How Much of a Growth Screen's Backtested Edge Is Survivorship Bias? Point-in-Time Index Testing in Python

July 17, 2026

What's the question?

A growth screen ranks companies by revenue growth and buys the fastest growers. Testing one requires a universe, and most people reach for the index as it looks today. That choice rigs the test: today’s S&P 500 members survived and grew enough to still be there, while firms acquired or destroyed are absent and recent joiners are present because they had already won. Screening "high growth in 2021, hold to 2026" over that list guarantees the winners are in the sample.

Survivorship bias is the name for this distortion, usually asserted rather than measured. The question is how large it is: run one identical screen over two universes differing only in whether membership is point-in-time or contaminated. The return gap is the survivorship premium.

The approach

xfl.index("sp500", as_of="2021-06-30") returns membership as it stood on that date, including firms later removed; without as_of, today’s.

1. Take the S&P 500 on 30 June 2021 (honest universe) and on 16 July 2026 (contaminated).
2. Size the difference: count the names present then but gone now.
3. Screen both identically — top decile by revenue growth over the last fiscal year public at formation.
4. Hold to 16 July 2026, equal-weighted, against the S&P 500 itself.

Two details decide the result. The first is lookahead: growth must be knowable at formation, so the screen uses the last fiscal year ending on or before 31 March 2021 — a 90-day reporting lag. Only restated figures exist here; the top-decile picks were verified against filings.

The second is the crux: names that leave. Dropping one because its price series ends reintroduces the bias being measured. Every name is held from formation to its last available price; if it stops trading, the proceeds redeploy equally across survivors. Nothing is discarded. Here the rule never handles a termination — all 48 point-in-time picks still trade — but it absorbs one data gap: FIS has a 16-month hole, treated as a pause; excluding it moves the headline to 978 basis points.

"Left the index" rarely means failure: of the 75 leavers, 30 stopped trading, mostly through acquisition. Another 44 still trade — including First Republic, seized in May 2023, now changing hands as FRCB at a fraction of a cent — and one is unreachable by ticker. Entity-keyed matching captures that near-total loss.

Code

import xfinlink as xfl
import pandas as pd

xfl.set_api_key("YOUR_API_KEY")  # free at https://xfinlink.com/signup

FORM, END = "2021-06-30", "2026-07-16"

pit = xfl.index("sp500", as_of=FORM)   # membership AS IT STOOD in June 2021
now = xfl.index("sp500")               # membership today
pit_e, now_e = set(pit["entity_id"]), set(now["entity_id"])

# Revenue growth from the last fiscal year public at formation (90-day lag)
allt = sorted(set(pit["ticker"]) | set(now["ticker"]))
fund = pd.concat([xfl.fundamentals(allt[i:i+40], period_type="annual", fields=["revenue"],
                                   start="2018-01-01", end="2021-03-31")
                  for i in range(0, len(allt), 40)], ignore_index=True)

rows = []
for eid, g in fund[fund["revenue"].notna()].sort_values("period_end").groupby("entity_id"):
    g = g.drop_duplicates("period_end", keep="last")
    if len(g) < 2:
        continue
    t0, t1 = g.iloc[-1], g.iloc[-2]
    if not (300 <= (t0["period_end"] - t1["period_end"]).days <= 430) or t1["revenue"] <= 0:
        continue
    rows.append({"entity_id": eid, "growth": t0["revenue"] / t1["revenue"] - 1})
gr = pd.DataFrame(rows)

def screen(universe):
    g = gr[gr["entity_id"].isin(universe)]
    return set(g.nlargest(int(round(len(g) * 0.10)), "growth")["entity_id"])

s_pit, s_now = screen(pit_e), screen(now_e)

# prices() caps at 50,000 rows and truncates silently — chunk the request
tick = sorted(pd.concat([pit, now]).drop_duplicates("entity_id")
              .set_index("entity_id").loc[sorted(s_pit | s_now), "ticker"])
px = pd.concat([xfl.prices(tick[i:i+25], start="2021-06-25", end=END,
                           fields=["close", "return_daily"])
                for i in range(0, len(tick), 25)], ignore_index=True)
px["date"] = pd.to_datetime(px["date"])
px = px[px["entity_id"].isin(s_pit | s_now)]          # entity guard: recycled tickers
investable = set(px[px["date"] == FORM]["entity_id"])  # must be buyable at formation
wide = px[(px["date"] > FORM) & (px["date"] <= END)].pivot_table(
    index="date", columns="entity_id", values="return_daily")

def backtest(ids):
    # equal weight across names still trading; leavers' proceeds redeploy to survivors
    r = wide[[c for c in wide.columns if c in ids & investable]].mean(axis=1, skipna=True)
    cum = (1 + r.fillna(0)).cumprod()
    yrs = (cum.index[-1] - cum.index[0]).days / 365.25
    return cum, cum.iloc[-1] ** (1 / yrs) - 1

cum_pit, cagr_pit = backtest(s_pit)
cum_now, cagr_now = backtest(s_now)
print(f"honest {cum_pit.iloc[-1]-1:.2%} ({cagr_pit:.2%}/yr)   "
      f"naive {cum_now.iloc[-1]-1:.2%} ({cagr_now:.2%}/yr)   "
      f"gap {(cagr_now-cagr_pit)*1e4:.0f} bp/yr")

