Understand the data and predictions

Cat and dog intake episodes with no confidently matched recorded outcome as of the snapshot.

Unit:

stays

Formula or cohort:

Count of companion-animal intake rows whose intake is on or before the snapshot and which have no matched outcome on or before it. A repeat arrival is a new stay.

Window:

As of the snapshot time shown in the data details.

Sample and exclusions:

Wildlife and livestock are excluded. Ambiguous joins are excluded. Records open for more than a year are counted but flagged as possibly stale.

Practical use:

A rough size for the population that may need review.

Limitation:

Not verified adoption availability and not exact kennel occupancy: Austin publishes intake and outcome events, not a roster.

Source:

Derived from the two current ShelterBuddy feeds.

Whole days from this stay's intake to the snapshot, for an inferred open stay.

Unit:

days

Formula or cohort:

Snapshot time minus intake time, both read on Austin's clock, rounded down to whole days.

Window:

Intake to snapshot.

Sample and exclusions:

Open stays only.

Practical use:

Finding long waits.

Limitation:

A record the shelter never closed keeps accruing days; past a year it is flagged as possibly stale rather than trusted.

Source:

Intake `source_date` in the current feed; the snapshot time is this site's.

The middle wait among animals actually adopted in the cohort: about half waited less and half more.

Unit:

days

Formula or cohort:

Median of intake-to-adoption durations over adopted, confidently matched stays.

Window:

Stays that began in the trailing twelve months.

Sample and exclusions:

Adoptions only. Transfers, returns to owner, euthanasia and open stays are not in the sample.

Practical use:

Comparing observed adoption timelines across periods or groups.

Limitation:

It is not every animal's expected wait: animals still waiting and animals who left another way are not counted.

Source:

Current feeds; adoption variants are folded into one bucket.

Eligible arrivals adopted by day 30, divided by arrivals that have had a full 30 days in which to be adopted.

Unit:

share of eligible arrivals

Formula or cohort:

Adopted with duration ≤ 30 days ÷ arrivals with at least 30 days between intake and the snapshot.

Window:

Arrivals in the trailing twelve months, excluding the most recent 30 days.

Sample and exclusions:

Arrivals too recent to have had 30 days are excluded, not counted as failures. Other exits within 30 days count as not adopted.

Practical use:

Historical early-adoption frequency.

Limitation:

A recent change in practice shows up only after 30 days have elapsed.

Source:

Current feeds.

Completed stays lasting more than 60 days, divided by completed stays.

Unit:

share of completed stays

Formula or cohort:

Completed stays with duration > 60 ÷ completed stays in the cohort.

Window:

Stays that began in the trailing twelve months.

Sample and exclusions:

Completed, confidently matched cat and dog stays, whatever the outcome.

Practical use:

How often long stays happened in completed history.

Limitation:

It is not the fraction of animals currently here who are past 60 days, and it is not a prediction.

Source:

Current feeds.

The model's estimate, made from what was known at intake, that the total stay exceeds 60 days.

Unit:

probability

Formula or cohort:

Calibrated output of the long-stay model for the stay's intake-time inputs. Bands: 60% and above, 40%, 25%.

Window:

Answered at intake; not re-asked as the stay continues.

Sample and exclusions:

Cats and dogs with a recorded date of birth. Others show as unavailable.

Practical use:

Ranking which arrivals may deserve earlier attention.

Limitation:

A 70% estimate means roughly 70 of 100 comparable scored cases would pass 60 days if calibration holds. It is not certainty, temperament or an animal's value, and for an animal already past 60 days it is history, not a forecast.

Source:

Published model artifact; version and evaluation on the How it works page.

Estimated probability of adoption by day 30 counting from intake.

Unit:

probability

Formula or cohort:

Calibrated output of the 30-day adoption model for intake-time inputs.

Window:

Day 0 to day 30 from arrival.

Sample and exclusions:

Cats and dogs with a recorded date of birth.

Practical use:

A second view of the same arrival: fast exit rather than long stay.

Limitation:

Not "within the next 30 days" for an animal already here, and not one minus the long-stay probability.

Source:

Published model artifact.

