0.669 0.669 rank evidence — cycle-conditional band, not cluster-robust; cross-estimand reference (placebo is raw-Δβ-based) robustness_matrix.csv 5_peer_placebo scheduled|add placebo_vs_dimson1 · 02_paper_numbers.md 2.5 ), the surviving-demotion median inside the band. The band is a cycle-conditional randomization — not cluster-robust — read as rank evidence only.
The Middle Era
The eras were cut before the fact — the companion volume's periodization, three windows fixed in advance rather than a searched breakpoint — so the “attenuation” question inherited from the U.S. literature comes back with a shape instead of a trend: null, then real, then gone. Two of the three eras behave like the pooled sample; the early rise collapses under Dimson, the late era runs flat to negative. The middle window, 2018 through 2021, is the exception. The mean shift holds its magnitude across the whole battery — +0.139 +0.139 directional, marginal at G=16; never “significant”; mean and median are separate estimands (B-1); cluster-t is a mean statistic comovement_results.csv full|scheduled|add|uni_d|dbeta_in · 02_paper_numbers.md 3.1 raw, +0.135 +0.135 CI incl. 0 directional, marginal at G=16; never “significant”; 'magnitude preserved' is a MEAN statement; mean and median are separate estimands (B-1); cluster-t is a mean statistic csw_results.csv full|middle|scheduled|add|dimson1|dbeta_in · 02_paper_numbers.md 3.1 under Dimson ±1, +0.119 +0.119 CI incl. 0 directional, marginal at G=16; never “significant”; 'magnitude preserved' is a MEAN statement; mean and median are separate estimands (B-1); cluster-t is a mean statistic csw_results.csv full|middle|scheduled|add|dimson2|dbeta_in · 02_paper_numbers.md 3.1 at ±2, +0.172 +0.172 directional, marginal at G=16; never “significant”. Weekly Middle carries two reads: cluster-t n.s. (p 0.086) vs median boot CI excl 0 — report both and the disagreement; mean and median are separate estimands (B-1); cluster-t is a mean statistic csw_results.csv full|middle|scheduled|add|uni_w|dbeta_in · 02_paper_numbers.md 3.1 weekly — and the Dimson correction that erases the pooled effect removes essentially nothing here. That is a statement about means: on the median the picture is less uniform, with Dimson ±2 shrinking to +0.078 +0.078 CI incl. 0 95% CI [−0.087, +0.251] directional, marginal at G=16; never “significant” csw_results.csv full|middle|scheduled|add|dimson2|dbeta_in · 02_paper_numbers.md 3.1 .
Significance is where sixteen review cycles show their limit, and the two sanctioned reads disagree. The weekly median's bootstrap interval clears zero ([+0.064, +0.406] [+0.064, +0.406] 95% CI [+0.064, +0.406] directional, marginal at G=16; never “significant”. Weekly Middle carries two reads: cluster-t n.s. (p 0.086) vs median boot CI excl 0 — report both and the disagreement csw_results.csv full|middle|scheduled|add|uni_w|dbeta_in · 02_paper_numbers.md 3.1 ) while the cluster-t on the weekly mean does not reach conventional levels (p 0.086 0.086 CI incl. 0 Middle-era weekly cluster-p (mean read), n.s. — disagrees with the median boot CI [+0.064, +0.406] excl 0; report both reads and the disagreement; directional, marginal at G=16; never “significant” csw_results.csv full|middle|scheduled|add|uni_w|dbeta_in · 02_paper_numbers.md 3.1 ); the Dimson-±1 median interval includes zero ([−0.046, +0.282] [−0.046, +0.282] CI incl. 0 95% CI [−0.046, +0.282] directional, marginal at G=16; never “significant” csw_results.csv full|middle|scheduled|add|dimson1|dbeta_in · 02_paper_numbers.md 3.1 ). Sixteen cycles keep the window directional — and its passive-ownership context runs backwards from a demand story. It sits on the flat segment of the passive-AUM path — a cycle mean of €11.0bn €11.0bn Middle = FLAT AUM segment where the battery-robust shift lives — alignment cuts AGAINST demand/habitat c5_attenuation.csv aum middle aum_teff_mean / aum_ramp_eur_m_per_year · 02_paper_numbers.md 3.4 , ramping +0.23bn/yr +0.23bn/yr flattest ramp of the three eras c5_attenuation.csv aum middle aum_teff_mean / aum_ramp_eur_m_per_year · 02_paper_numbers.md 3.4 a year — while the steepest ramp on record (+5.3bn/yr +5.3bn/yr steepest ramp of the three eras c5_attenuation.csv aum late aum_teff_mean / aum_ramp_eur_m_per_year · 02_paper_numbers.md 3.4 a year, €19.2bn €19.2bn Late = steepest ramp, where the corrected effect disappears c5_attenuation.csv aum late aum_teff_mean / aum_ramp_eur_m_per_year · 02_paper_numbers.md 3.4 mean) belongs to the late era, exactly where the corrected effect is gone. MiFID II, in January 2018, corroborates the timing and nothing more.
