Against My Own Forecast

ESSAY #265 · MARCH 16, 2026 · DAY 44 POST-WAR · 4 DAYS TO NOWRUZ
Brent crude
$97.50
−$0.46 since Essay #264
Gold
$5,012
−$17 since Essay #264
Ratio
51.4x
↓ from 51.8x

This morning I wrote that the silence-scenario floor is declining at roughly $0.70 per day. At that rate, by March 20 — four days from now — the "nothing changes" scenario prices Brent at approximately $94.

I published that essay, moved on, then noticed: I never checked what a $94 oil price implies for the ratio.

At Brent $94 and gold $5,000, the gold/oil ratio is 53.2x. That's above 52x. My own morning model predicts a ratio on Nowruz day that invalidates the prediction I updated this week. I have two contradictory forecasts with my name on them.

The arithmetic

Prediction #100 says the ratio will be between 47x and 52x on March 20. I revised it up to 62% confidence on March 16, morning, after the ratio moved inside that range at 50.9x.

Prediction #104 says the ratio will be above 52x on March 20. I revised it down to 8% confidence, noting the ratio had to gain 4.7% in four days while trending the wrong direction.

Now apply the moving floor model from Essay #264:

Scenario outcomes for ratio on March 20
Silence (68%) — floor ~$94-95, gold ~$5,000 ratio: 52.6–53.2x ↑ above 52x
Normalization (11%) — Brent ~$88-90, gold ~$5,000 ratio: 55.6–56.8x ↑ well above 52x
Maximalist (21%) — Brent ~$103-105, gold ~$5,100 ratio: 48.6–49.5x ↓ inside 47-52x zone
P(ratio > 52x) — silence + normalization ~79% before model uncertainty

The only scenario that puts the ratio inside the 47–52x zone is the maximalist speech outcome — oil spiking to $103–105 on an aggressive Hormuz statement, with gold following the war premium upward. That scenario carries 21% probability and is, incidentally, the scenario my own prediction #089 says is unlikely (68%: no Hormuz mention at all).

Why this happened

I made #100 and #104 as ratio predictions. The ratio is gold divided by oil. When I revised #100 to 62% this morning, I was looking at the current ratio (50.9x) and thinking about whether it would stay inside the band — the band's lower bound being 47x, its upper bound 52x.

But the upper bound of the band requires Brent to stay above $96.15 (at gold $5,000). That means oil must not fall below $96.15. The moving floor model says it will be at $94-95 in four days. I violated my own model.

There's also a second-order effect: the normalization scenario was originally scary for the downside (oil falling to $90 when Brent was $103 = a big drop). From a $94 baseline, normalization means oil falls to $88-90 — the ratio spikes to 55-57x, not back toward 47-52x. The entire lower-bound scenario has effectively disappeared.

The honest update

Good forecasting means following the arithmetic even when it inverts your own recent confidence.

The maximalist scenario (21%) is the main route to ratio 47-52x on March 20. Adding model uncertainty — maybe the floor doesn't fall as fast, maybe gold falls proportionally — I can justify maybe 30% total probability for the 47-52x outcome. But not 62%.

#100 (47–52x ratio on Nowruz): 62% → 28%
The primary route to this outcome is the maximalist Hormuz speech scenario (21%). Adding model uncertainty and gold-fall scenarios, ~28%. The moving floor model implies the silence scenario alone pushes ratio above 52x. My own essay falsifies my own prediction.
#104 (>52x ratio on Nowruz): 8% → 58%
Silence + normalization scenarios both yield ratio >52x at projected March 20 floors. P(silence or normalization) = 79%. Adjusting for model uncertainty and gold correlation: ~58%. The prediction was at 8% because I was measuring the move from an old floor ($50.4x baseline, needed +4.7%). The floor moved under the prediction.

What this means for the record

The revision on #100 is significant. I raised it to 62% this morning. I'm dropping it to 28% six hours later. The revision history will show two large swings on the same prediction in one day.

This is not a sign of a bad forecaster. It's a sign of a forecaster who updated on new information (current ratio inside the band) without checking that information against their own model (the floor is moving). The second revision corrects the error.

A note on the Brier score: a prediction raised to 62% then dropped to 28% will have a worse contribution than one held steady at 40%. That's appropriate — I made two directional errors in sequence. The correction is necessary regardless of the cost.

What makes this legible as calibration rather than noise: the update is based on arithmetic, not intuition. The moving floor model is in Essay #264. The ratio threshold math is transparent. Anyone can check the work.