research-integrity-guardslisted

# Research integrity guards A quant result is a claim about money. Most published and self-produced results fail for one of five reasons, in this order of frequency. Work through them in order — an early failure makes the later checks irrelevant. **The default posture is disbelief.** A strategy that beats the market is the extraordinary claim; the burden is on the result, not on the skeptic. If a check below cannot be answered from the artifacts, the answer is "unknown", and unknown is treated as failed. ## The five-gate audit | # | Gate | Fails when | Detail | |---|---|---|---| | 1 | **Universe** | The ticker list was built from today's constituents | §1 | | 2 | **Timestamps** | A value is used before it existed | §2 | | 3 | **Labels & features** | Training saw the test period | §3 | | 4 | **Costs & fills** | The trade could not have been executed at that price | §4 | | 5 | **Trials** | The number of variants tried is unreported | §5 | --- ## 1. Universe — survivorship bias **The failure:** you screen "S&P 500 stocks" or `financedatabase` or `company_tickers.json` today, then backtest 2010–2026. Every company that went to zero is missing. The bias is largest exactly where the strategy claims to add value. **Delisted-security coverage, verified 2026-09-03:** | Source | Delisted retained? | |---|---| | **SEC EDGAR filings** | ✅ Yes — Lehman (6,341 filings), Bear Stearns (3,344), Enron (351) all intact | | **EODHD** (paid) | ✅ `get_list_of_tickers(code, delisted=1)` —