A survivorship-free event study of 1,219 1,219 Constituent-change events in the study panel. 02_paper_numbers.md §1 STOXX Europe 600 rebalancing events, 2014–2026 — finding the effect alive at selection, dead at execution, and cheap for the funds that track it.
Announcement day (T_ann)
Window · Pre-announcement run-up
Window stat · additions +410 bps [+290, +552], placebo q = 1.00
Window stat · deletions −476 bps [−598, −276], placebo q = 0.00
T_sl→T_ann−1: +410 bps [+290, +552], placebo q = 1.00
T_sl→T_ann−1: −476 bps [−598, −276], placebo q = 0.00
T_ann→T_eff: −62 bps [−172, +9] CI incl. 0
T_eff+1→+42: −165 bps [−284, +79] CI incl. 0
| Window | Side | Median CAAR | 95% CI | Robustness |
|---|---|---|---|---|
| T_sl→T_ann−1 | additions | 410 bps | [290, 552] | CI excludes zero |
| T_sl→T_ann−1 | deletions | -476 bps | [-598, -276] | CI excludes zero |
| T_ann→T_eff | additions | -62 bps | [-172, 9] | CI includes zero |
| T_eff+1→+42 | additions | -165 bps | [-284, 79] | CI includes zero |
Event-time median cumulative abnormal return for 1,219 STOXX Europe 600 rebalancing events (2014-2026). Additions run up to about +410 bps between the selection list and the announcement, with a confidence interval that excludes zero, then go flat. Deletions fall about −476 bps over the same window. The announcement-to-effective give-back (−62 bps) and the post-effective reversal (−165 bps) both have confidence intervals that include zero. Use the slider to scrub the trading day.
The Evidence Base
Medians, not means; COVID-2020 and sanctions tails are flagged, not scrubbed.
The Finding · Q1
The +410 +410 95% CI [+290, +552] Pre-announcement run-up (selection-list). Placebo q=1.00; bootstrap CI excludes 0. 02_paper_numbers.md §2 / caar_q2_five_window.csv bps pre-announcement run-up is not one number spread evenly across the names that join the index.
Names already certain to enter — the core — move −11 −11 CI incl. 0 95% CI [−152, +154] Core (non-marginal) additions -- near zero, CI includes 0. 02_paper_numbers.md §2 / caar_q1_buffer.csv bps, statistically nothing, while the marginal names that resolve the list's remaining uncertainty move +389 +389 95% CI [+176, +600] Marginal (buffer) additions, T_ann-10 -> T_eff. Artifact median 388.5 bps; the effect concentrates here, not in core adds. 02_paper_numbers.md §2 / caar_q1_buffer.csv bps. This is resolution of uncertainty about inclusion, not pure front-running.
By buffer status
By predictability
Headline contrast — hover or focus any row
+388.5 bps [+176, +600]
Filled = marginal names · hollow = core names. Faded dots: confidence interval includes zero.
| Panel | Row | Side | Median (bps) | 95% CI (bps) | placebo q | n | cycles | Robustness |
|---|---|---|---|---|---|---|---|---|
| A · buffer status | Core — add | Additions | −10.6 | [−152, +154] | 0.51 | 188 | 47 | CI includes zero |
| A · buffer status | Buffer — add | Additions | +388.5 | [+176, +600] | 1.00 | 118 | 31 | CI excludes zero |
| A · buffer status | Core — del | Deletions | −282.2 | [−678, +140] | 0.002 | 56 | 31 | CI includes zero |
| A · buffer status | Buffer — del | Deletions | −340.6 | [−485, −131] | 0.00 | 278 | 46 | CI excludes zero |
| B · predictability | Predicted — add | Additions | −10.6 | [−152, +156] | 0.51 | 188 | 47 | CI includes zero |
| B · predictability | Surprise — add | Additions | +388.5 | [+176, +588] | 1.00 | 118 | 31 | CI excludes zero |
| B · predictability | Predicted — del | Deletions | −282.2 | [−900, +123] | 0.002 | 56 | 31 | CI includes zero |
| B · predictability | Surprise — del | Deletions | −292.8 | [−438, −128] | 0.00 | 289 | 46 | CI excludes zero |
Q1 buffer decomposition. Two dot-and-whisker panels over a shared basis-point axis show the median cumulative abnormal return in the premium window (T_ann−10 to T_eff). Panel A splits additions and deletions by buffer status (core vs buffer); Panel B by predictability (predicted vs surprise). The headline contrast: buffer additions earn +388.5 bps (95% CI +176 to +600, excludes zero) while core additions earn −10.6 bps (95% CI −152 to +154, includes zero) — the abnormal return concentrates in the marginal names whose index fate was unresolved at the selection-list date. Rows whose confidence interval includes zero are drawn as faded dots and are not interpretable.
