How these forecasts are made, and how they are marked

What Kestrel forecasts

When a company applies for a mining permit, the question everyone around it needs answered is when the regulator will decide. Kestrel forecasts that date.

It does not forecast whether the permit will be granted. In these registers outright refusals are rare, so a refusal call carries almost no information, while the timing varies from weeks to years and is what changes a decision anyone is actually making.

What a forecast is

For every application, a range of dates and a middle estimate.

The range is an 80% credible interval, which means the model puts an eight-in-ten chance on the decision landing between those two dates. It is a claim that can be wrong, and the calibration record is where you find out how often it has been: if the ranges are honest, about eight in every ten decided files should land inside.

A range is not a promise, and a wide range is not a hedge. Where the model is genuinely uncertain the range is wide, and saying so is more useful than a confident date that is wrong.

Where the numbers come from

A survival model fitted on the register’s own history: every application that register has decided, how long each took, and the recorded facts about it. The model learns the shape of the wait and how named factors move it, then applies that to files still pending.

The factors are things the register publishes: how long a file has already been waiting, whether the registry has put it on hold, what the office’s recent throughput has been, and which rules were in force when it was applied for. There is no private information in a forecast, and nothing comes from talking to a regulator or a company.

The two clocks

A file has been waiting since the day it was applied for. A forecast measures from the Friday the file entered our coverage. These are different numbers and the site labels both, because a file pending for four years can still be forecast to be decided next month.

Held files

Held means the registry has stopped the clock on a file. In Ontario exploration permits that is almost always consultation with an Indigenous nation.

Kestrel forecasts the timing consequence of a hold and says nothing else about it. Nothing on this site is a claim about whether a consultation is adequate, how it should resolve, or what any party should do. Held and unheld files are never averaged together, because a paused clock and a running one do not average into anything; each file’s row on the front page says whether its clock is running or stopped.

Permit family timelines

Every permit type in the Bulletin has its own timeline below: the steps an application goes through, and what can happen at each one. They are drawn from the states the forecasting model for that permit type uses, so a step here is a step the model knows about. Where an application moves from one box to another, an arrow shows it.

Ontario exploration permits

1. Posted. A notice for the application appears on the registry. Kestrel forecasts it from the first Friday after it is seen, so a file is forecast during its comment period too.
2. Public comment. The registry takes comments for 30 days, the period the law sets.
3. Waiting for a decision. Once comments close, the registry's clock runs until the ministry decides. This is the wait the model measures. From here a file can be decided, be put on hold, or keep waiting.
4. On hold: consultation. The registry stops the clock, almost always to consult an Indigenous nation. Kestrel forecasts the effect of a hold on timing and says nothing about the consultation itself. In the model a file leaves a hold only by being decided; it does not return to waiting.
5. On hold: temporary. A temporary hold placed by the ministry, or one the company asked for. The clock is also stopped. Some temporary holds never end, and the model allows for that.
6. How a file can end. Approved, meaning the permit is issued; not approved, meaning refused, withdrawn or terminated; or still undecided when the period the model looks ahead over ends. A file on hold can end any of these three ways, and the model does not move a file from one kind of hold to the other.

Which date we forecast

The date on this site is the day a decision becomes public, not the day the ministry made it.

Amended 6 October 2026. Kestrel’s principal ruled that day that the date worth forecasting is the day the ministry issues the permit, because that is the day the applicant can act, and that how long the registry then takes to show the decision is a separate fact to report beside each forecast. The paragraphs below describe the rule the record still runs under: the live Ontario exploration model was fitted on posting dates and is graded on them, and it stays that way until the model is refitted on issue dates at the next quarterly refit. When that happens the change will be dated here and in the scoring rulebook, and the lag between issue and posting will be printed with every forecast.

For an Ontario exploration permit, the fact anyone outside the ministry can actually observe is the day the Environmental Registry posts the decision notice. That is the date Kestrel forecasts and the date it is graded against.

Usually the two are the same. Across the 1,557 decided exploration permits whose notice states the ministry’s own issue date, the register posted a median of one day later, and within a week for nine in ten. But the tail is long and it is not rare enough to ignore: about three in every hundred were posted more than six months after the ministry issued the permit, about two in a hundred more than a year, and the slowest 1,301 days. The backlog clears in bursts rather than steadily.

We grade against the posting date because it is the first day the decision exists for anyone but the ministry, and because nothing visible at the time of a forecast distinguishes a file that is genuinely still pending from one already decided and not yet posted.