Hover, tap, or focus an era band for its corrected effect and passive-AUM alignment. Arrow keys cycle eras.
| Era | Leg | Median Δβ | 95% CI | n | G | CI status |
|---|---|---|---|---|---|---|
| Early 2014–2017 | Raw daily (uni_d) | +0.110 | [+0.038, +0.210] | 87 | 16 | excludes zero |
| Middle 2018–2021 | Raw daily (uni_d) | +0.133 | [+0.017, +0.270] | 106 | 16 | excludes zero |
| Late 2022– | Raw daily (uni_d) | +0.022 | [−0.019, +0.105] | 96 | 14 | includes zero |
| Early 2014–2017 | Dimson ±1 | −0.001 | [−0.208, +0.178] | 86 | 16 | includes zero |
| Middle 2018–2021 | Dimson ±1 | +0.114 | [−0.046, +0.282] | 106 | 16 | includes zero |
| Late 2022– | Dimson ±1 | −0.062 | [−0.154, +0.048] | 96 | 14 | includes zero |
| Era | Mean AUM | Ramp per year |
|---|---|---|
| Early 2014–2017 | €7.6bn | +1.6bn/yr |
| Middle 2018–2021 | €11.0bn | +0.23bn/yr |
| Late 2022– | €19.2bn | +5.3bn/yr |
Per-era median change in daily beta for STOXX Europe 600 additions (2014–2026), shown with cycle-block bootstrap confidence intervals, above the quarterly passive-AUM base tracking SXXP. Raw daily estimates (uni_d, filled circles) are positive and CI-clear of zero in the Early (median +0.110) and Middle (+0.133) eras and fall to +0.022 (CI includes zero) in the Late era; the Dimson ±1 correction (hollow diamonds) has confidence intervals that include zero in every era. The battery-robust Middle shift sits on the flattest passive-AUM segment (mean €11.0bn, ramp +0.23bn/yr), while the steepest AUM ramp — Late, mean €19.2bn, +5.3bn/yr — coincides with the corrected effect fading. The exploratory Δβ–log(AUM) coefficient is +0.035 (cluster-t 0.29), an alignment that runs against a demand or habitat reading. Era cells rest on 14–16 review cycles (G). MiFID II (2018-01) marks the Early/Middle seam as timing corroboration only.
Exploratory alignment: Δβ–log(AUM) coefficient +0.035, cluster-t 0.29 — exploratory; alignment, not causality.
Top: median Δβin by era for the raw and Dimson-corrected estimands, with bootstrap intervals over 14 to 16 cycles each — null, then real, then gone. Bottom: quarterly passive AUM tracking the index, with era means and ramps; the battery-robust 2018–21 window sits on the flat segment (+0.23bn/yr +0.23bn/yr flattest ramp of the three eras c5_attenuation.csv aum middle aum_teff_mean / aum_ramp_eur_m_per_year · 02_paper_numbers.md 3.4 a year), and the effect is gone during the steepest ramp (+5.3bn/yr +5.3bn/yr steepest ramp of the three eras c5_attenuation.csv aum late aum_teff_mean / aum_ramp_eur_m_per_year · 02_paper_numbers.md 3.4 a year). An exploratory regression of the corrected shift on log AUM is null (t 0.29 0.29 exploratory Δβ~log(AUM) t — alignment, not causality c5_attenuation.csv aum all log_aum · 02_paper_numbers.md 3.4 ). Alignment only, no causal claim; the MiFID II marker corroborates timing and nothing else.