Separate counts of arrival events and adoption events in each month.

Unit:

events per month

Formula or cohort:

Intakes by intake month; adoptions by outcome month.

Window:

Every complete month in the current feeds; the current month is marked partial.

Sample and exclusions:

Cats and dogs, confident joins.

Practical use:

Seasonality and pressure.

Limitation:

The two lines involve different animals. Dividing them is not an adoption rate; the tile uses a cohort rate instead.

Source:

Current feeds.

Median days to adoption among adopted animals in each age group.

Unit:

days

Formula or cohort:

Median over adopted stays, grouped by life stage at intake.

Window:

Stays that began in the trailing twelve months.

Sample and exclusions:

The adopted count is shown beside each median; the completed count is the share's denominator.

Practical use:

Comparing observed timelines by age.

Limitation:

A short median among successful adoptions can coexist with a low adoption share: the two answer different questions.

Source:

Current feeds.

Each recorded ending, divided by completed stays in the window.

Unit:

share of completed stays

Formula or cohort:

Completed stays with that outcome ÷ completed stays.

Window:

Stays that began in the trailing twelve months.

Sample and exclusions:

Adopted folds every adoption variant. Transfer, return to owner, euthanasia and other statuses keep the feed's own wording.

Practical use:

Where stays actually end.

Limitation:

Transfer is not adoption. Open stays have no outcome and are not included.

Source:

Current outcomes feed.

The number of eligible stays or adoptions behind that specific figure.

Unit:

stays

Formula or cohort:

Count of rows contributing to the calculation, after exclusions.

Window:

Same as the figure it accompanies.

Sample and exclusions:

Stated as "based on … stays" or "… adoptions" where the distinction matters.

Practical use:

Judging how much to trust a figure.

Limitation:

A small sample can move sharply when a few records change.

Source:

Computed alongside each figure.

How often the model ranks a randomly chosen positive case above a randomly chosen negative one.

Unit:

0.5 to 1

Formula or cohort:

Area under the ROC curve on the held-out test block.

Window:

Held-out test arrivals, dates on the How it works page.

Sample and exclusions:

Stays with an observable outcome.

Practical use:

Ranking quality.

Limitation:

0.694 does not mean 69.4% of predictions are correct.

Source:

Published model artifact.

Summarises precision against recall as the selection threshold changes.

Unit:

0 to 1

Formula or cohort:

Average precision (scikit-learn), not the trapezoidal precision–recall area. Compare with the event rate: a model with no skill scores about the base rate.

Window:

Held-out test block.

Sample and exclusions:

Same as ROC AUC.

Practical use:

How well the top of the ranking concentrates true cases when the event is rare.

Limitation:

False alarms and missed cases trade off; this is one summary of that trade.

Source:

Published model artifact.

The average gap between predicted and observed rates across probability groups, weighted by group size.

Unit:

probability points

Formula or cohort:

Ten equal-width bins; per-bin |observed − predicted| weighted by bin share.

Window:

Held-out test block, after calibration.

Sample and exclusions:

Same as ROC AUC.

Practical use:

Whether the printed percentages mean what they say.

Limitation:

0.047 is about 4.7 percentage points on that evaluation, not a ±4.7-point interval on any one animal.

Source:

Published model artifact.

A straight line fitted to observed rate against predicted rate across ten deciles; ideal is slope 1, intercept 0.

Unit:

unitless / probability points

Formula or cohort:

observed = intercept + slope × predicted, fitted by least squares on decile means. A slope below 1 means predictions are more spread than outcomes; a positive intercept means observed rates sit above predictions at the low end.

Window:

Held-out test block.

Sample and exclusions:

Same as ROC AUC.

Practical use:

Naming the shape of a calibration problem.

Limitation:

Neither value alone proves probabilities are trustworthy.

Source:

Published model artifact; the fit is on the untransformed probability scale.

Average squared difference between predicted probabilities and binary outcomes.

Unit:

0 (best) to 1

Formula or cohort:

mean((p − y)²).

Window:

Held-out test block.

Sample and exclusions:

Same as ROC AUC.

