Predicting index changes: how likely is a company to be added to or deleted from an index?
Predicting index changes means estimating, for each company, the probability that it will be added to or deleted from an index at one of the next reviews. This works because the rules are public and the rankings of the candidates follow from market capitalizations; the only unknown is how prices move until the cutoff date. Here you will learn how the distance to the threshold, price volatility and the time to the cutoff date turn into a probability, how a Monte Carlo simulation runs through entire rankings, how the model can be calibrated on past changes and where it reaches its limits. Why prices react to index changes is covered in the article Index effect.
Why index changes can be calculated
Index providers review their indices on fixed dates and by published rules: companies are ranked by market capitalization, and whoever crosses a threshold moves up or down. So the rule is known, but the ranking on the cutoff date is not, because prices keep moving until then. A forecast is therefore never a yes or no but a probability between 0 and 1. It depends on three things: how far a company is from the threshold today, how much its price can fluctuate until the cutoff date and how much time is left.
An index change has three phases: the time before the announcement, when only probabilities are possible; the time between announcement and implementation, when the change is certain and the demand from index funds matters; and the implementation day itself. This article covers the first phase. The other two are described in the article Index effect.
The rules: thresholds, buffers and the special case of the S&P 500
Most indices work with a ranking and a buffer: a newcomer needs a better rank than a member, which only drops out once it falls below another, lower rank. This keeps companies near the threshold from switching back and forth at every review. For a forecast, what counts is exactly where these ranks lie:
| Index | Measure | Inclusion | Deletion |
|---|---|---|---|
| DAX | Free-float market capitalization, average price (VWAP) of the last 20 trading days | Rank 33 or better (quarterly), in March and September rank 40 or better | Rank worse than 60 (quarterly), in March and September worse than 53 |
| FTSE 100 | full market capitalization | Rank 90 or better | Rank 111 or worse |
| MSCI, standard index of a market | Market capitalization against a size cutoff that follows from 85% coverage | new companies from the size cutoff, small-cap members from 1.5 times the cutoff | below two thirds of the size cutoff |
| S&P 500 | Committee; new members need a market capitalization of at least $22.7 billion | no rule, the committee decides | no rule |
Sources: methodologies of STOXX (DAX, DAX Methodology (2026)), FTSE Russell (FTSE 100, FTSE Russell, UK ground rules (2026)), MSCI (MSCI GIMI Methodology 2024) and S&P Dow Jones Indices (S&P DJI, U.S. indices methodology (2026); size threshold since July 1, 2025, S&P DJI (2025)). New DAX members also need positive EBITDA in the last two financial years. Rules change; the methodology in force at the time is what counts.
For the DAX, FTSE 100 and MSCI, the threshold is not a fixed amount but the market capitalization of the company at the threshold rank. This value also moves with the market. The S&P 500 has no ranking rule at all: a committee decides when and how many constituents are replaced, and the eligibility criteria apply only to new members (S&P DJI, U.S. indices methodology 2026). A model can only calculate who meets the criteria there and has to assume the number of replacements.
The model: how far away, how uncertain?
In the standard model, a company's market capitalization follows a geometric Brownian motion, so prices are lognormally distributed (Hull 2021). For the logarithm of the market capitalization on the cutoff date:
- market capitalization today
- annual volatility of the price
- time to the cutoff date in years
- random number from the standard normal distribution
The term −σ² T / 2 makes the expected market capitalization equal today's; the same effect is behind volatility drag. If the threshold B is fixed, the probability of being above it on the cutoff date has a closed form:
- cumulative distribution function of the standard normal distribution
- logarithmic distance to the threshold; positive if the company is already above it
- dispersion of the logarithm until the cutoff date, i.e. the unit in which the distance counts
Formula from the example, T = 0.25 years, fixed threshold. The more the price fluctuates, the flatter the curve: far below the threshold, volatility helps; far above it, volatility hurts. Exactly at the threshold, the probability is just under 50% because the drift −σ² T / 2 lowers the logarithm.
With ranking rules, however, the threshold itself moves, and prices partly move together. A market factor that shifts all market capitalizations by the same factor does not change any rank. The simulation captures this with a one-factor model: a company's random number consists of a common part and its own part.
- random number of company i in the formula above
- common random number of all companies (market factor)
- company i's own random number
- co-movement of prices; at ρ = 0 all move independently, at ρ = 1 all move the same
In each run of the Monte Carlo simulation, the model draws the market capitalizations of all companies up to the cutoff date of each of the next reviews, applies the rules to the resulting ranking and carries membership forward from review to review (Glasserman 2003). The probability of an inclusion is the share of runs in which it occurs. The probability for any of the reviews is not the sum of the individual ones, because a run counts only once. The random error falls with the square root of the number of runs: at a probability of 50%, the standard deviation of the estimate is 1.6 percentage points with 1,000 runs and 0.7 percentage points with 5,000 runs.
