Half Your Index Rebalance Happens In The Last 60 Seconds - And Arbs Are Waiting
Introduction
MSCI index reconstitutions move roughly half of the day’s trading volume into the final minute before the effective date. Index-tracking funds are so terrified of tracking error that they pile into closing auctions like lemmings, creating a 5.4% round-trip arbitrage opportunity over two weeks. Chang et al. (2025) document this across 56 markets from 2006-2023, and the numbers are frankly embarrassing for passive fund managers.
What problem are they really solving?
Index funds need to buy additions and sell deletions when MSCI reshuffles its indices. The question is: when exactly do they trade? Theory says they should spread execution to minimize impact. Reality says they cram everything into the closing auction on ED-1 to hit the closing price and minimize tracking error.
The paper quantifies this rigidity and asks whether it creates exploitable patterns.
Quick summary:
3,291 additions and 2,638 deletions across 56 markets (2006-2023)
Average 14 trading days between announcement (AD) and effective date (ED)
ED-1 abnormal return (AR(ED-1)): +0.9% for additions, -1.3% for deletions
41-50% of ED-1 volume happens in the final minute
Long-short strategy from AD to ED-1 earns 5.4% net of lending fees
The prediction target and what it means for trading
The authors are not building a predictive model in the ML sense. They are documenting a mechanical pattern: when MSCI announces an index change, prices drift predictably until the effective date, then reverse.
For additions:
CAR peaks at ED-1 close
AR(ED-1) = +0.9% on average
AR(ED) = -0.7% (reversal)
For deletions:
CAR troughs at ED-1 close
AR(ED-1) = -1.3% on average
AR(ED) = +1.2% (reversal)The trading implication is dead simple: go long additions and short deletions at AD, close everything at ED-1 close. You are front-running the index trackers who must trade at ED-1 close, then stepping aside before the reversal.
Data and features
Data setup:
MSCI Global Standard Index reconstitution announcements from MSCI website
Daily prices from Compustat Global and WorldScope
Intraday minute-level data from Bloomberg
Equity lending data from Markit Security Finance
US index fund holdings from CRSP Mutual Fund Database
The “feature” is just the announcement itself. Once MSCI says “stock X is being added,” you know index trackers will buy it before ED. That is the entire alpha.
Key variables constructed:
AR(t) = R(t) - R_market(t) # market = MSCI ACWI + Frontier
CAR[t1,t2] = sum(AR(t)) for t in [t1, t2]
AbnormalTurnover(t) = Turnover(t) / AvgTurnover(AD-200, AD-100)
LMVol/Vol = LastMinuteVolume / DailyVolumeThe last-minute volume ratio is the key metric for measuring index-tracker rigidity. On normal days, the final minute accounts for 7-8% of volume. On ED-1, it is 41-50%.
Does it actually work?
Gross returns:
The long-short strategy S[AD, ED-1] earns:
Long leg (additions): +2.5%
Short leg (deletions): +3.6%
Total: +6.2%
Holding period: ~14 trading days
After costs:
Subtracting average daily lending fees for the short leg:
Total net: +5.4%
The one-day strategy S[ED-1] (enter at ED-2 close, exit at ED-1 close) earns 1.7% gross.
Regional breakdown tells you where the edge is fattest:
Asia-Pacific and “Other” markets (Middle East, Africa, South America) show the strongest effects. North America is weakest - US index trackers are apparently less religious about hitting the closing price exactly.
Region | Total Return S[AD,ED-1]
Asia-Pacific | 6.1%
Europe | 4.1%
North America | 2.8%
Others | 5 .4%Robustness:
The authors run cross-sectional regressions controlling for size, momentum, beta, asset growth, book-to-market, leverage, liquidity, and profitability. The effects survive. They also show that the change in lendable shares around ED (a proxy for index-tracker holdings) predicts both ED-1 returns and volume concentration.
Time trend warning:
For additions, CAR[AD, ED-1] has been declining over time. Index trackers are getting slightly less rigid, or arbs are getting more aggressive, or both. For deletions, the pattern has actually intensified - more short-selling activity is piling in.
How you would actually test this
The setup is straightforward because MSCI publishes reconstitution announcements publicly:
1. Pull MSCI index change announcements from msci.com/index-review
2. For each addition/deletion, get daily prices and volume
3. Calculate CAR from AD to ED-1
4. Calculate AR on ED-1 specifically
5. For intraday analysis, pull minute-level data for ED-1
6. Compute last-minute volume concentration
Simple backtest:
- On AD close: long additions, short deletions (equal weight or cap weight)
- On ED-1 close: flatten everything
- Track gross and net (after borrow costs) PnLThe main implementation questions:
Borrow availability: Can you actually locate shares to short for deletions? The paper uses Markit data showing utilization rises but does not discuss hard-to-borrow situations.
Execution quality: The strategy assumes you can execute at AD close and ED-1 close. In reality, you are competing with index trackers in the closing auction on ED-1. Your sell orders for additions are hitting the same bid as their buys.
Capacity: With 50-100 events per quarter globally, and many in small/mid-cap names, realistic capacity is probably low tens of millions USD per quarter at best.