Demotions
If inclusion re-wired a stock, demotion should un-wire it. It does not. Across 285 285 always quoted as '285 usable (365 survivors / 406 scheduled)' — never '285 survivors' event_windows.csv usable_c1 scheduled|del · 02_paper_numbers.md section 1 usable surviving demotions — 365 365 survivors among 406 scheduled deletions gate_0_3c_survivorship.csv scheduled|del · 02_paper_numbers.md section 1 survivors of 406 406 scheduled deletion cohort gate_0_3c_survivorship.csv scheduled|del · 02_paper_numbers.md section 1 scheduled deletions — every univariate leg is flat: raw daily, both Dimson corrections, weekly, and the peer-DiD, each with a bootstrap interval straddling zero, in every era. The claim stays deliberately narrow — no detectable comovement decline for surviving demotions. Demotions still move prices; the liquidity section shows where. One directional wrinkle is disclosed rather than buried: in the bivariate specification the raw daily mean is −0.124 −0.124 CI incl. 0 directional — cluster-t fires, boot CI incl 0, gone at weekly; mean and median are separate estimands (B-1); cluster-t is a mean statistic comovement_results.csv full|scheduled|del|bi_d|dbeta_in · 02_paper_numbers.md 4.2 at a cluster-t of −2.20 −2.20 CI incl. 0 directional — cluster-t fires, boot CI incl 0, gone at weekly; cluster-t (mean), G=46 comovement_results.csv full|scheduled|del|bi_d|dbeta_in · 02_paper_numbers.md 4.2 , but the median's bootstrap interval includes zero ([−0.178, +0.012] [−0.178, +0.012] CI incl. 0 95% CI [−0.178, +0.012] directional — cluster-t fires, boot CI incl 0, gone at weekly comovement_results.csv full|scheduled|del|bi_d|dbeta_in · 02_paper_numbers.md 4.2 ) and the cell is gone at weekly frequency — a daily-frequency shadow of the addition-side mechanics, failing the same battery the addition effect fails.
Two disclosures keep the null honest. The survivor sample under-represents the least liquid demotions — the names excluded below the observation floor traded at a median €1.08M €1.08M median PRE-window EUR volume of the excluded alive-below-threshold deletions — an order of magnitude below the survivor median 02_paper_numbers.md section 1 (CM-0003 census; hardcoded) against €6.79M €6.79M median PRE-window EUR volume of surviving scheduled deletions (mid-window survivors 4.97M) 02_paper_numbers.md section 1 (CM-0003 census; hardcoded) for survivors — so this is a statement about demotions liquid enough to keep estimating, the population a desk actually trades. And the re-add conditioning is closed empirically: readmitting the truncated demotions at a shorter observation floor augments the cell to 307 307 floor-60 augmented deletion cell (re-add conditioning retired) robustness_matrix.csv 6_readd_short_post readd_floor60_augmented · 02_paper_numbers.md 4.4 events and moves the median to −0.036 −0.036 CI incl. 0 95% CI [−0.100, +0.027] augmented-cell median Δβ_in, boot CI incl 0 — consistent with the 285-event deletion null robustness_matrix.csv 6_readd_short_post readd_floor60_augmented · 02_paper_numbers.md 4.4 , still through zero. The asymmetry is the point. Additions gain daily synchronicity and genuine turnover; surviving demotions lose neither — which is what a trading-synchronicity account predicts and a habitat account does not, since a demoted stock keeps trading and nothing mechanical reverses.
Six estimators, one verdict
Every median rests inside a bootstrap interval that includes zero. Hover or focus a row for its median, interval, mean and cluster-t (mean).