Advanced: the average price for the DAX
The DAX ranking is not based on the price on the cutoff date but on the average price (VWAP) of the last 20 trading days. An average fluctuates less than a single price, so the uncertainty until the cutoff date is smaller than the time T suggests. For a Brownian motion, the mean over a window of length L (20 trading days, i.e. L = 20 / 252 years) has the variance σ² (T − 2L / 3) once the window lies entirely in the future. If it starts today or earlier, only the part in the future counts:
- effective time; it replaces T in σ √T
- length of the window in years, 20 / 252
- time to the data cutoff date in years
At three months (T = 0.25), τ = 0.197 instead of 0.25, and at 30% volatility the dispersion σ √τ falls from 0.150 to 0.133. Candidates near the threshold thus become somewhat more predictable. The effect is largest shortly before the cutoff date (T below 20 trading days): at T = 0.05, τ is only 0.0066.
Calibration: logit and probit on past changes
The pure model assumes that index providers apply their rules as written. In practice there are exceptions: takeovers, corrections to the free float, interpretations. That is why the model can be calibrated on past changes. For each past review, each level and each company, the standardized distance to the threshold in the direction of the event is calculated, along with whether the event occurred. From this, a probability curve is estimated:
- standardized distance; positive if the company is already on the side of the event
- distribution function: normal distribution (probit) or logistic distribution (logit)
- coefficients estimated from the history (maximum likelihood)
With probit, a = 0 and b = 1, the result is exactly the formula above, apart from the drift. This makes a and b easy to read: if a differs from 0, changes at the threshold occur more often (a > 0) or less often (a < 0) than price volatility alone would suggest, for example because of exceptions made by index providers. If b differs from 1, the outcome depends more (b > 1) or less on the distance. The logistic distribution has fatter tails than the normal distribution; with few events, the data can hardly tell which of the two fits better.
The estimate uses Newton's method and a small penalty on b, because with perfectly separable data b would go to infinity. Without at least three events and three non-events, there is no estimate; distances above 6 standard deviations carry no information and are left out.
What research shows
The building blocks of this model are well known: lognormal prices, Monte Carlo simulation and binary regression models. This article does not cite a study that tests this kind of forecast of index changes; it is an application of known methods to published rules. Research on the index effect does, however, show when a forecast is worthwhile. For the S&P 500, the excess return on inclusion fell from an average of 7.4% in the 1990s to below 1% in the last decade; in the working paper version, the authors name the better predictability of inclusions as a possible additional reason (Greenwood and Sammon 2025; Greenwood and Sammon 2022). The better inclusions can be predicted, the more likely part of the effect is already in the price before the announcement.
For the Russell reconstitution, Madhavan (2003) found that rebalancing on the cutoff date was costly for funds, while liquidity providers earned money but bore risk. For the S&P 500, Patel and Welch (2017) show that price reactions over longer windows have reversed since the 2000s and that no permanent shift in demand is visible anymore. A probability of inclusion is therefore a statement about the index, not a recommendation to buy the stock.
Limits, misconceptions and common mistakes
- Committee instead of rule: For the S&P 500, a committee determines the timing and number of replacements. The model only calculates who would be eligible and needs the number of replacements per review as an assumption. The probabilities depend heavily on it.
- Estimate instead of an official list: Where no ranking is available, members and sizes come from the holdings of funds that report to the US Securities and Exchange Commission (SEC). That is not the official composition of the index, and the reports arrive quarterly with a delay of several weeks. The sizes are therefore uncertain; the model adds an estimation error for this.
- Events outside the rules: The model does not know about takeovers, mergers, IPOs, changes in free float or special decisions by index providers. A company that is taken over disappears from the index without any ranking rule requiring it.
- Constant assumptions: Volatility and co-movement are treated as constant per security and period, and prices are lognormally distributed. Real prices have fatter tails, so rare large jumps are underestimated.
- Probability is not certainty: 22% means that out of many identical cases, roughly one in five occurs. For a single case, the number says nothing. Whether a forecast is well calibrated only becomes clear over many reviews.
- Only with data: The Index changes page calculates with estimated or imported rankings, not with the lists of the index providers. It does not offer indices without a daily estimate and without an import. This is not investment advice.
How Portfolio Lab calculates
The Index changes page in the Research section knows the rules of eleven index families: DAX, MDAX and SDAX, TecDAX, S&P 500, Nasdaq-100, EURO STOXX 50, STOXX Europe 600, FTSE 100, Russell 1000 and 2000, and MSCI standard for one market, plus MSCI Germany and MSCI USA. You can choose the families for which there is a daily estimate or an imported ranking; for these the page calculates the three phases of an index change. Before the announcement, it shows for each candidate the probability of inclusion or deletion at each of the next up to four regular reviews, at most twelve months ahead. The rules are stored with source and date.
