Survivorship Bias

Survivorship bias

The companion to lookahead-safety: that one stops you from using data that didn't exist yet. This one stops you from using a UNIVERSE that didn't exist yet. Together they're the two halves of "what was actually knowable, on what tickers, on a given date."

Core principle

A backtest universe must be reconstructed at every rebalance / query date from the membership snapshot AT that date — never from today's resolution.

Today's S&P 500 contains companies that joined in 2024. Today's Russell 3000 excludes companies that delisted in 2022. Today's "all US small-caps" is filtered by who survived to today.

Using any of those for a 2018 backtest is the single most common way to manufacture alpha that doesn't exist.

Where survivorship bias enters silently

1. Universe construction

Bug: "Pull all small caps with mkt cap < $300M today, then run the strategy on their 2018-2023 history." Why wrong: The bankrupt, merged, and reverse-split-delisted small caps from that period are gone. You're testing only on companies that survived. Returns inflated, drawdowns understated. Fix: point-in-time universe — at each rebalance date, query the universe AS OF that date including names that subsequently delisted.

2. Ticker resolution

Bug: Vendor API returns "no data" for a delisted ticker, or silently returns the SUCCESSOR entity's data. Why wrong: Some yfinance / Polygon resolutions for delisted tickers map to merger-acquirer prices, fabricating the equity trajectory. (Documented for HDFCBANK demerger, FAANG-era M&A patterns, several SPAC-merger flips.) Fix: require a delisting / effective-date table; never trust ticker-only resolution. If a vendor returns prices for a "delisted" ticker, treat as suspect until verified.

3. Index membership

Bug: "Use the Russell 3000 components for my 2018-2023 backtest." Why wrong: Russell rebalances in June each year. Today's R3000 is the June 2025 reconstitution; June 2018's was different and is the correct snapshot for any 2018-06-25 to 2019-06-24 query. Fix: historical constituents (LSEG, Siblis Research, Bloomberg's PORT) keyed by reconstitution date.

4. Strategy ranking / leaderboards

Bug: "Backtested top 10 strategies on names with > 5 years of history." Why wrong: "5 years of history" filters by survival. The strategies that worked on subsequently-delisted names are pruned. Fix: include short-history names with synthetic "strategy-not-applicable" outcomes for periods before they listed.

5. Manager / fund track records

Bug: Hedge fund index based on "currently reporting funds". Why wrong: Closed and blown-up funds drop out of the index. Inflates the asset class. Fix: include the dropouts at their last-reported value, not zero and not removed.

Small-cap-specific traps (the lane this skill cares most about)

Reverse-split-then-delist phantom returns

Pattern: Small cap does a 1:20 reverse split to maintain Nasdaq listing compliance, then delists 30–90 days later anyway. Why dangerous: Adjusted-price feeds apply the 20x split factor to all pre-split prices. A name that traded $0.05 → $1.00 (from reverse split) → $0.10 (delisting) shows in adjusted feeds as a $1.00 → $1.00 → $0.10 trajectory. If you exit at the last available price BEFORE delisting, the adjusted feed makes it look like a 20x phantom return on the holding period. Fix: raw OHLCV + split table + delisting table. Treat any name that reverse-split + delisted within ~120 days as a survivorship- adjustment candidate; verify the actual cash-out value.

ATM-into-delisting

Pattern: Heavy ATM dilution + Nasdaq compliance failure + delisting → shareholders end up with shares of nothing. Why dangerous: Last-trade price on the last trading day overstates realizable value. Pink-sheet quotes after delisting can be 80–95% discount to last Nasdaq print. Fix: mark delisted positions to ZERO unless there is a documented post-delisting realization (cash distribution, M&A consideration).

Reg SHO threshold residency as leading indicator

Pattern: Names that sit on the Reg SHO threshold list for >13 consecutive days have an elevated probability of subsequent forced delisting (failure to meet listing standards) within 6–12 months. Use: when constructing a small-cap short universe, names that satisfy "Reg SHO threshold >13 days" should NOT be filtered out at the rebalance — keeping them in (with realistic borrow cost from transaction-cost-modeling) is what tests the strategy on the candidates that mattered.

SPAC pre-merger / post-merger discontinuity

Pattern: SPAC trades at $10 NAV pre-merger; post-merger the combined entity often sees significant deviation. Many SPACs that merged in 2021 are now trading <$1 or have delisted. Why dangerous: Backtests using "SPAC tickers" today miss the ~30%+ that have delisted; backtests starting from "pre-merger SPAC" need to handle the ticker change at merger date and the failure mode if the deal didn't close.

Reverse-stock-split anti-pattern detection

A 1:N reverse split where N >= 5 on a sub-$300M issuer is, per SEC and Nasdaq listing data, more often than not a compliance move that precedes one of:

1. Continued dilution to fund operations
2. Delisting within 12 months
3. Merger of convenience with a private operator (de-SPAC variant)

Treat such names as higher-than-baseline survivorship-adjustment risk in any backtest covering the period.

Data sources (where to get survivorship-bias-free universes)

Free reconstruction path: SEC's quarterly Form 25 delisting filings + EDGAR's company-tickers feed snapshot per quarter + manual ticker- change tracking from 8-K item 5.07 disclosures.

Workflow when reviewing a backtest universe

1. Identify the universe definition (filter, index, manual list).
2. Confirm membership is point-in-time, not today's snapshot.
3. Confirm delisted names are present at their delisting date with either a realized cash value or a marked-to-zero treatment.
4. Confirm ticker-level data is sourced from a delisting-aware feed, not yfinance/info-style resolution.
5. For small-cap universes specifically: count the delisting rate over the backtest window. If it's <5%, the universe is suspiciously filtered. Reality is 10–30%+ over 5-year windows for that universe.
6. Stamp the analysis with the universe-as-of date and the delisting- data source for reproducibility.

Composition with other skills

• lookahead-safety: same principle (no future data) applied to a different dimension (universe vs. dataset).
• dilution-event-scoring: high scores predict delisting risk; delistings are exactly the names survivorship bias erases.
• transaction-cost-modeling: HTB borrow + extreme spreads are precursors to delisting; if the cost model is realistic, you naturally include the delisted names with their actual exit costs.
• sec-filing-types: NT 10-K / NT 10-Q / Form 25 are the filings that signal listing-compliance failures. A skill-aware pipeline tracks these as leading indicators.

Phrases that should trigger this skill

• "survivorship bias" / "survivor bias" / "survival bias"
• "backtest universe" / "all small caps from 2018"
• "I pulled all tickers with X" → followed by historical analysis
• "delisted" / "delisting" / "removed from index"
• "Reg SHO threshold"
• "reverse split" + small cap
• "SPAC merger" + historical
• "Russell 3000 components" / "S&P 500 components" + historical date
• "manager track record" / "fund index"

What this skill is NOT

This is not a delisting database. It does not provide the historical membership data — that requires a paid feed or careful SEC reconstruction. It encodes the rules to spot when a universe is secretly survivor-filtered, with specific attention to the small-cap patterns (reverse-split-delist, ATM-into-delisting, SPAC-merger flips) where the bias is largest.