Survivorship-free backtest research — corpus expansion & literature grounding

1 · What was already in the corpus (2025-07→ harvest) — now citable for this work

BM25 sweeps (10 themed queries) over the 114k-doc corpus found real coverage in three clusters:

Selection/overfitting protocols (measure the inflation):

• 2606.01650 Post Selection Estimation of Sharpe Ratios — estimators for the true Sharpe of

the asset picked for having the max in-sample Sharpe (polyhedral lemma, shrinkage, empirical Bayes). The direct modern tool for goal (b).

• 2603.09219 AlgoXpert: rigorous IS/WFA/OOS protocol for mitigating overfitting;

2602.00080 GT-Score: robust objective reducing overfitting; 2602.10785 double out-of-sample parameter optimization — protocol templates for re-running the backtest.

• 2604.08356 Measuring Strategy-Decay Risk: minimum regime performance; 2512.11913 *Not

All Factors Crowd Equally: alpha decay* — durability framing for whatever survives.

• 2606.19550 Which Portfolios? Construction Dependence of Factor Model Performance — the

result depends on universe construction; exactly the sensitivity we're testing.

Survivorship/look-ahead measurement (named, recent, quantified):

• 2603.19380 Survivorship Bias in Emerging Market Small-Cap Indices (NIFTY Smallcap 250) —

the corpus's one direct index-survivorship quantification; template for measuring ours.

• 2603.20237 Temporal Coverage Bias in Financial Panel Data — coverage-aware panel

structuring; the panel-level generalization of our missing-dead-tickers problem.

• LLM-era look-ahead cluster (relevant because our loop uses LLMs): 2601.13770

Look-Ahead-Bench (point-in-time LLMs), 2603.11838 DatedGPT, 2512.23847 Detecting Lookahead Bias in LLM Forecasts, 2510.11677 Chronologically Consistent LMs, 2605.24564 Mitigating Look-Ahead Bias in Financial Backtesting with LLMs.

Strategy-adjacent: 2601.06074 Clarifying Note on Long-Horizon Investment and Dollar-Cost Averaging; 2603.01298 Single-Asset Adaptive Leveraged Volatility Control; 2607.00883 Tail Risk Management with Puts and Trend Following; 2607.01550 Is Trend Still Your Friend? (microstructural demise of short-term trend).

The gap the sweeps exposed: zero pre-2025 canon — no CFM long-history papers, no max-Sharpe selection-bias inference, no drawdown-significance work, nothing on constituent reconstruction. Hence the tranche.

2 · The 11-paper tranche (added 2026-07-18, PDFs in F:\arxiv-corpus\pdfs\)

All ids web-verified, then title-matched against the arXiv API before ingest (arxiv_add_papers.py --dry-run; the dry-run also caught PowerShell double-parsing eating trailing zeros fromunquoted ids — quote id lists).

Not addable (not on arXiv) — cite directly: Bailey, Borwein, López de Prado, Zhu Pseudo-Mathematics and Financial Charlatanism (AMS Notices 2014; SSRN 2308659) — expected max in-sample Sharpe grows with configurations tried, and under memory effects overfit backtests have negative expected OOS returns; The Deflated Sharpe Ratio (SSRN 2460551); Pav A Short Sharpe Course (SSRN 3036276); Bessembinder Do Stocks Outperform Treasury Bills? (JFE 2018); Ammann, Burdorf, Liebi, Stöckl Survivorship and Delisting Bias in Cryptocurrency Markets (SSRN 4287573); Brown-Goetzmann-Ross (1995) and Elton-Gruber-Blake (1996) on fund survivorship.

3 · Free-data recipe for the point-in-time universe (goal a)

Not arXiv material — engineering references, recorded here:

• Wikipedia "List of S&Pcompanies" maintains a selected changes table

(additions/removals with dates); the page's revision history gives point-in-time snapshots back to ~2007-2008. Two worked implementations: teddykoker/survivorship-free-spy (constituent CSVs + Python walkthrough) and riazarbi's reconstruction (Wikipedia-revision method, documents its gaps).

• Delisted-ticker roster: Alpha Vantage LISTING_STATUS (free key, memory

options-iv-data-sources) returns active and delisted symbols with listing/delisting dates — the roster of the dead, though not their price history.

• Dead-company prices remain the hard part on free data: Yahoo drops delisted tickers

entirely. Stooq carries some; Alpaca (when :4177 is back) covers ~2016+ including many since-delisted symbols; IBKR historical data needs Alex to refresh stale creds (memory ibkr-local-owner-creds-stale). Anything unreachable gets bounded, not ignored: delisting-return assumptions bracketed per Shumway-style corrections (−30%/−55% NYSE/AMEX vs ~−90% Nasdaq performance-delistings are the literature's brackets; verify before using).

• Honesty rule from the repo's own precedent (dca_walkforward_sim.py): instruments enter

the sim only when they verifiably existed with history; where the universe can't be reconstructed, the claim shrinks to the reconstructable window instead of extrapolating.

4 · Published bias magnitudes (goal b's comparison row)

Priors the measured hand-picked-winners inflation should be compared against — each [claim → source → confidence]:

• Equity mutual-fund survivorship inflates annual returns ~1-2%/yr (BGR 1995) /

~1.4%/yr, growing with horizon (EGB 1996) → journal classics, abstract-level → 0.9 (magnitudes widely replicated).

• Crypto universe truncated by survivorship+delisting: **+0.93%/yr value-weighted, +62%/yr

equal-weighted; size premium overstated ~50% (Ammann et al., 3,904 coins 2014-21) → SSRN→ 0.85**. The EW number is the cleanest warning for small hand-picked universes.

• **~4% of US stocks account for all net wealth creation above T-bills since 1926; the median

stock's lifetime buy-and-hold return trails one-month T-bills (Bessembinder JFE 2018) → 0.9**. This is the direct counterfactual to "the GEs, Ciscos, Lehmans, Nokias": a uniform draw from the historical universe most likely underperforms cash; hand-picking today's survivors samples the 4%.

• Expected max in-sample Sharpe of N unskilled configurations grows ~√(2·ln N); overfit

selection ⇒ negative expected OOS under memory effects (Pseudo-Mathematics, AMS 2014) → 0.9.

• Index-membership look-ahead (using today's constituents historically) is quantified for an

EM small-cap index in-corpus (2603.19380) — read before citing numbers → 0.6 until read.

• What honest long-history looks like: trend on indices t-stat ≈over two centuries

(1404.3274) — significance from breadth×time, not from picking winners → 0.85.

What this tranche does NOT license: any alpha claim. It grounds methodology and magnitudes. The measured $2M-deflation number must come from our own reconstruction (Σ₀: our deployments' evidence must be our own measurements), with the literature as the sanity band — mirroring how PR #2728 validated the overlay against published trend work.

5 · Next steps (for the follow-on iteration on this branch)

1. Rebuild the universe: Wikipedia-revision S&Pmembership (2008→) + Alpha Vantage

delisted roster; freeze as a point-in-time JSONL under data/ with provenance per row.

1. Re-run the single-stock DCA on (i) hand-picked winners, (ii) point-in-time index members

with delisting brackets — the spread between (i) and (ii) IS the measured inflation; deflate (i)'s Sharpe per the max-of-N corrections (1906.00573, 2606.01650).

1. Log the convergence record: hypothesis (inflation ≥ X%/yr), evidence (both runs + corpus

cites), confidence split (measured vs literature-bracketed), sources (this doc).