What's the question?
Concentration in the S&P 500 is normally discussed in terms of market value, and the ten largest members now carry roughly a third of the index weight. Market value is a price, though, and a price is an opinion about future profits that can be revised overnight.
Reported profit is different. It arrives in an audited annual filing and cannot be revised by sentiment. So the question is whether the profit behind the index has concentrated as fast as the price. If the ten largest earners still produce the share of aggregate profit they produced in 2010, rising index concentration is a valuation phenomenon and should mean-revert like one. If their share has climbed too, the concentration is structural.
Two measures answer this: the top-ten share of the profit pool, being the sum of positive net income across members, and the effective number of companies, the reciprocal of the Herfindahl index of profit shares. The second reports how many equally sized earners would produce the observed concentration.
The approach
Any comparison across sixteen years falls apart if the roster is wrong. Today’s membership list is a list of survivors, and survivors were disproportionately profitable, so backdating it decides the answer in advance.
- Pull S&P 500 membership at each 31 December from 2010 to 2025, keyed to entity identifiers rather than tickers. A ticker is a lease rather than a name: S, PX and SLE carried Sprint, Praxair and Sara Lee in 2010 and carry SentinelOne, P10 and Super League Enterprise today.
- Pull one annual income statement per member per roster. Fiscal year Y covers period ends from June of Y to May of Y+1, which places Microsoft’s June year end and Walmart’s January year end in the year they describe.
- Screen the panel. A member without both figures for that fiscal year, or without positive revenue, leaves that year’s sample. Where a fiscal-year change yields two statements inside one window, the later is kept.
- Compute both measures for revenue and for net income separately, since revenue is a scale measure and profit a margin measure, then recompute fiscal 2010 using today’s roster to size what survivorship bias does here.
Coverage runs from 467 of 486 members in fiscal 2010 to 485 of 496 in fiscal 2025.
Code
import numpy as np
import pandas as pd
import xfinlink as xfl
xfl.set_api_key("YOUR_API_KEY") # free at https://xfinlink.com/signup
def roster(as_of):
df = xfl.index("sp500", as_of=as_of).dropna(subset=["entity_id"])
return sorted(set(df["entity_id"].astype(int)))
def fiscal_year(ids, year):
frames = [xfl.fundamentals(entity_id=ids[i:i + 100], period_type="annual",
start=f"{year}-06-01", end=f"{year + 1}-05-31",
fields=["revenue", "net_income"], max_rows=100000)
for i in range(0, len(ids), 100)]
f = pd.concat([x for x in frames if len(x)], ignore_index=True)
f = f.dropna(subset=["revenue", "net_income"])
f = f[f["revenue"] > 0]
return f.sort_values("period_end").groupby("entity_id", as_index=False).tail(1)
def concentration(values):
pos = np.sort(np.asarray(values, dtype=float)[np.asarray(values) > 0])[::-1]
w = pos / pos.sum()
return 100 * w[:10].sum(), 1.0 / np.sum(w ** 2)
for year in range(2010, 2026):
f = fiscal_year(roster(f"{year}-12-31"), year)
rev_top10, rev_neff = concentration(f["revenue"])
ni_top10, ni_neff = concentration(f["net_income"])
print(f"{year} revenue {rev_top10:.1f}% / {rev_neff:.1f} "
f"profit {ni_top10:.1f}% / {ni_neff:.1f}")Full script with formatting and visualisation: sp500-earnings-concentration-point-in-time-python.py
Output
S&P 500 earnings vs revenue concentration, point-in-time rosters
FY members covered rev top10 rev Neff NI top10 NI Neff profit pool loss-makers
2010 486 467 21.8% 102.9 23.4% 98.3 757b 25
2011 488 468 22.0% 99.9 25.7% 90.4 862b 27
2012 491 470 21.3% 103.0 27.6% 79.3 845b 36
2013 491 470 21.1% 105.4 23.4% 101.8 979b 18
2014 492 469 20.8% 107.9 23.6% 103.6 959b 20
2015 495 475 20.7% 110.9 23.3% 97.2 972b 41
2016 496 483 20.8% 111.6 22.4% 103.7 987b 47
2017 498 486 21.0% 112.0 26.7% 89.1 1,080b 42
2018 499 490 21.6% 110.1 22.6% 104.1 1,213b 28
2019 500 493 22.0% 109.3 27.2% 79.1 1,301b 33
2020 500 493 23.4% 99.5 29.0% 76.8 1,107b 79
2021 500 494 23.7% 102.4 28.3% 76.8 1,840b 20
2022 499 497 23.7% 101.9 27.5% 80.1 1,688b 40
2023 499 499 24.5% 99.3 30.9% 70.5 1,813b 32
2024 499 495 25.2% 95.5 34.8% 63.3 1,966b 24
2025 496 485 25.4% 93.9 35.9% 56.1 2,193b 26
Top 10 earners, FY2010 vs FY2025 ($m)
# FY2010 net income FY2025 net income
1 XOM 30,460 GOOGL 132,170
2 T 19,864 NVDA 120,067
3 CVX 19,024 AAPL 112,010
4 MSFT 18,760 MSFT 101,832
5 JPM 17,370 AMZN 77,670
6 WMT 16,389 BRK 66,968
7 IBM 14,833 META 60,458
8 AAPL 14,013 JPM 57,048
9 JNJ 13,334 BAC 30,509
10 BRK 12,967 XOM 28,844
FY2010 measured two ways
point-in-time 2010 roster : 467 names, top-10 share 23.4%, Neff 98.3
today's roster backdated : 456 names, top-10 share 24.5%, Neff 88.2What this tells us
Revenue barely moved. The ten largest sellers took 21.8 percent of index revenue in fiscal 2010 and 25.4 percent in fiscal 2025, and the effective number of revenue generators fell only from 102.9 to 93.9. Profit moved a long way: the top-ten share went from 23.4 percent to 35.9 percent and the effective count fell from 98.3 to 56.1, a decline of 43 percent. Since the revenue base did not concentrate, the gap is margin. Alphabet converted 32.8 percent of its fiscal 2025 revenue into net income and NVIDIA converted 55.6 percent, against 10.8 percent at Amazon, the largest seller in the index.
Net income absorbs impairments, settlements and tax charges, so the profit series is noisy, and the effective count swings between 76 and 104 through the middle of the decade without trending. What follows fiscal 2022 differs in kind: three consecutive falls, from 80.1 to 70.5 to 63.3 to 56.1, with no reversal.
Measuring fiscal 2010 with today’s members returns a top-ten share of 24.5 percent instead of 23.4 percent and an effective count of 88.2 instead of 98.3, because the companies that later left the index were mid-sized earners whose absence tightens the distribution. That biased start shrinks the measured decline in effective count from 42.2 companies to 32.1, understating the change by 24 percent.
So what?
An index-level earnings forecast is mostly a forecast about ten companies and should be built that way. A bottom-up model aggregating 500 independent estimates assigns weight to names that no longer move the total, and its error is dominated by whichever of the five largest earners misses.
The result also constrains the bubble argument. Price concentration that outruns profit concentration will unwind on sentiment; price concentration that tracks it will not, and both have risen here.
For equal-weight investors the number to watch is the effective count rather than the top-ten weight. An equal-weighted S&P 500 holds roughly 0.2 percent in each of the ten companies producing 35.9 percent of the profit, an underweight to the earnings base that should be sized rather than inherited. And rebuild the roster point in time: the shortcut costs about a quarter of the effect here, always in the direction that makes the past resemble the present.
Built with xfinlink — free financial data API for Python.