# Winner strategy backtest audit

Audited 12 October 2026. Prices through 9 October 2026. This is a dated audit of the reconstruction, not a claim that the original author’s unpublished inputs have been recovered.

## Findings

The calculation engine agrees with a separately written Python implementation. No change to the historical portfolios or their returns was required by the checks performed. The automatic updater needed stronger validation, which has been added. Historical source coverage remains incomplete in the places described below.

MRNA gained 662.97% during 2026 through 9 October, versus SNDK’s 566.37%. The strategy chooses holdings from the **previous calendar year**. MRNA fell 29.08% in 2025; SNDK gained 559.39% from its first available 2025 close. Therefore the current-year comparison does not invalidate the 2026 selection. The page now explicitly labels the selection year and the year whose returns determined it.

## Calculations and prices

- Independently recalculated 1,800 combinations: every available start year and Top 1/2/3 portfolio, thread/total/price return bases, 10/35/60% stops, and permitted December-close/January-open and ideal/gap fill choices. Starting cash was $1,234.56, separate from the $100 default.
- Compared ending values, CAGR, drawdown, every annual return and stop record, and weighted sums of every daily valuation. There were 4,270,221 valuations in those runs. Relative numerical tolerance: 2e-11.
- Compared every one of the default reconstruction’s 6,742 observations individually, including strategy, no-stop portfolio, SPY and cash values.
- Checked all 6,738 daily trading dates against the NYSE calendar, including the 2001 attacks, presidential funerals and Hurricane Sandy. Checked exact prior-December anchors and positive, aligned prices. The remaining three observations are annual records; the initial cash observation is separate.
- Reconciled 3,526 older held-day bars with preserved raw WIKI records and independently accumulated share splits. Rebuilt every 2014–2026 held-year dataset from its preserved Yahoo/archive inputs and checked the live rows. Later dividend adjustment rescaling was compared by normalized growth, rather than treating an equivalent constant factor as a price disagreement.
- Confirmed cash scaling, equal initial allocations with subsequent weight drift, individual stops, no redistribution of stopped cash, and annual re-entry. The stop is measured from entry, not from the stock’s later peak.

The default Top 1 strategy stops in 2000, 2002, 2008 and 2011, matching the thread. Reconstructed daily stop dates are 28 January 2000 (QCOM), 28 March 2002 (NVDA), 16 September 2008 (NOV) and 29 September 2011 (NFLX). NOV opens below its stop; the default still assumes the ideal stop fill, as explicitly labelled.

## Selection checks

The 2025 year-end universe contains 503 equity share classes. It was obtained from dated official iShares IVV holdings, not today’s constituents. Fresh provider histories ranked 492 of them. Eleven renamed, delisted or unavailable histories were recovered using official 31 December 2024 and 31 December 2025 fund unit prices matched by CUSIP/ISIN. Their raw closing-price gains are below the selected top three; their vanished daily histories were not substituted with acquirers.

For securities present in both official snapshots with available provider histories, annual returns matched the official closing prices after the provider’s reported corporate-action factors. There were no discrepancies beyond the 0.01-percentage-point tolerance. Securities absent from the prior snapshot retain the provider history check; the first-listed SNDK history is an explicit exception.

| 2026 holding | 2025 measurement | Adjusted price gain |
| --- | --- | ---: |
| SNDK | 13 February–31 December 2025 | 559.389% |
| WDC | 31 December 2024–31 December 2025 | 282.212% |
| MU | 31 December 2024–31 December 2025 | 239.128% |

SNDK’s February 13 observation is a when-issued quote, before regular-way trading began on February 24. This matches the thread’s first-available-quote convention. WDC’s history includes the provider’s 1.323 corporate-action factor for the Sandisk separation. The displayed ranking is a corporate-action-adjusted price comparison; ordinary cash dividends are excluded. It is not the return of retaining both spin-off companies as separate holdings. [Sandisk/Western Digital SEC separation filing](https://www.sec.gov/Archives/edgar/data/106040/000119312525033383/d847507d8k.htm), [Western Digital distribution filing](https://www.sec.gov/Archives/edgar/data/106040/000010604025000012/wdc8-khardxspin.htm).

Earlier Top 1 holdings match the thread and dated historical membership. Earlier Top 2/3 selections use contemporary performance tables, membership filters and recalculated prior-year price gains. A broad WIKI screen corroborated the 2014–2018 held-year selections, but is not a complete independent constituent ranking: it lacks some delisted securities and contains corporate-action errors. For example, EVHC’s December 2016 merger exchanged each old share for 0.334 new shares; its uncorrected archive jump would falsely place it near the winners. That record was not used to change the portfolio. [SEC merger terms](https://www.sec.gov/Archives/edgar/data/1678531/000119312516787523/d288696d8k.htm).

**A complete independent all-constituent certification of every historical winner is not claimed.** The original thread supplies the older Top 1 choices; public archives and contemporary tables provide corroboration. Full historical identity/corporate-action coverage is a remaining source limitation, particularly before 2014. The daily portfolio calculations have been checked for the selected holdings.