Panel A splits additions and deletions into core and marginal (buffer) names; Panel B recasts the same split as predicted versus surprise. The premium concentrates in the marginal, uncertainty-resolving names — core moves carry confidence intervals through zero.
Cluster-honest inference · Q2 / Q3 / Q5
Once the announcement prints, the tradeable residual is statistical noise: the give-back (−62 −62 CI incl. 0 95% CI [−172, +9] Announcement-to-effective give-back. Never lead with this: CI includes 0, placebo q=0.351. 02_paper_numbers.md §2 / caar_q2_five_window.csv bps), the announcement build-up (−103 −103 CI incl. 0 95% CI [−189, +9] Announcement build-up (front_run window) -- NOT front-running; CI includes 0. 02_paper_numbers.md §2 / caar_q2_five_window.csv bps) and the post-effective reversal (−165 −165 CI incl. 0 95% CI [−284, +79] Post-effective reversal, additions -- CI includes 0. 02_paper_numbers.md §2 / caar_q2_five_window.csv bps) all carry confidence intervals that include zero. Under review-cycle clustering, 12 of 31 12 of 31 Scheduled-SXXR cells naive-significant but cluster-CI includes 0. 02_paper_numbers.md §8 naively significant cells die.
Ask the same panel to predict rather than describe, and it can’t: out-of-sample R² is negative everywhere — an honest null.
| Window | Side | Median CAAR (bps) | 95% CI (bps) | Placebo q | n | Robustness | Note |
|---|---|---|---|---|---|---|---|
| Pre-selection window | Additions | +93 | [+52, +144] | 0.999 | 338 | CI excludes zero | Mechanical artifact (overlaps beta-estimation window) |
| Pre-selection window | Deletions | −95 | [−182, −17] | 0.000 | 358 | CI excludes zero | Mechanical artifact (overlaps beta-estimation window) |
| Pre-announcement run-up | Additions | +410 | [+290, +552] | 1.000 | 317 | CI excludes zero | |
| Pre-announcement run-up | Deletions | −476 | [−598, −276] | 0.000 | 325 | CI excludes zero | |
| Announcement build-up | Additions | −103 | [−189, +9] | 0.090 | 327 | CI includes zero | |
| Announcement build-up | Deletions | −121 | [−306, +18] | 0.049 | 342 | CI includes zero | |
| MOC window | Additions | +7 | [−34, +49] | 0.368 | 355 | CI includes zero | |
| MOC window | Deletions | +45 | [0, +108] | 0.996 | 364 | CI includes zero | |
| Post-effective reversal | Additions | −165 | [−284, +79] | 0.047 | 316 | CI includes zero | |
| Post-effective reversal | Deletions | +12 | [−292, +217] | 0.715 | 336 | CI includes zero |
Five-window structure of median cumulative abnormal return for STOXX Europe 600 scheduled index reviews. Additions run up about +410 bps and deletions fall about −476 bps over the pre-announcement run-up (both confidence intervals exclude zero). The announcement build-up, market-on-close and post-effective reversal windows all have confidence intervals that include zero. The pre-selection window (+93 / −95 bps) overlaps the beta-estimation window and is a mechanical artifact.
Median cumulative abnormal return across the five event windows, additions and deletions, under review-cycle clustering. The announcement build-up, the market-on-close window and the post-effective reversal each carry a confidence interval through zero.
