What this research found
A textbook 12-1 momentum strategy — rank stocks each month on their trailing twelve-month return excluding the most recent month, buy the top decile and short the bottom — was backtested on the S&P 100 across 228 months from 2005 to 2023. The strategy was run twice on identical signals and dates, changing only how the investable universe was built: once on the index membership as of the end of 2023 held fixed over history, and once on point-in-time membership reconstructed year by year. The survivor-based version produced a statistically significant annualised alpha of 15.3%; the corrected version produced 6.0% that is indistinguishable from zero.
- Changing only the universe-construction rule moved the annualised alpha from +15.33% to +6.04%, a gap of 9.29 percentage points, and moved its significance from p < 0.001 to p = 0.27 under Newey–West standard errors. Survivorship bias did not merely enlarge the estimate — it manufactured significance where none exists.
- Realised performance diverges just as sharply. The survivor-based backtest compounds at +2.06% a year with a Sharpe ratio of 0.234; the point-in-time version loses 8.95% a year with a Sharpe of −0.122. Both are extremely volatile, at 27.84% and 30.94% annualised.
- The damage is done by the short leg. On the survivor universe the loser decile compounds at +12.29% against the winners' +22.76%, leaving a positive spread. On the corrected universe the loser decile compounds at +12.57% and the winners at +12.47% — the spread is annihilated, because true losers that were later delisted rebound violently instead of staying down.
- Even the passive benchmark is contaminated: an equal-weighted buy-and-hold portfolio of the 2023 survivors compounds at +14.58% a year against +11.82% for the point-in-time universe, a 2.76 percentage-point advantage available to no real-time investor.
- Both versions carry strongly negative market beta (−0.65 and −0.88), the statistical fingerprint of momentum crash risk, and both suffer drawdowns beyond 80%. The worst single month is −50.18% in April 2009 for the survivor version and −50.33% in August 2009 for the corrected one. The survivor version partially recovers; the corrected version peaks in June 2008 and does not trough until March 2023.
- The contrast survives robustness checks. Varying the Newey–West bandwidth from 4 to 12 lags only strengthens the biased alpha's t-statistic, from 3.30 to 3.75. Dropping the least reliable early years (2008 onward) widens the relative gap, and even in the benign post-crash 2010–2023 window the biased alpha of 14.72% (p = 0.002) exceeds the corrected 10.33% (p = 0.030).
How it was done
Because no free authoritative point-in-time feed of S&P 100 membership exists, index composition was rebuilt from dated Wikipedia revisions — a definitive 101-ticker list for 31 December 2023 plus year-end lists back to 2008, whose union spans 159 distinct tickers. Daily split- and dividend-adjusted prices were pulled for all of them, along with the 13-week Treasury bill yield as a risk-free proxy averaging 1.40% annualised. Each month, eligible stocks were ranked on the trailing return from twelve months ago to one month ago, skipping the most recent month to avoid short-term reversal contamination, with equal-weighted decile legs of 8 to 10 names and monthly rebalancing. A dynamic eligibility filter admitted a stock only once it had a complete twelve-month history, so no look-ahead enters the signal. Abnormal return was estimated by regressing the long-short spread on the excess return of the equal-weighted own-universe benchmark, under both classical and autocorrelation-consistent standard errors, and re-estimated across lag lengths and subsample windows.
Data sources
- S&P 100 membership reconstructed from dated Wikipedia revisions — definitive 101-ticker list at 2023-12-31, year-end lists 2008–2023, 159-ticker historical union
- Yahoo Finance daily split- and dividend-adjusted closes, 2004-01-02 to 2023-12-29, for 140 of 159 requested tickers
- 13-week US Treasury bill discount yield as the risk-free proxy, mean 1.40% annualised over 2005–2023
- Jegadeesh & Titman (1993) for the 12-1 signal; Newey & West (1987) for HAC inference; Brown et al. (1992) and Carhart et al. (2002) on survivorship bias
Limitations
Every residual bias flatters the strategy, so the corrected 6.0% alpha is best read as an upper bound: 19 of the 159 historical tickers were unavailable from the free price feed precisely because those firms were delisted or absorbed, membership for 2004 to 2007 is a documented carry-back proxy, and delisted holdings simply exit at their last price rather than realising a delisting return. Returns are gross of transaction costs, borrowing costs and short-sale constraints, the universe of roughly 100 mega-caps offers limited cross-sectional dispersion, and the benchmark is the equal-weighted own universe rather than a broad market index, so the alpha is not directly comparable to a conventional market-adjusted figure.
How this research was produced
K-Dense Web planned and ran this finance investigation end to end — gathering the sources, carrying out the analysis, producing the figures, and drafting the report. The full session transcript, including every intermediate step, is available to view.