## Annual observations and execution limits

Reebok 2001, Providian 2003 and original Avaya 2004 have sourced year-end prices and reported annual extrema. Their daily paths and exact stop dates remain unavailable. Annual extrema were rechecked against company filings, including Avaya’s fiscal-to-calendar-quarter alignment. All three had no common cash dividends.

| Held year | Stock | Entry | Year-end close | Annual low | Annual high |
| --- | --- | ---: | ---: | ---: | ---: |
| 2001 | RBK | $27.34 | $26.50 | $18.50 | $35.75 |
| 2003 | PVN | $6.49 | $11.64 | $5.25 | $13.00 |
| 2004 | AV | $12.94 | $17.20 | $11.95 | $19.00 |

Reported lows allow the ideal fixed-entry stop outcome to be determined, but not its execution date or gap fill. Those years remain clearly marked as annual observations. January-opening and gap-fill controls are disabled for periods containing them. Top 2/3 begin in 2014. Mixed-frequency drawdown can understate intrayear losses; fully daily drawdown is based on closing valuations, not intraday portfolio peaks and troughs. [Reebok ranges](https://www.sec.gov/Archives/edgar/data/770949/000091205702011726/a2073450z10-k405.htm), [Providian ranges](https://www.sec.gov/Archives/edgar/data/1035713/000104746905008606/a2154671zex-13.htm), [Avaya fiscal-quarter ranges](https://www.sec.gov/Archives/edgar/data/1116521/000110465905060564/a05-21334_1ex13.htm). Full close-price provenance is attached to the downloadable dataset.

The default December closing entry uses the same closing observation that determines the winner and is idealized. The optional January opening entry provides a later executable entry. Ideal stop fills can overstate proceeds across gaps; the optional gap setting uses the lower opening price. Daily bars cannot establish every intraday fill or slippage. Dividends use adjusted-close factors, including for stop proceeds; they are a return-series approximation, not an exact dividend-payment/reinvestment ledger. Fees, taxes, cash interest, market impact and other slippage are excluded.

## Comparison with the original thread

Same period, January 1997 through 9 October 2026; $100; Top 1; 35% fixed-entry stop; ideal December-close convention; stock price returns through 2011 and dividend-adjusted returns from 2012; SPY total returns throughout.

| Portfolio | Thread | Reconstruction | Difference |
| --- | ---: | ---: | ---: |
| 35% stop | $1,514,581 | $1,528,103.87 | +0.893% |
| No stop | $89,418 | $90,189.79 | +0.863% |
| S&P proxy | $1,752 | $1,753.76 | +0.101% |

The holdings and four stop years agree. The thread does not publish its full price/adjustment inputs, and many annual bars are rounded. Its exact endpoint has not been reconciled. The chart preserves sourced prices and exposes this difference; it is not calibrated to force agreement. Tighter stop choices can differ further because the thread itself marks some lows as unconfirmed and treats at least one gap specially. [Original stop-strategy post](https://x.com/gurgavin/status/2109399651870720030).

## Changes to future updates

- Normalize the official iShares share-class tickers BRKB/BFB to the provider’s BRK-B/BF-B. Test against the actual 503-row holdings response.
- Require every expected completed NYSE session, including half-days. A missing interior day or newest completed session blocks publication. Known extraordinary closures are included; a future unannounced closure needs calendar review and will block rather than invent a price.
- Require the exact final prior-December session for entry and ranking. An older December close cannot silently stand in for it.
- Permit the first-available listing close only when the provider’s first-trade metadata confirms that date and ranking year. Missing historical anchors for an existing stock block selection.
- Keep the complete-constituent ranking requirement. Missing members or prices block annual selection, rather than reducing the eligible universe.
- Lock simultaneous updater runs and use unique temporary files for atomic replacement.
- Publish refresh status. A failed refresh is shown on the chart immediately, retaining the last valid data. The existing seven-day stale-data notice remains.
- Validate browser snapshots for prices, dates, alignment, coverage and ranking order; invalid live snapshots fall back to the saved valid snapshot.

The schedule remains 06:57, 12:57 and 18:57 America/New_York daily. Only sessions before the current New York date are included. Scheduled update service and an actual refresh were checked after deployment. Public data access, ticker changes, acquisitions and exceptional closures can still require maintenance; these are reported as failures rather than silently altering history.

## Reproducible evidence

Local audit scripts: `audit-engine.cjs`, `audit-reference.py`, `audit-inputs.py`, `audit-rankings.py`, `audit-official-rankings.py`, `audit-wiki-rankings.py`; regression checks: `test-model.cjs`, `test-update.py`, `test-audit-guards.py`. Results and preserved provider/official responses are in the project’s `audit` folder. The dated official 2024/2025 IVV response hashes are saved in `official-ranking-2025.json`. The published data snapshot retains per-stock source hashes and links.