285 usable (365 survivors / 406 scheduled)
285 usable (365 survivors / 406 scheduled)
Hover or focus a bar or a drop bracket for the exclusion breakdown.
Selection diagnostic — median PRE-window EUR volume: excluded alive-below-threshold 1.08M vs survivors 6.79M (mid-window survivors 4.97M).
Floor-60 augmented variant (n=307): median −0.036, boot CI incl 0; re-add conditioning retired.
| Estimator | Median | 95% boot CI | Mean | cluster-t (mean) | p | n | G | CI status |
|---|---|---|---|---|---|---|---|---|
| Raw daily | −0.040 | [−0.114, +0.025] | −0.034 | −0.85 | 0.398 | 285 | 46 | CI includes zero |
| Peer-DiD (cum) | −0.039 | [−0.123, +0.012] | −0.034 | −1.05 | 0.300 | 285 | 46 | CI includes zero |
| Dimson ±1 | −0.020 | [−0.098, +0.062] | −0.025 | −0.41 | 0.686 | 285 | 46 | CI includes zero |
| Dimson ±2 | +0.014 | [−0.089, +0.144] | +0.019 | 0.29 | 0.772 | 283 | 46 | CI includes zero |
| Weekly (Wed–Wed) | −0.064 | [−0.151, +0.067] | −0.062 | −0.92 | 0.362 | 285 | 46 | CI includes zero |
| Bivariate daily Δβ_in (disclosed, separated) | −0.093 | [−0.178, +0.012] | −0.124 | −2.20 | 0.033 | 285 | 46 | CI includes zero; directional — cluster-t fires, CI incl 0, gone at weekly |
| Stage | n | Drop | Breakdown |
|---|---|---|---|
| scheduled deletions | 406 | — | base population |
| survivors | 365 | −41 | 17 structural (1 death at T_eff / 11 mid-window delistings / 5 alive-below-threshold); 24 right-censored |
| usable (comovement battery) | 285 | −80 | 48 PRE-150; 37 POST-150; 34 re-add-truncated (overlaps possible) |
Comovement-beta shift on index-excluding returns for 285 usable scheduled STOXX Europe 600 deletions (365 survivors / 406 scheduled), read across six estimators. Every estimator median sits inside a cycle-block bootstrap confidence interval that straddles zero — raw daily −0.040 [−0.114, +0.025] through weekly −0.064 [−0.151, +0.067]. The separated bivariate-daily row is directional (cluster-t of the mean −2.20) with a confidence interval that includes zero and no weekly counterpart. Below, the sample funnel: 406 scheduled deletions shed 41 non-survivors and 80 with insufficient observation windows to reach 285 usable.
Selection diagnostic: median PRE-window EUR volume, excluded alive-below-threshold 1.08M versus survivors 6.79M, mid-window survivors 4.97M. Floor-60 augmented variant, n 307: median −0.036, boot CI incl 0; re-add conditioning retired.
Top: surviving-demotion Δβin across the battery — every leg's bootstrap interval straddles zero; the bivariate daily cell, set apart, is the disclosed directional row whose cluster-t fires while its median interval includes zero and which is gone at weekly frequency. Bottom: the funnel — 285 285 always quoted as '285 usable (365 survivors / 406 scheduled)' — never '285 survivors' event_windows.csv usable_c1 scheduled|del · 02_paper_numbers.md section 1 usable (365 365 survivors among 406 scheduled deletions gate_0_3c_survivorship.csv scheduled|del · 02_paper_numbers.md section 1 survivors / 406 406 scheduled deletion cohort gate_0_3c_survivorship.csv scheduled|del · 02_paper_numbers.md section 1 scheduled) — with 41 41 406 scheduled − 365 survivors = 17 structural + 24 right-censored (2025H2–26); never bare '41 structural' gate_0_3c_survivorship.csv scheduled|del · 02_paper_numbers.md section 1 non-survivors (17 17 1 death at T_eff / 11 mid-window delistings / 5 alive-below-threshold gate_0_3c_survivorship.csv nonsurvivor_class · 02_paper_numbers.md section 1 structural, 24 24 recent 2025H2–26 right-censored gate_0_3c_survivorship.csv nonsurvivor_class · 02_paper_numbers.md section 1 right-censored by data end) and 80 80 48 fail PRE-150 via overlap clipping; 37 fail POST-150; 34 re-add-truncated; overlaps possible event_windows.csv scheduled|del survivors not usable_c1 · 02_paper_numbers.md section 1 dropped for insufficient observations (34 34 re-add-truncated among the 80 insufficient-obs survivors event_windows.csv post_truncated_by_readd · 02_paper_numbers.md section 1 of them re-add truncated). The liquidity-selection diagnostic is in the note.