Practical use:

Overall probability quality; lower is better in a comparable population.

Limitation:

Reflects both ranking and calibration, and depends on the base rate.

Source:

Published model artifact.

Probability error that penalises confident wrong predictions strongly.

Unit:

nats, lower is better

Formula or cohort:

−mean(y·ln p + (1−y)·ln(1−p)).

Window:

Held-out test block.

Sample and exclusions:

Same as ROC AUC.

Practical use:

Comparing models on the same evaluation problem.

Limitation:

Not a percentage or an adoption rate.

Source:

Published model artifact.

Of the 50 highest-ranked evaluation cases, the share that actually had the outcome.

Unit:

share

Formula or cohort:

Positives among the top 50 by predicted probability ÷ 50.

Window:

Held-out test block.

Sample and exclusions:

The 50 highest scores in the block.

Practical use:

What a short worked list would have contained, historically.

Limitation:

Retrospective evaluation, not demonstrated outreach success on today's queue.

Source:

Published model artifact.

The top-50 outcome rate divided by the evaluation base rate.

Unit:

multiple

Formula or cohort:

Precision in the top 50 ÷ overall positive rate in the block.

Window:

Held-out test block.

Sample and exclusions:

Same as precision in the top 50.

Practical use:

How concentrated the outcome is at the top of the ranking.

Limitation:

A lift of 5× describes concentration of the outcome. It does not mean 5× more adoptions caused by the tool, and today's queue is not guaranteed the same figure.

Source:

Published model artifact.

Mean predicted rate minus observed rate within a group, in percentage points.

Unit:

percentage points

Formula or cohort:

mean(p) − mean(y) for the group's test rows.

Window:

Held-out test block.

Sample and exclusions:

Group sizes are shown beside each gap; small groups move easily.

Practical use:

Whether the model over- or under-states risk for a kind of animal.

Limitation:

Uncertainty is not estimated for small groups, so a small gap is not proof of fairness.

Source:

Published model artifact.

How much the model uses an input across the evaluated population.

Unit:

share of shown gain

Formula or cohort:

LightGBM gain importance, normalised over the inputs shown.

Window:

Training data.

Sample and exclusions:

Whole training population.

Practical use:

Which inputs the model leans on.

Limitation:

Not causation, not direction of effect, and not an explanation of any single animal.

Source:

Published model artifact.

The share estimated to have been adopted by a given day, accounting for competing exits.

Unit:

share

Formula or cohort:

Aalen–Johansen cumulative incidence with transfer, return and euthanasia as competing events.

Window:

Archive study, October 2013 to March 2022.

Sample and exclusions:

Legacy archive stays under the previous shelter system.

Practical use:

Historical context for how quickly groups were adopted.

Limitation:

Archived values describe that period; they are not current rates.

Source:

Earlier archive study.

The relative instantaneous adoption rate under a fitted survival model.

Unit:

multiple

Formula or cohort:

Cox proportional-hazards coefficient, exponentiated.

Window:

Archive study.

Sample and exclusions:

At-risk archive stays.

Practical use:

Historical association between a characteristic and adoption speed.

Limitation:

Not an adoption probability and not proof that an action changes outcomes.

Source:

Earlier archive study.

The sum of assumed task times for the actions in a plan.

Unit:

minutes

Formula or cohort:

Σ assumed minutes per assigned action.

Window:

This plan.

Sample and exclusions:

The rows in the plan.

Practical use:

Rough workload planning.

Limitation:

Coordination time, not the time to place an animal or complete care.

Source:

Assumptions in the intervention catalogue.

A hypothetical count from explicit assumptions about benefit and eligible animals.

Unit:

animals

Formula or cohort:

Σ over eligible animals of min(assumed percentage-point increase, 100% − baseline), where baseline is the intake-time 30-day estimate if the window is still open and zero otherwise.

Window:

The 30-day horizon from intake, for animals still inside it; hypothetical for others.

Sample and exclusions:

Animals in the plan.

Practical use:

Sizing an assumption, not measuring an outcome.

Limitation:

Never achieved impact or a validated forecast.

Source:

Planner assumptions set by the user.