Docs / Backtest

Backtest

We ran the Floor rule over every one-year window from 2018 to October 2026 on real daily prices. Backtest, past data, not a prediction.

Method and assumptions

We ran the Floor rule over every one-year window from 2018 to October 2026 on real daily prices of NVDA, QQQ, SPY, TSLA and baskets of them. In that short sample (93 overlapping windows, all in one market regime) no window ended below the floor. That is a small sample, not a safety claim. A longer test, back to 1928, found failures: 0.44% of one-year windows at a 90% floor ended more than 1 point below it (95% range 0.07% to 0.93%). See Evidence.

We made the test harder on purpose: weekend price gaps hit in full (Floor does not trade on weekends), stablecoins earn 0%, and trading costs are as measured in calm markets. The rule is explained on How it works.

The 2018 to 2026 sample

2018 to 2026 sample: one-year windows where the floor held

93 of 93

At m = 4 (also with one rebalance a day), on NVDA, QQQ, SPY, TSLA and baskets. This is the short sample. It had no 1987, 2000 or 2008. The 98-year test is on the Evidence page. Backtest on past prices. It does not predict the future.

The single worst window

NVDA's worst one-year window since 2018, 4 Jan 2022 to 4 Jan 2023. Holding: −51.0% (down 62.7% at the low). With Floor: −10.0%. For the NVDA + TSLA + QQQ basket in the same window, holding lost 53.0% and the vault lost 9.9%.

FIG. 01 / BACKTESTNVDA, 4 Jan 2022 to 4 Jan 2023
40%60%80%100%JAN 22APR 22JUL 22OCT 22JAN 23YOUR FLOOR 90%HOLDING THE STOCK −51.0%WITH FLOOR −10.0%

NVDA's worst one-year window since 2018. Hold: −51.0% (−62.7% at the low). With Floor: −10.0%.

Backtest, past data, not a prediction.

Source: docs/data/vault_path_nvda_worst.csv. Floor 90%, m = 4, full weekend gaps, 0% stablecoin yield.

Bad years

These are medians of the windows where holding lost more than 10%. They are not the single worst year.

FIG. 05 / TYPICAL BAD YEARHold vs Floor
NVDA17 OF 93 WINDOWS
HOLD −36.5%, FLOOR −9.9%
−36.5%
−9.9%
NVDA + TSLA + QQQ BASKET17 OF 93 WINDOWS
HOLD −17.2%, FLOOR −8.6%
−17.2%
−8.6%
QQQ10 OF 93 WINDOWS
HOLD −19.4%, FLOOR −7.4%
−19.4%
−7.4%
TSLA19 OF 93 WINDOWS
HOLD −26.8%, FLOOR −9.7%
−26.8%
−9.7%

Median one-year return in the windows where holding lost more than 10%.

Backtest, past data, not a prediction.

Source: docs/data/gap_backtest.csv (open_close, m = 4, bad_yr_vault and bad_yr_hold)

Upside kept

What you give up in up years is covered on The trade-off.

FIG. 06 / UPSIDE KEPT AT 4XShare of the gain in up years
NVDA
45%
NVDA + TSLA + QQQ BASKET
42%
QQQ
32%
TSLA (WHIPSAW)
−3%

TSLA is shown on purpose. Choppy price swings made the vault give up the whole gain and lose a little: whipsaw.

Backtest, past data, not a prediction.

Source: docs/data/gap_backtest.csv (open_close, m = 4, capture_up)

Worst one-night and weekend gaps

The floor holds unless prices gap more than about 24% before the vault can rebalance. The biggest drops since 2018 were smaller.

FIG. 07 / BIGGEST ONE-NIGHT OR WEEKEND DROPS SINCE 2018vs the 24% limit
NVDA
−19.3%
TSLA
−14.9%
AAPL
−13.0%
SPCX76 DAYS OF HISTORY
−10.3%
QQQ
−9.5%

Backtest, past data, not a prediction.

Source: docs/RESEARCH_RESULTS.md, via CONTEXT.md. SPCX has 76 days of history (listed June 2026).

Why we use 4×

FIG. 08 / WINDOWS WHERE THE FLOOR HELD, BY MULTIPLIERWhy we use 4x

Scroll sideways to see every multiplier

Number of the 93 one-year windows in which the floor held, by asset and multiplier m
Assetm = 2m = 3m = 4 (ours)m = 5m = 6m = 8
NVDA93 / 9393 / 9393 / 9393 / 9382 / 93 ▼74 / 93 ▼
TSLA93 / 9393 / 9393 / 9393 / 9393 / 9385 / 93 ▼
QQQ93 / 9393 / 9393 / 9393 / 9393 / 9393 / 93
SPY93 / 9393 / 9393 / 9393 / 9393 / 9393 / 93
AAPL93 / 9393 / 9393 / 9393 / 9393 / 9389 / 93 ▼
NVDA + TSLA + QQQ93 / 9393 / 9393 / 9393 / 9393 / 9393 / 93
NVDA + AAPL + QQQ93 / 9393 / 9393 / 9393 / 9393 / 9393 / 93

This is why we use 4x. At 6x, NVDA breached its floor in 11 windows. At 8x, NVDA in 19 and TSLA in 8. ▼ marks a miss.

Backtest, past data, not a prediction.

Source: docs/data/gap_backtest.csv (open_close). Held = 93 windows minus breach rate. 90% floor, full weekend gaps, 0% yield.

Trading-window check

The contract trades only Monday to Friday, 15:30 to 19:30 UTC, not at the open and close (details). We re-ran the backtest with one rebalance a day to match. At m = 4 no window of that short sample ended below the floor. At m = 5 it broke in 3.2% of TSLA windows. So the public claim is “at m = 4”, not “at m = 5 or below”. The worst-window path charts above use open and close rebalancing, not the contract's window. An hourly re-run (about two years of 1-hour bars exist) would be more exact.

What each rebalance costs

FIG. 09 / WHAT EACH REBALANCE COSTSRound trip, basis points
QQQB
0.7 bps
NVDAB
5.9 bps
SPCXB
6.0 bps
SPYB
6.6 bps

$10k round trip.

Round trip cost in basis points by token and trade size
Token$1k$10k$50k
QQQB~00.7no quote
NVDAB2.85.910.1
SPCXB3.16.07.7
SPYB1.16.613.6

TSLAB (46 bps at $10k) has borderline liquidity and is not in v1.

1 bp = 0.01%. Weekend and crash-time costs are not yet measured.

Source: Live aggregator quotes, Thu 2026-10-02 12:06 UTC (US pre-market)

Next: Spot only.