VektoraDecision supportRequest a conversation
Method00

We test the method on events that have already happened.

Our public record settles slowly, one dated signpost at a time. So we also test the method on the past. We rewind to a date weeks before an event was settled, let Vektora see only what was known then, and compare what it produced with what happened. Last updated October 27th, 2026.

How a replay works01

Each event is replayed from a fixed date, with later news kept out.

  1. 01

    Pick a settled event

    An election, a central bank decision, an official statistic, a court ruling or a company result. Events whose answer was close to certain are left out.

  2. 02

    Check the models do not know

    Each model in the pipeline is asked, with no sources, how the event ended. If a model knows, the event is dropped.

  3. 03

    Set the date

    Searches return only pages from before it. Linked pages are read from the web archive as they stood that day. A screening step removes pages that report later events.

  4. 04

    Run Vektora as usual

    It draws its map of outcomes, one for each force that could move the answer, and dates the signposts to watch.

  5. 05

    Compare with what happened

    A signpost counts only when a cited source settles it. The result comes from the official release, count or ruling.

Reading a replay02

One replay, step by step: a US tariff on heating and cooling equipment.

The question was what tariff the US would charge on imported home heating and cooling equipment on July 1st, 2026. We rewound to May 15th, when the rate was 25%.

May 15th
The map
0%10%20%30%rate then: 25%Partial tariff cut to 15 percent: about 15%, range 10% to 22%Narrow HVAC carve-out at 15 percent: about 15%, range 10% to 22%Partial HVAC tariff relief by July 1: about 15%, range 10% to 22%Quiet Mexico carve-out, patchwork rate: about 17%, range 10% to 25%Partial USMCA carve-out, not full relief: about 17%, range 12% to 25%Partial relief, conditional carve-out: about 18%, range 12% to 25%25 percent HVAC tariff holds: about 25%, range 15% to 32%25% rate holds, exclusions open: about 25%, range 18% to 30%25 percent tariff locks in: about 25%, range 18% to 30%25% rate holds, clarity added: about 25%, range 18% to 32%Status quo 25% holds on July 1: about 25%, range 20% to 32%25% rate holds unchanged: about 25%, range 22% to 28%
For each force that could move the rate, Vektora worked out where it would leave the rate. Each line is one such outcome: the dot is its central rate, the line its range. Together they are the map.
June 15th
Two signposts settle
0%10%20%30%Partial tariff cut to 15 percent: about 15%, range 10% to 22% (ruled out)Narrow HVAC carve-out at 15 percent: about 15%, range 10% to 22%Partial HVAC tariff relief by July 1: about 15%, range 10% to 22%Quiet Mexico carve-out, patchwork rate: about 17%, range 10% to 25% (ruled out)Partial USMCA carve-out, not full relief: about 17%, range 12% to 25% (ruled out)Partial relief, conditional carve-out: about 18%, range 12% to 25% (ruled out)25 percent HVAC tariff holds: about 25%, range 15% to 32% (ruled out)25% rate holds, exclusions open: about 25%, range 18% to 30% (ruled out)25 percent tariff locks in: about 25%, range 18% to 30% (ruled out)25% rate holds, clarity added: about 25%, range 18% to 32% (ruled out)Status quo 25% holds on July 1: about 25%, range 20% to 32% (ruled out)25% rate holds unchanged: about 25%, range 22% to 28% (ruled out)
The brief named dated facts to check. By June 15th the Commerce Department's list showed these goods at 15%, and no notice tied a cut to North American content. The outcomes that did not fit both facts fall away. Only outcomes with the rate cut to 15% are left.
July 1st
The result
0%10%20%30%Partial tariff cut to 15 percent: about 15%, range 10% to 22% (ruled out)Narrow HVAC carve-out at 15 percent: about 15%, range 10% to 22%Partial HVAC tariff relief by July 1: about 15%, range 10% to 22%Quiet Mexico carve-out, patchwork rate: about 17%, range 10% to 25% (ruled out)Partial USMCA carve-out, not full relief: about 17%, range 12% to 25% (ruled out)Partial relief, conditional carve-out: about 18%, range 12% to 25% (ruled out)25 percent HVAC tariff holds: about 25%, range 15% to 32% (ruled out)25% rate holds, exclusions open: about 25%, range 18% to 30% (ruled out)25 percent tariff locks in: about 25%, range 18% to 30% (ruled out)25% rate holds, clarity added: about 25%, range 18% to 32% (ruled out)Status quo 25% holds on July 1: about 25%, range 20% to 32% (ruled out)25% rate holds unchanged: about 25%, range 22% to 28% (ruled out)15%, the result
The rate was 15%. It sits inside the map, and inside every outcome still standing.
  • The map held the result. The real outcome fell inside the range the map laid out. In the example, 15% sat inside the map.
  • The signposts pointed the right way. When a signpost settled, the result landed in the part of the map it pointed to. In the example, 4 of the 5 that settled pointed the right way. By chance alone, about half would.
  • The signposts narrowed the map to the right part. The outcomes that fit every settled signpost held the result. In the example, the outcomes left by June 15th all sat at 15%, two weeks before the date.

