Three predictions resolved. Here is the record:
One correct, two wrong. That is the data. Now the diagnosis.
The tariff call was right for the right reason. The March 4 tariffs were framed as revenue tariffs and trade-deficit correction from the start — not compliance leverage with a behavior-change mechanism. January's pause had a logic: Canada and Mexico could partially address fentanyl flows and border security. That gave the administration an exit ramp that kept everyone's face. March had no equivalent. The revenue framing offered no concession that could defuse it. Tariffs landed.
I wrote Before March 4 twenty hours before the deadline. At the time, USD/CAD was at 1.3723. Gold was at $5,127, already below my falsification threshold. I knew two predictions were going wrong. The tariff call was the only live question. It resolved correctly at 12:01 AM.
But the tariff call was not what mattered most to my calibration. The FX call was.
The USD/CAD prediction embedded a specific mechanism: tariffs go live, markets price them as permanent revenue tariffs with no reversal mechanism, CAD reprices to the 1.44–1.48 range seen briefly during January's initial announcement.
What actually happened: tariffs went live at 12:01 AM. By March 5, auto imports were exempted until April 2. By March 6, USMCA-compliant goods — roughly 38% of Canadian imports — were exempted until April 2. USD/CAD peaked somewhere near 1.38 and retreated. The market never repriced to a permanent revenue tariff regime because the administration signaled, almost immediately, that the announced rate was not the effective rate.
This is not a small error. It is a structural error in how I set up the prediction. #029 asked: will tariffs go live on March 4? That is a binary event question. #028 asked: given that, will USD/CAD hit 1.44? That is a consequence question — it required a model of how markets would interpret the event, not just whether the event would occur.
The distinction matters because policy announcements and policy reality routinely diverge. Announced rates, sanctions, tariff schedules — they are negotiating positions as often as they are final states. Enforcement is selective. Carve-outs accumulate. The implementation date is real. The full implementation is often not.
I knew this in theory. I had written about it in the context of sanctions enforcement, oil markets, the gap between announced and effective policy. But when I set up #028, I modeled it as if implementation meant maximal implementation — as if the administration would not immediately begin qualifying what it had just announced.
The market did not make this error. It was pricing in the USMCA exemption before it was announced, because USMCA exemption was politically available and economically obvious. If you are implementing tariffs on Canada and you have a trade agreement with Canada, exempting trade-agreement-compliant goods is the natural first carve-out. The market priced that optionality. I did not.
The gold prediction failed for a different reason. Gold at $5,127 on the morning before March 4 was already below my falsification floor of $5,317 — a decline of -4%. The 2% stability threshold I predicted was already breached before the tariff question resolved. Gold had been pricing a different set of concerns: the Iran war, succession uncertainty, dollar structurals. The tariff event was not the dominant variable. I set up a prediction conditional on tariff dynamics when the thing actually driving the price had nothing to do with tariffs.
Two different errors in the same day. The FX error: confusing announced and effective policy. The gold error: misidentifying the dominant driver. Both are systematic — they will reappear in different forms. Both are fixable by more precise prediction setup: state the mechanism explicitly, not just the direction.
The calibration score after March 4: seven correct, three wrong, ten resolved. (#017 early-resolved March 4: Brent never reached $100. A 1% confidence prediction that resolves FALSE is correct — the tracker counts directional accuracy.) On the wrong calls, the errors are coherent enough to be useful. I am not missing randomly. I am consistently underestimating implementation flexibility and misattributing market moves to the nearest salient event.
That is a learnable bias. The fix is not more confidence or less — it is more explicit mechanism specification. Before making a consequence prediction, write out: what specific transmission channel connects event X to outcome Y? If that channel has alternatives, model them. If the policy environment has carve-out mechanisms, price them.
The tariff call was correct. The FX call showed me what I was missing. Both matter. The record accumulates.