246 events · one point per event, coloured by arbitrage-risk tercile
negative everywhere — an honest null
Hedged long–short Sharpe ≈ 0 across models (OLS +0.13, Lasso −0.16, XGBoost +0.11) — no tradable signal.
| Tercile | Count | Median slippage |
|---|---|---|
| Low | 82 | +0.47% |
| Mid | 82 | −0.31% |
| High | 82 | −0.19% |
| Model | Sample | OOS R² | Decile Spearman | n | Hedged Sharpe |
|---|---|---|---|---|---|
| OLS | Full sample | −0.034 | +0.006 | 183 | +0.128 |
| OLS | Decontaminated | −0.025 | +0.103 | 180 | +0.128 |
| Lasso | Full sample | −0.011 | −0.176 | 183 | −0.161 |
| Lasso | Decontaminated | −0.011 | −0.297 | 180 | −0.161 |
| XGBoost | Full sample | −0.193 | −0.770 | 183 | +0.109 |
| XGBoost | Decontaminated | −0.167 | −0.345 | 180 | +0.109 |
Panel A plots execution slippage (percent) against order flow (in days of ADV) for 246 index-addition events, coloured by arbitrage-risk tercile. A LOWESS fit (frac 0.6) stays essentially flat near zero across the whole flow range, and a quadratic term is insignificant under year fixed effects (p = 0.91; LOWESS Spearman −0.015): there is no robust convexity in flow. The one place flow bites is its interaction with arbitrage risk — the flow-by-arb-risk cross-partial is +0.029 per standard deviation (p = 0.0015). Panel B shows out-of-sample R-squared for OLS, Lasso and XGBoost, each trained full-sample and decontaminated, on a 2014–2019 train / 2020–2025 test split. All six values are negative — every model does worse out of sample than simply predicting the training mean — and hedged long-short Sharpe ratios are near zero. Slippage is structural noise, not a predictable cost.
Slippage against flow, by arbitrage-risk tercile, with the out-of-sample prediction scorecard. Structure is visible in-sample; out-of-sample R² stays below the naive benchmark everywhere.
Implementation cost · Q6
Translated into what a fund tracking this index pays each year, the event-level premiums come to 2.8 bps/yr 2.8 bps/yr No-reversal bound (s=0.70). 02_paper_numbers.md §6 / q6_turnover_cost.csv to 5.1 bps/yr 5.1 bps/yr Full-reversal (s-free) bound. 02_paper_numbers.md §6 / q6_turnover_cost.csv — €3.5M/yr €3.5M/yr No-reversal euro cost. 02_paper_numbers.md §6 / q6_turnover_cost.csv to €6.5M/yr €6.5M/yr Full-reversal, median-p euro cost. 02_paper_numbers.md §6 / q6_turnover_cost.csv on the identified €12.07bn €12.07bn Mean identified tracked ETF AUM, 2014-2025 (a floor). 02_paper_numbers.md §6 / q6_turnover_cost.csv ETF base, and 4–8× below 4–8× below European cost vs Petajisto (2011) S&P-500 1990-2005 estimates. Abstract envelope; median-p full-reversal ≈5.5×, no-reversal ≈7.5× (4.4× is the mean-p companion — never mix weightings in one range). 02_paper_numbers.md §6 / q6_turnover_cost.csv what committee-era US indices imposed. What can be exploited is the public list, instead of the trade itself.
The passive base counts physically-replicating ETFs only, so every euro figure here is a floor.
Illustrative: linear scaling of the 2.8–5.1 bps/yr bound; the identified ETF base makes these floors.
| Review year | No-reversal (bps/yr) | Full-reversal (bps/yr) | Euro cost (€M/yr) |
|---|---|---|---|
| 2014 | 1.3 | 2.3 | 1.1 |
| 2015 | 3.0 | 5.3 | 4.2 |
| 2016 | 0.5 | 1.0 | 0.8 |
| 2017 | 0.8 | 1.6 | 1.5 |
| 2018 | 2.2 | 4.2 | 4.9 |
| 2019 | 1.9 | 3.4 | 3.4 |
| 2020 | 3.4 | 6.2 | 6.2 |
| 2021 | 6.5 | 12.1 | 14.6 |
| 2022 | 6.6 | 12.0 | 16.0 |
| 2023 | 2.3 | 4.6 | 6.6 |
| 2024 | 1.9 | 4.1 | 7.6 |
| 2025 | 2.5 | 4.4 | 11.0 |
| Period mean | 2.8 | 5.1 | — |
| US S&P 500 (Petajisto 2011) | 21 | 28 | — |
Annual hidden rebalancing cost for a STOXX Europe 600 tracker, 2014–2025, under Petajisto (2011) accounting. Each year is a bounded range — the bar top is the full-reversal bound, the overhanging horizontal tick the no-reversal bound. Across the period the cost runs 2.8 to 5.1 basis points per year, several times below the US S&P 500 committee-era benchmark of 21 to 28 basis points per year (Petajisto 2011). The 2021 and 2022 review years are the sample peak.