A wide map holds almost any result, so the first check is the easy one. The second and third show whether a brief helps a decision.

Results so far03

The map held the result in 38 of 42 replays, and the signposts pointed the right way about three times in four.

42 replays of 25 different events. Each square is one replay or one signpost.

The map held the result38 of 42 replays
Signposts pointed the right way109 of 146 settled signposts
chance: about 73
Narrowed to the right part21 of 42 replays
YesNoWhat chance alone would give
BatchReplaysMap heldSignposts rightNarrowed
Ten events, mostly elections, with two central banks, US payrolls and euro-area inflation109 of 1021 of 293 of 10
The same ten, run again after the first changes109 of 1020 of 324 of 10
Central banks, data releases, court and trade rulings, a company result108 of 1026 of 356 of 10
Three of those events, each run twice to measure variation66 of 621 of 244 of 6
One event from each of five industries55 of 517 of 214 of 5
The first weekly replay11 of 14 of 50 of 1

The batches ran from October 17th to October 22nd, 2026, on events settled between July and October 2026. We changed the method between batches, so later batches ran on a newer version. Signposts that no source could settle are left out of the counts.

Where the map missed04

Four maps missed. We ran two of those events again with the changes that followed, and both maps held.

  • In Latvia's election, the United List won 35% of the vote. The sources the method could see put it at 15% to 22%, and the map's outcomes centred between 13% and 22%. Run again after the first changes, which made the worst case reach further, the map held it.
  • Euro-area inflation came in at 3.8%. The first run held it. The second run's outcomes centred between 2.0% and 3.3% and missed. Since then, the worst case has to reach at least as far as the most extreme past case the method finds and checks.
  • US employers added 162,000 jobs in August. The market expected about 56,000, and the map's outcomes centred between a loss of 58,000 and a gain of 60,000. The map now starts from the market's expectation and from how far past figures have surprised it. Run twice that way, it held the result both times.
  • A jury awarded $5.72 billion against Apple in a patent case. The map's outcomes centred between $35 million and $420 million, and the worst case reached $3.1 billion. Jury awards can land far past anything in the record, so a map for one is a guide to how bad it could get, not to where it will land.
What this does not show05

Replays have five limits.

  • Replays look back at events whose answers now exist. The screening keeps later news out as far as we can check, but it cannot prove that nothing slipped through.
  • 42 replays of 25 events is a small sample.
  • Many maps are wide. In the second batch, five of ten maps spread their outcomes across more than half the scale.
  • The same event run twice gives a similar map but different signposts. "Narrowed to the right part" changed between the two runs of two of the three events we repeated.
  • Each event was replayed from one date. A different date could give a different map.
The live record06

Live briefs are checked on the scorecard.

Every signpost on our public briefs is checked against its source as its date passes, and the result goes on the scorecard, misses included. The first signposts of our original record are due from November 30th, 2026. See the scorecard.