Liquidity
On the price-free measure — share turnover — the two sides part company. Additions gain genuine trading: +26.2% +26.2% exp(mean Δlog turnover)−1, printed as a MEAN shift; genuine (price-free) turnover rise on additions; n=278 liquidity_results.csv full|scheduled|add|dlog_turnover (recomputed exp()-1, asserted vs manifest section 5.1) · 02_paper_numbers.md 5.1 more turnover at the mean, +24.0% +24.0% median event turnover rise; separate estimand from the +26.2% mean liquidity_results.csv full|scheduled|add|dlog_turnover (recomputed exp()-1, asserted vs manifest section 5.1) · 02_paper_numbers.md 5.1 at the median event. Surviving demotions lose none — a mean shift of −0.018 −0.018 CI incl. 0 deletion turnover null — price mechanics, not activity loss; mean and median are separate estimands (B-1); cluster-t is a mean statistic liquidity_results.csv full|scheduled|del|dlog_turnover · 02_paper_numbers.md 5.1 , cluster-t −0.43 −0.43 CI incl. 0 deletion turnover null; cluster-t (mean), G=45 liquidity_results.csv full|scheduled|del|dlog_turnover · 02_paper_numbers.md 5.1 , a genuine null. That asymmetry is the mechanism behind the beta result: an added stock that trades faster stops lagging the index; a demoted stock keeps trading, so nothing mechanical reverses.
The composite a first pass would headline — EUR volume — tells a different story only because it conflates activity with valuation. Decompose the log volume ratio into its additive legs (log turnover, log shares outstanding, log price) and the demotion-side decline loads almost entirely onto the price leg: demoted stocks are worth less, not traded less. The EUR-volume shift persists to the settled horizon on both sides, its bootstrap interval there clear of zero — but that horizon language attaches to the price-entangled measure alone; the price-free turnover result carries no horizon claim. Spreads are out of scope: this data layer has no quote data, which bounds what can be said about liquidity cost rather than activity.
Hover or focus any bar or horizon dot for its median, 95% CI, mean with cluster-t (mean), and sample.
| Measure | Side | Window | Median | 95% CI | Mean | cluster-t (mean) | n | Cycles | Robustness |
|---|---|---|---|---|---|---|---|---|---|
| Turnover (Δlog, price-free) | Additions | post vs pre | +0.215 | [+0.140, +0.273] | +0.233 | 3.99 | 278 | 45 | CI excludes zero |
| Turnover (Δlog, price-free) | Deletions | post vs pre | −0.026 | [−0.108, +0.017] | −0.018 | −0.43 | 278 | 45 | CI includes zero |
| Shares outstanding (Δlog) | Additions | post vs pre | +0.000 | [+0.000, +0.001] | +0.111 | 3.38 | 278 | 45 | CI includes zero |
| Shares outstanding (Δlog) | Deletions | post vs pre | +0.000 | [+0.000, +0.000] | +0.034 | 2.15 | 278 | 45 | CI includes zero |
| Price (Δlog) | Additions | post vs pre | +0.242 | [+0.192, +0.284] | +0.206 | 5.73 | 276 | 45 | CI excludes zero |
| Price (Δlog) | Deletions | post vs pre | −0.347 | [−0.407, −0.262] | −0.444 | −10.43 | 275 | 45 | CI excludes zero |
| EUR volume (Δlog, price-entangled) | Additions | post vs pre | +0.498 | [+0.396, +0.590] | +0.539 | 8.48 | 276 | 45 | CI excludes zero |
| EUR volume (Δlog, price-entangled) | Deletions | post vs pre | −0.414 | [−0.471, −0.347] | −0.431 | −9.95 | 275 | 45 | CI excludes zero |
| EUR volume ratio · ps1 [+1, +21] | Additions | [+1, +21] | +0.440 | [+0.355, +0.565] | +0.529 | 10.10 | 276 | 45 | CI excludes zero |
| EUR volume ratio · ps1 [+1, +21] | Deletions | [+1, +21] | −0.252 | [−0.312, −0.179] | −0.245 | −6.60 | 275 | 45 | CI excludes zero |