The scoring rules

The record is graded by Kestrel’s scoring rulebook. Its rules were written before the results they govern were known, and each later addition carries its date. Where a number on this site and the rulebook disagree, the rulebook governs and the disagreement is a defect to report. In plain words:

What a forecast claims. A timing claim: “there is an 80% chance the decision lands between date X and date Y”, with a 50% range inside it. Beside it, probability claims graded on their own: the chance the file is decided within 13 weeks, and within 26 weeks, of the forecast date; for a file not on hold, the chance it is put on hold before its decision; and, where the model’s look-ahead runs out first, the chance the file is still undecided on the last date the model looks to (“62% chance still undecided on 15 January 2031”). No claim is made about whether a permit will be refused. The calibration record on the front page shows the timing claim; the probability claims are graded in the same record.

The decision date is the day the register posts the decision, not the day the ministry made it; the section above says why. A decision that predates the forecast, or was posted between the Friday the forecast was made for and the moment it was generated, is listed with that reason and counts neither as a hit nor a miss.

When a forecast is graded, and how. Once a week the record reads the decisions the register has posted and grades them against the numbers exactly as published, never a recomputed version. A timing claim is a hit if the posted date falls inside the published 80% range and a miss if not; the 50% range is graded the same way. A probability claim is scored by how far the stated chance sat from what happened (saying 90% and being right costs 0.01, being wrong 0.81). Every file is graded against the first range published for it, however many times it is forecast afterwards: a file about to be decided is always being forecast from a few days out, so scoring the last forecast would measure how close to the end we were standing.

A forecast made during the comment period is an ordinary forecast. A file is forecast from the day it first appears on the register, comment period included, and that forecast is in the record and graded like any other, with no provisional class beside the headline. The model takes the comment period’s close from the notice’s stated closing date.

A forecast is never revised. It is frozen in version control on the day it is published; a new week’s forecast is a new commitment beside the old one. A single forecast is withdrawn only for a defect in how it was produced that can be shown without reference to how it scored, never for scoring badly; the withdrawal is published with its date and reason, its grades go with it, and a corrected forecast is graded from the day it exists. A grade that was the same grading event written twice is struck as a duplicate: the earliest copy stands and no number moves.

Until the record is formally launched, any part of it can be taken out. Amended 8 October 2026. Amended 20 October 2026. The record has not yet been formally launched. Until it is, any part of it that Kestrel no longer stands behind, whether a single forecast, a model version, a whole permit type or one kind of claim within a permit type, can be removed from the record entirely on the decision of Kestrel’s principal, and its grades go with it. These removals are not listed on this site; on the day the record is formally launched, the list of everything removed will be published beside it. A new model can also appear in the Bulletin on the principal’s decision to release it, without first running privately for some weeks; if it does not hold up, it is taken out the same way. Everything the record still holds stays exactly as published, and the day the record is formally launched, this exception ends.

The record was reset on 14 August 2026. Every rule stayed as written; what restarted was the population. A permit type enters only after a written admission review, its record starts the day it is admitted, and nothing graded before the reset is pooled with the numbers here.

One headline per permit type, across model versions. The count on the front page is per permit type, pooled over the model versions that have forecast it. A replacement model’s forecasts join the same record beside its predecessor’s, and nothing already graded is restated.

No accuracy percentage until 50 graded files. On a handful of files a percentage reads as a verdict, and a short record is not one in either direction. Below 50 the tally is published instead. The rule was fixed on 2 September 2026, before any permit type came close to meeting it, and it applies whether the count flatters us or not.

Refusals are not scored on their own until there are 30 of them. Outright refusals are rare in these registers. Until a permit type has 30 posted refusals, a refusal is a decision like any other for the timing claim, dated the day it is posted, and no claim is made about how well refusals are called.

What this record is, and what it is not

It began on 14 August 2026. Before that date Kestrel kept a prediction record that it found could not be trusted, and reset it. Nothing from before that date is pooled with the numbers on this site.

The record is therefore short, and the site says so everywhere rather than dressing it up. One permit type is admitted today.

Retrospective tests against history never appear here and are never pooled with live forecasts. A model that was only ever run backwards over data it was fitted on has proved nothing.

Not advice

Nothing on this site is legal, financial or investment advice. It is a published estimate of when a public body will decide a matter already before it.