The defaults are 1,000 runs (up to 5,000), a volatility of 30% for securities without their own figure, a co-movement of 0.3 and the logistic distribution; volatility can also be estimated from prices. For the S&P 500, the model assumes at most three replacements per review, in which the largest eligible non-members replace the smallest members; the number can be changed between 0 and 20. An imported history of past changes calibrates logit or probit, and the tool shows simulation and model side by side, along with a chart of the probability over the distance to the threshold. Rankings can be imported as CSV; without an import, the tool estimates them as far as reports from suitable funds are available. For MDAX, SDAX, TecDAX, STOXX Europe 600 and FTSE 100 there is no estimate: they only calculate with an imported ranking.
Between announcement and implementation, the Index changes page shows the announced changes with their schedule, the price path since the announcement against the index and the expected demand from index funds in days of trading volume. On the implementation day, the purchases and sales at the closing price and an event study of past changes are added. The calculation behind this is explained in the article Index effect. You can download the page's report as PDF or HTML, save it, share it and embed it.
Frequently asked questions
Can index changes be predicted?
As a probability, yes; as a certain statement, no. The rules and the rankings are public; what is unknown is how prices move until the cutoff date. The closer a company is to the threshold and the more its price fluctuates, the more uncertain the result. For the S&P 500, a committee also decides.
How do you calculate the probability of an index inclusion?
With a fixed threshold, using Φ((ln(market capitalization / threshold) − σ² T / 2) / (σ √T)). With ranking rules, the threshold moves, so you simulate the market capitalizations of all companies until the cutoff date, apply the rules to the rankings and count how often the company is added.
Why is the S&P 500 harder to predict than the DAX?
The DAX follows ranking rules with fixed ranks for inclusion and deletion. For the S&P 500, a committee chooses when and how many constituents are replaced; the criteria apply only to new members. A model can therefore only calculate eligibility and needs an assumption about the number of replacements.
What is the use of knowing about an inclusion in advance?
A lot in the past, little today. For the S&P 500, the excess return on inclusion fell from 7.4% in the 1990s to below 1%, and better predictability is seen as a possible additional reason. Specialized traders are faster, and costs and forecast errors add up. This is not investment advice.
How accurate is such a forecast?
A probability of 22% means that out of many identical cases, roughly one in five occurs. On top of that come the random error of the simulation (about 1.6 percentage points at 50% with 1,000 runs) and the model error: constant volatility, lognormal prices, estimated sizes and events outside the rules.
Where does the data on index changes come from?
The rules come from the index providers' methodologies, with source and date. You can import rankings as CSV; without an import, the tool estimates them where possible from the holdings of funds that report to the US Securities and Exchange Commission (SEC). That is not the official composition of the index.
Literature
- FTSE Russell (2026): Ground Rules: FTSE UK Index Series, v17.4. LSEG/FTSE Russell, October 2026; accessed November 3, 2026. Source
- Glasserman, Paul (2003): Monte Carlo Methods in Financial Engineering. Springer.
- Greenwood, Robin; Sammon, Marco (2022): The Disappearing Index Effect. NBER Working Paper 30748 (version later published in the Journal of Finance). doi:10.3386/w30748
- Greenwood, Robin; Sammon, Marco (2025): The Disappearing Index Effect. The Journal of Finance 80 (2), pp. 657–698. doi:10.1111/jofi.13410
- Hull, John C. (2021): Options, Futures, and Other Derivatives. Pearson, 11th edition (Global Edition).
- Madhavan, Ananth (2003): The Russell Reconstitution Effect. Financial Analysts Journal 59 (4), pp. 51–64. doi:10.2469/faj.v59.n4.2545
- MSCI Inc. (2024): MSCI Global Investable Market Indexes Methodology. MSCI, November 2024; accessed October 25, 2026. Source
- Patel, Nimesh; Welch, Ivo (2017): Extended Stock Returns in Response to S&P 500 Index Changes. The Review of Asset Pricing Studies 7 (2), pp. 172–208. doi:10.1093/rapstu/rax012
- S&P Dow Jones Indices (2025): S&P Dow Jones Indices Announces Update to S&P Composite 1500 Market Cap Guidelines. Press release, July 1, 2025; accessed November 3, 2026. Source
- S&P Dow Jones Indices (2026): S&P U.S. Indices Methodology. S&P Dow Jones Indices; accessed November 3, 2026. Source
- STOXX Ltd. (2026): DAX Equity Index Methodology Guide. STOXX, August 2026 version; accessed October 26, 2026. Source
This article explains a method; it is not investment or tax advice. Worked examples are simplified and rounded. All terms in the glossary.