Full script with formatting and visualisation: growth-screen-survivorship-bias-python.py

Output

S&P 500 members on 2021-06-30: 499
S&P 500 members on 2026-07-16: 499
In the 2021 index, gone today: 75
In today's index, absent in 2021: 75
  of those 75: 44 still have a live price series, 30 stopped trading (acquired, taken private or failed),
  1 unreachable by ticker (recycled to an unrelated company)

Screenable (revenue history available): point-in-time 481/499, today 489/499
Naive screen picks that were NOT index members at formation: 16/49
    MRNA  MODERNA INC                    FY growth +1234%
    HOOD  ROBINHOOD MARKETS INC          FY growth +247%
    DASH  DOORDASH INC                   FY growth +226%
    COIN  COINBASE GLOBAL INC            FY growth +139%
    SQ    BLOCK INC                      FY growth +101%
    CRWD  CROWDSTRIKE HOLDINGS INC       FY growth +82%

                                          total     CAGR
Honest screen (point-in-time)     n=48    52.57%    8.74%
Naive screen (today's index)      n=48   138.22%   18.79%
  picks common to both            n=33    89.31%   13.50%
  honest-screen-only picks        n=15    -7.39%   -1.51%
  naive-screen-only picks         n=15   242.67%   27.68%
S&P 500 ETF (SPY)                 n=1     84.52%   12.92%

SURVIVORSHIP PREMIUM: +85.65% over the window, +1005 bp per year

Sector labels resolved: members 498/499, removed 74/75, added 75/75

Index churn by sector:
                         in_2021  removed  added  churn_rate
Communication Services       22      6.0      4    0.454545
Information Technology       63      9.0     18    0.428571
Consumer Discretionary       58     16.0      6    0.379310
Energy                       21      3.0      4    0.333333
Financials                   73     10.0     12    0.301370
Materials                    28      5.0      3    0.285714
Health Care                  64     11.0      6    0.265625
Industrials                  76      8.0     11    0.250000
Real Estate                  30      3.0      4    0.233333
Consumer Staples             35      3.0      4    0.200000
Utilities                    28      0.0      3    0.107143

What this tells us

The survivorship premium is 1,005 basis points per year — 85.65 percentage points over five years. Same screen, same date, same rules; the only difference is which list the screen could look at.

The direction of the error matters more than its size. Measured honestly, the screen returned 8.74% a year against the S&P 500’s 12.92% — losing to a passive index fund by 418 basis points. Measured naively, it returned 18.79% and appears to beat the index by 587. The backtest does not overstate a real edge; it manufactures one where the truth is negative.

The decomposition locates the damage. The 33 shared picks returned 89.31%, roughly the index. The 15 held only by the honest screen returned −7.39%; the naive screen’s 15 exclusive picks returned 242.67%. And 16 of the 49 naive picks were not members at formation: Moderna joined three weeks later, Palantir in 2024, AppLovin and Carvana in 2025.

The dominant channel is therefore not failures being deleted but winners being inserted. Committees add companies that have already succeeded, so today’s membership is assembled with knowledge of the future. Deletion contributes less: removal is mostly acquisition, and acquired companies leave at a premium.

So what?

The universe is part of the strategy, not background scenery. A backtest built on today’s constituent list is not a weak test of a growth screen; it tests a different strategy — one that knew which companies the committee would later admit. Reconstruct membership at each formation date, and let names leave when they did.

Contamination scales with churn, so a sector-tilted screen inherits phantom edge in proportion to its tilt. Churn runs from 45% in Communication Services to 11% in Utilities, across the 498 of 499 members carrying a sector label, and the mechanisms differ: Information Technology added 18 names while losing 9, whereas Consumer Discretionary lost 16 and added 6. Apparent alpha gaps across sectors may measure index turnover rather than signal.

Two safeguards cost little: key on entity identifiers rather than tickers, and confirm bulk requests were not truncated — this pull spans 80,802 rows against a 50,000-row default cap.

Built with xfinlink — free financial data API for Python. pip install xfinlink