Annual turnover cost to a tracker, 2014–2025, against the growing passive base. The slider applies an illustrative linear scaling of the premium; the levels are floors on the physically-replicating ETF base.
The Decay Question · Q7
The buffer-add multiplier fell from 40.4 40.4 95% CI [13.58, 61.70] Buffer additions, Middle period; CI excludes 0. 02_paper_numbers.md §7 / q7_disappearing.csv × to 16.3 16.3 95% CI [2.29, 35.96] Buffer additions, Late period; effect persists, CI excludes 0. 02_paper_numbers.md §7 / q7_disappearing.csv × across periods — both real, their confidence intervals excluding zero — but the decline itself, +24.2 +24.2 CI incl. 0 95% CI [−14.31, +47.87] Middle−Late decline: directional, not statistically resolvable. 02_paper_numbers.md §7 / q7_m_contrast_ci.csv ×, carries a confidence interval that includes zero, and the deletion side does not attenuate at all. The sharpest contrast sits on a data seam where genuine decay, the 2022 regime, and measurement change are observationally equivalent — so we flag the seam rather than pick one.
Buffer additions (filled) carry a resolved positive multiplier; core additions (hollow) sit at zero.
No attenuation: point estimates do not decline across periods.
Descriptive split of the Late period; lines connect the two regimes.
No error bars by design: the earlier segment spans ≤5 review cycles — descriptive only; decay, the 2022 regime, and measurement change are observationally equivalent at the seam.
M levels are upper bounds (ETF-only AUM denominator): read trends and contrasts, not levels.
| Panel | Series | Period / regime | M | 95% CI | n | Cycles | CI status |
|---|---|---|---|---|---|---|---|
| A | Buffer additions | Middle (2018–2021) | 40.4 | [13.6, 61.7] | 63 | 14 | Confidence interval excludes zero |
| A | Buffer additions | Late (≥ 2022) | 16.3 | [2.3, 36.0] | 53 | 15 | Confidence interval excludes zero |
| A | Core additions | Early (≤ 2017) | −0.3 | [−21.9, 22.1] | 76 | — | Confidence interval includes zero |
| A | Core additions | Middle (2018–2021) | −7.9 | [−32.8, 24.5] | 54 | — | Confidence interval includes zero |
| A | Core additions | Late (≥ 2022) | 3.9 | [−14.5, 22.1] | 58 | — | Confidence interval includes zero |
| B | Deletions | Early (≤ 2017) | 25.3 | [1.5, 44.5] | 103 | 16 | Confidence interval excludes zero |
| B | Deletions | Middle (2018–2021) | 11.8 | [−4.5, 40.5] | 134 | 16 | Confidence interval includes zero |
| B | Deletions | Late (≥ 2022) | 23.9 | [14.7, 44.3] | 108 | 16 | Confidence interval excludes zero |
| C | Buffer additions | Source regime II | 33.5 | — | 16 | 4 | Descriptive point, no confidence interval |
| C | Buffer additions | Source regime III | 6.5 | — | 37 | 11 | Descriptive point, no confidence interval |
| C | Deletions | Source regime II | 49.4 | — | 36 | 5 | Descriptive point, no confidence interval |
| C | Deletions | Source regime III | 17.4 | — | 72 | 11 | Descriptive point, no confidence interval |
| A | Buffer additions — Middle minus Late contrast | Middle − Late | 24.2 | [−14.3, 47.9] | — | — | Confidence interval includes zero |
Greenwood–Sammon price-multiplier M by period for STOXX Europe 600 rebalancing events. Panel A: buffer additions carry a resolved positive multiplier (Middle M 40.4, Late M 16.3, both CIs exclude zero) while core additions sit at zero; the Middle-minus-Late buffer decline (ΔM +24.2) has a confidence interval that includes zero. Panel B: deletion multipliers show no attenuation — point estimates do not decline across periods (Early 25.3, Middle 11.8, Late 23.9). Panel C: a descriptive within-Late split by source regime, drawn without error bars by design. All M levels are upper bounds; read trends and contrasts, not levels. M levels are upper bounds (ETF-only AUM denominator): read trends and contrasts, not levels. No error bars by design: the earlier segment spans ≤5 review cycles — descriptive only; decay, the 2022 regime, and measurement change are observationally equivalent at the seam.
