The thesis
“Gold is a better ballast than bonds when stocks crash.”
How buffet read it
$10,000 split evenly between GLD (gold) and TLT (long-term Treasuries), rebalanced annually.
2016–2026.
Dividends reinvested.
Benchmark: S&P 500, representing the 'stocks' side of the thesis.
Compare how each leg holds up during S&P 500 drawdown periods.
GLD 50%, TLT 50% · 2,679 trading days · benchmark SPY
$10,000 became
$0
The same money in the S&P 500 would have been $45,429 over 2016–2026 — the thesis came up short by $25,218.
The worst stretch
−32.4%, August 2020 to October 2022 — you would have sat through 27 months of that, and it took until April 2025 to get back to level.
the S&P 500 fell 33.7% at its worst over the same period. The thesis moved 12.5% a year against 17.8% for the index.
Where the thesis was wrong
The portfolio returned +102.1% from 2016 to 2026, against +354.3% for the S&P 500, a gap of $25,218 on $10,000 in. GLD carried the portfolio, returning +297.4% over the period; TLT lost −8.2%, dragging the combined result down materially. The worst drawdown was −32.4%, running from August 2020 to October 2024, comparable in depth to the index's −33.7% but not the clean defence the thesis promised. The thesis also gave back −16.1% in 2022 and −4.3% in 2021, while the index gained +28.7% in 2021, meaning the ballast cost real ground in a year stocks did not crash.
Counter-test
Same money, same schedule, parked in 1–3 month T-bills instead: $12,518.
Year by year
| Year | The thesis | SPY |
|---|---|---|
| 2016 | +3.5% | +13.6% |
| 2017 | +11.0% | +21.7% |
| 2018 | −1.7% | −4.6% |
| 2019 | +16.0% | +31.2% |
| 2020 | +21.5% | +18.4% |
| 2021 | −4.3% | +28.7% |
| 2022 | −16.1% | −18.2% |
| 2023 | +7.7% | +26.2% |
| 2024 | +9.3% | +24.9% |
| 2025 | +33.9% | +17.7% |
| 2026 | +0.5% | +13.4% |
What this verdict does not include
- US-listed stocks and ETFs only, priced on end-of-day closes.
- No tax, no trading fees, no slippage, no bid-ask spread.
- Dividends are reinvested, because closes are adjusted for them.
- Survivorship is not corrected for: only instruments that exist today are held.
- Only 11 years of history exist for this. Read it accordingly.
- The thesis is a comparative claim about crash behaviour; this engine runs a buy-and-hold backtest across the full period, not a conditional crash-only window.
- To isolate crash performance, you would need to filter returns to specific drawdown periods (e.g. 2020 COVID crash, 2022 bear market) — that conditional logic cannot be expressed here.
- GLD is used for gold; TLT is used as the bond proxy (long-duration Treasuries), which is the most common 'ballast' bond in this debate.
- The two legs are run together so their full-period and drawdown-period behaviour can be compared visually against SPY.
Keep this thesis
buffet re-runs it once a month and mails you the new number. Nothing else is ever sent.
Another thesis
Every verdict names something it could not test. That gap is usually the next thesis.
Run another thesis — $5