How much of this survives contact with the real world?
Costs and impact:
The paper subtracts lending fees but does not model market impact or spread costs. For liquid large-caps in developed markets, this is probably fine. For small-caps in frontier markets, you are likely paying several percent round-trip in execution costs alone.
Execution assumptions:
The strategy requires executing at closing prices. In practice, closing auctions are competitive. On ED-1, you are trying to sell additions into a closing auction where index trackers are aggressive buyers. You might not get the closing price - you might get filled early at worse prices or not filled at all.
Latency and data availability:
MSCI announcements are public and well-covered. No speed advantage exists. The edge comes from willingness to take the position and hold it for two weeks, not from being fast.
Look-ahead and survivorship:
The sample uses realized additions and deletions. In practice, you need to predict which stocks will be added/deleted. MSCI’s methodology is rule-based and somewhat predictable, but there is noise. The paper does not quantify prediction accuracy for the announcement itself.
Regime sensitivity:
The authors show the effect exists globally across 2006-2023, including through the 2008 crisis and COVID. That is reassuring. The declining time trend for additions is less reassuring - the trade may be getting more crowded.
Blunt verdict:
This is a real effect with solid documentation. The strategy is implementable for someone with global equity infrastructure and prime brokerage for shorting. Realistic net returns after all costs are probably 2-3% per event rather than 5%+. Capacity is limited. This is a nice supplemental strategy for a global equity desk, not a standalone fund.
What traders should actually do with this
If you trade global equities:
Track MSCI reconstitution announcements (they publish a calendar)
Run your own backtest on your universe with your cost assumptions
Focus on Asia-Pacific and EM where the effect is strongest
Avoid North America unless you have very low execution costs
If you run an index fund:
Stop being so rigid. You are leaving money on the table by cramming all your rebalancing into the closing auction on ED-1. The arbs are front-running you.
If you are building execution algos:
The paper documents that half the volume on ED-1 hits in the final minute. If you are trying to execute large orders on reconstitution days, you need to account for this concentration.
Expected edge:
Small but real. This is not a get-rich-quick trade. It is a 2-3% edge over two weeks that requires operational infrastructure to capture. If you already have the infrastructure, it is free money. If you do not, the fixed costs probably do not justify building it for this alone.
Minimal Reproduction Plan
Core hypothesis: MSCI index additions earn positive abnormal returns from AD to ED-1; deletions earn negative abnormal returns. Returns reverse on ED.
Minimal data:
MSCI reconstitution announcements (free from msci.com/index-review)
Daily price data for affected stocks (Yahoo Finance, or paid: Refinitiv, Bloomberg)
Start with one market (e.g., US or Japan) for simplicity
2-3 years of data is enough for initial validation
Feature construction:
# There is no feature engineering - the “signal” is the announcement itself
is_addition = 1 if stock added to index else 0
is_deletion = 1 if stock deleted from index else 0Target construction:
# For each event
CAR_AD_to_ED1 = sum(daily_return - market_return for t in [AD, ED-1])
AR_ED1 = return(ED-1) - market_return(ED-1)
AR_ED = return(ED) - market_return(ED)Model specification:
There is no model. This is an event study. You are testing whether mean CAR and AR are significantly different from zero.
# Simple t-test
from scipy.stats import ttest_1samp
t_stat, p_value = ttest_1samp(CAR_additions, 0)
**Training and evaluation protocol:**
This is not a train/test setup. You are measuring average effects across events.
# Split by time to check stability
Period 1: 2015-2018
Period 2: 2019-2022
Period 3: 2023-present
Check that effect exists in all periodsSanity checks:
Effect exists in both additions and deletions
Effect exists in multiple markets
Reversal on ED confirms temporary price pressure
Volume spike on ED-1 confirms mechanical flow
Cost model for reality check:
net_return = gross_return - total_round_trip
**Success criteria:**
SUCCESS if:
- CAR[AD, ED-1] > 2% for additions, < -3% for deletions (statistically significant)
- AR[ED-1] > 0.5% for additions, < -0.5% for deletions
- Effect exists in multiple time periods
- Net return after costs > 1%
PARTIAL SUCCESS if:
- Gross metrics match paper but costs kill it
- Effect exists but only in certain markets
FAILURE if:
- Cannot replicate basic CAR patterns
- Effect only exists in-sample (older periods)If reproduction succeeds:
Add intraday analysis (requires Bloomberg or similar)
Expand to all 56 MSCI markets
Build announcement prediction model to front-run official announcements
Test variations: different entry/exit timing, position sizing by expected volume
Conclusion
Index-tracking funds cram half their reconstitution trades into the final minute before MSCI changes take effect, creating predictable price pressure that arbs can harvest for 2-3% net returns over two weeks. The effect is global but strongest in Asia-Pacific and emerging markets, weakest in North America. The addition side is fading over time, but deletions remain juicy targets for short-sellers. As long as passive funds treat tracking error as the cardinal sin, this scheduled wealth transfer from indexers to arbitrageurs will persist.