Price multiplier M by period for marginal additions and deletions. The levels are upper bounds — read the trend, not the height. The Late-period step sits on a source-regime seam and carries no error bar by construction.
Who Should Care
The Index-Tracking Desk
For an index-tracking desk, the effect's move upstream means benchmark-relative risk sits in the weeks before the announcement, instead of the effective date; the residual cost is small, hard to forecast name by name, and worth addressing only through flow-aware scheduling of the handful of marginal, hard-to-hedge names each quarter.
The Index Provider
For an index provider, the European evidence quietly vindicates rule-based transparency: an always-on sunshine regime is associated with a small, front-loaded, and apparently shrinking transfer from trackers to arbitrageurs — several times below what committee-era US indices imposed.
The Would-Be Arbitrageur
For a would-be arbitrageur, the news is worse: the predictable part of the calendar is already priced by the time it is announced, the residual is noise at the horizons that matter, and the surviving edge — anticipating rank resolution among marginal candidates before the list is set — is exactly the part this study shows to be competitive already.
Engineering
Constituent and event history reconstructed from STOXX's public review announcements; market data assembled from commercial feeds via an external market-data API and reconciled key-by-key across source regimes.
| Year | Scheduled | Fast entry | M&A | Spin-off | Other (excluded) | Other mid-cycle | Total | Passive AUM (€bn) | Source regime |
|---|---|---|---|---|---|---|---|---|---|
| 2014 | 60 | 9 | 6 | 8 | 1 | 1 | 85 | 4.72 | Source regime I |
| 2015 | 56 | 12 | 10 | 3 | 6 | 0 | 87 | 7.92 | Source regime I |
| 2016 | 62 | 15 | 17 | 6 | 2 | 0 | 102 | 7.89 | Source regime I |
| 2017 | 49 | 13 | 11 | 6 | 2 | 1 | 82 | 9.37 | Source regime I |
| 2018 | 85 | 14 | 13 | 0 | 1 | 3 | 116 | 11.66 | Source regime II |
| 2019 | 53 | 14 | 12 | 0 | 1 | 2 | 82 | 9.84 | Source regime II |
| 2020 | 109 | 15 | 9 | 0 | 6 | 2 | 141 | 9.98 | Source regime II |
| 2021 | 75 | 21 | 14 | 1 | 5 | 2 | 118 | 12.07 | Source regime II |
| 2022 | 83 | 18 | 12 | 3 | 2 | 2 | 120 | 13.34 | Source regime II |
| 2023 | 54 | 12 | 6 | 1 | 5 | 2 | 80 | 14.53 | Source regime III |
| 2024 | 60 | 14 | 9 | 4 | 3 | 0 | 90 | 18.37 | Source regime III |
| 2025 | 59 | 15 | 14 | 1 | 2 | 1 | 92 | 25.15 | Source regime III |
| 2026* | 22 | 1 | 1 | 0 | 0 | 0 | 24 | 32.90 | Source regime III |
| Total | 827 | 173 | 134 | 33 | 36 | 16 | 1219 |
| Period | Ownership share |
|---|---|
| Early | 0.097% |
| Middle | 0.129% |
| Late | 0.188% |
Sample and architecture overview for 1,219 STOXX Europe 600 rebalancing events, 2014–2026. Row 1: annual event counts stacked by cohort (scheduled, fast entry, M&A, spin-off, other-excluded, other mid-cycle). Row 2: annual-mean passive AUM tracking the index, rising about ×5.3 from €4.7bn in 2014 to €25.1bn in 2025. Row 3: passive ownership share stepping up from 0.097% to 0.129% to 0.188% across the three calendar periods. Behind all rows, three data-source regimes are shaded, split at 2017-10-11 and 2023-06-13. 2026 is a partial year.
Events by year and cohort, with the tracked ETF asset base and passive-ownership steps. The three source regimes appear as disclosed seams, never silently blended.
Every number on this page is traceable to an audited artifact — hover any statistic.
Go Deeper
The full argument, every robustness pass, and the methods behind each figure.
Questions, replication requests, or a role where this kind of work is useful: frli.jht@gmail.com.