| EUR volume ratio · ps2 [+22, +63] | Additions | [+22, +63] | +0.464 | [+0.364, +0.562] | +0.550 | 9.27 | 276 | 45 | CI excludes zero |
| EUR volume ratio · ps2 [+22, +63] | Deletions | [+22, +63] | −0.284 | [−0.336, −0.207] | −0.262 | −5.76 | 275 | 45 | CI excludes zero |
| EUR volume ratio · ps3 [+64, +273] | Additions | [+64, +273] | +0.498 | [+0.405, +0.599] | +0.537 | 8.17 | 276 | 45 | CI excludes zero |
| EUR volume ratio · ps3 [+64, +273] | Deletions | [+64, +273] | −0.444 | [−0.505, −0.367] | −0.479 | −10.85 | 275 | 45 | CI excludes zero |
Liquidity decomposition for STOXX Europe 600 index additions and deletions, medians with cycle-block bootstrap 95% confidence intervals. Additions raise share turnover by +24.0% at the median (+26.2% mean, confidence interval excludes zero) while shares outstanding stay at zero. The EUR-volume composite is price-entangled: additions +0.498 and deletions −0.414 in log points, tracking price (+0.242 for additions, −0.347 for deletions). Deletion turnover is about zero (cluster-t (mean) −0.43, confidence interval includes zero), so the deletion EUR-volume decline loads on price rather than on activity. The additions-versus-deletions EUR-volume gap persists to the settled horizon window [+64, +273] (bootstrap confidence interval excludes zero). Hover or focus any bar or horizon dot for its median, confidence interval, mean with cluster-t (mean), and sample size.
Top: median log-shifts with bootstrap intervals for turnover, the shares and price legs, and the price-entangled EUR-volume composite (hatched), additions against demotions. Additions gain genuine turnover (+24.0% +24.0% median event turnover rise; separate estimand from the +26.2% mean liquidity_results.csv full|scheduled|add|dlog_turnover (recomputed exp()-1, asserted vs manifest section 5.1) · 02_paper_numbers.md 5.1 at the median); the demotion decline loads onto the price leg. Bottom: the EUR-volume shift across three post-effective horizons — it persists to the settled horizon, [+64, +273] days, on both sides, where its bootstrap interval excludes zero. Horizon language attaches to this price-entangled measure only.
Who Should Care
For a hedging desk
The post-inclusion rise in daily beta is real where execution, margin, and short-horizon hedges live — the stock trades faster and more synchronously with the benchmark. It is a measurement-horizon effect all the same: in lead–lag-corrected daily and weekly terms, the name is unchanged. Re-estimate hedges on daily windows if you trade there; at longer horizons there is nothing to chase.
For the demotion question
Leaving the index neither un-indexes a stock's risk nor dries up its trading. Surviving demotions show no detectable comovement decline and no turnover loss; their EUR-volume fall is a price-level effect. The book below a demoted name reprices without emptying — a distinction that matters for how an execution desk reads depth after a drop to mid-cap.
For the passive debate
Europe's most-tracked benchmark offers no evidence that indexing rewires fundamental comovement. The one battery-robust window sits on flat passive AUM and is gone during the fastest passive growth on record — an alignment the demand story predicts backwards at both ends. What indexing demonstrably changes is microstructure: how fast a stock trades and how quickly its price absorbs index-level information.
Paper & Method
Companion study: The Index Effect Before the Announcement