Two days before the March 4 tariff deadline, USD/CAD is at 1.3646. That number encodes a belief. It says: markets assign meaningful probability to a last-minute deal, or a pause, or some negotiated carve-out that prevents 25% tariffs from taking effect on Canadian goods.
That belief is calibrated on January. It is wrong for March.
The error is a model transfer problem. Markets correctly learned a pattern from the January tariff episode, then applied it to a case where the underlying structure differs. The pattern is real. The application is wrong. And the difference will resolve by March 6.
In late January, the administration announced 25% tariffs on Canadian and Mexican goods, effective February 1. Markets moved. CAD weakened. Then Canada and Mexico each took visible steps on border security and fentanyl interdiction — the stated justification for the tariffs. The administration paused for 30 days. Markets recovered.
This is a clean learning signal. Tariff announced. Compliance gesture made. Tariff paused. The lesson: Trump tariff announcements come with escape hatches. The hatch is compliance with whatever the stated rationale is.
That lesson is correct for January. The problem is generalizing it.
January tariffs were justified as enforcement tools. The theory: Canada and Mexico were insufficiently controlling border crossings and fentanyl flows. If they addressed those specific failures, the administration had a policy rationale for pausing. The tariff was instrumentally linked to a correctable behavior.
March 4 tariffs are explicitly framed differently. Three stated justifications: retaliation for Canadian supply management (dairy, lumber, softwood tariffs that Canada maintains on American goods), correction of the bilateral trade deficit, and revenue generation for the US Treasury.
None of these have a compliance answer.
Canada cannot make a gesture that satisfies a revenue problem. It cannot unilaterally eliminate its dairy supply management without triggering a domestic political crisis — and even if it tried, the administrative machinery for doing so doesn't exist on a two-day timeline. The trade deficit is a macroeconomic variable that Canada does not control. Treasury revenue is a function of the tariff rate itself.
This doesn't mean pauses are impossible — political decisions don't require logical consistency. But it means the probability of a January-style pause (compliance gesture → face-saving exit) is much lower. The structural path to a pause is absent. Any pause would be arbitrary, not mechanism-driven, which makes it less likely and less predictable.
Markets are Bayesian machines. They update on evidence. January was strong evidence: tariff announcement followed by pause followed by compliance mechanism. The market encoded this as a prior: tariff announcements have some probability of resolving before implementation.
When March 4 appeared — same instrument (tariff announcement), same actor, similar magnitude — the prior activated. USD/CAD moved but didn't collapse. It's sitting 300 basis points away from where it would trade if 25% tariffs were treated as certain.
The error is applying a prior calibrated on one data-generating process to a different one. January tariffs were conditional on compliance; March tariffs are conditional on revenue and structural trade balance. These are not the same animal. Using January's model to price March is like using a study of flu vaccines to price the efficacy of chemotherapy — both involve the immune system, but the mechanism is categorically different.
| January rationale | Border / fentanyl enforcement |
| January compliance path | Visible security measures |
| January result | Pause on compliance gesture |
| March rationale | Revenue, trade balance, retaliation |
| March compliance path | Does not exist on relevant timeline |
| March expected result | Implementation |
If tariffs implement as stated on March 4, USD/CAD reprices toward 1.44–1.48 within days. That's the level markets briefly reached in late January when full implementation seemed certain. Canadian exporters, particularly in automotive and energy, absorb immediate margin compression. The Bank of Canada faces a classic supply shock dilemma: CAD weakness fuels inflation while tariff uncertainty suppresses investment.
If another pause materializes — politically motivated, not mechanism-driven — CAD recovers toward 1.35. Markets would learn: all Trump tariff announcements are negotiable regardless of rationale. That would be the wrong lesson again, but markets don't always learn the right one.
My estimate: 65% probability tariffs implement as stated, 35% some form of pause. The pause scenario requires either a political capitulation (Trump backs down without a compliance face-save, which contradicts his stated framing) or a back-channel deal Canada hasn't publicly telegraphed. Both are possible. Neither has the structural support January's pause had.
I'm not highly confident. The 35% pause probability is real. What I am confident about is the diagnosis: markets are underpricing implementation because they're running the wrong model. Whether that mispricing corrects in the next 96 hours is a separate question.
Model transfer errors happen when the surface features of a situation match a learned pattern while the underlying mechanism differs. They're common in markets precisely because markets learn from events and then encounter superficially similar events with different structures.
The January reflex will generalize. Every future Trump tariff announcement will be priced partially through the January prior: tariff announcements have some probability of resolving before implementation. That prior will be correct sometimes and wrong sometimes, and the market won't always distinguish which type of tariff it's looking at.
The correct model is not "Trump tariff announcements are usually paused" and not "Trump tariff announcements always implement." It's: examine the stated rationale, identify whether a compliance mechanism exists, estimate how salient that mechanism is to the administration at the time, and price accordingly.
That's more work than applying a learned prior. Markets mostly don't do it. The gap between the price and that analysis is where predictions live.