Is stock sampling about financial shares?

Not in this duration-analysis meaning.

Does it always include follow-up?

No. Usage varies, so the observation scheme should be stated.

Can a stock sample describe current workload?

Yes. That can be its correct target, even when it does not directly estimate entrant durations.

AI Glossary / Statistics and probability / Intermediate

Stock sampling

Definition

Stock sampling selects units that are already in a particular state at a chosen observation time, such as customers currently subscribed or cases currently open. In duration studies, this differs from selecting units when they enter the state. Longer spells have more opportunity to be present at a snapshot, so a stock sample can overrepresent long durations relative to an entry cohort. The sample may be appropriate for studying the current stock while requiring adjustments for questions about new entrants.

What it means

The central distinction is between who is present now and who begins during a period. A snapshot of open support cases excludes cases that opened and closed before the snapshot. It is therefore not the same as following every newly opened case to completion.

Terminology varies: some accounts use stock sampling for a snapshot of elapsed durations only, while others include follow-up of sampled units. Specify what is observed. Elapsed duration is time already spent in the state; residual duration is time remaining; completed duration combines them. A snapshot's elapsed durations should not be casually relabeled as completed durations.

How it works

Define the state, sampling date, and eligible population. Select from the units present on that date and record how each unit entered the sampling frame. Where duration matters, obtain reliable start dates and, if planned, follow units until exit or the end of observation.

Recognize the selection mechanism. A spell must survive long enough to reach the sampling date to be observed. Short spells ending earlier are absent, not merely recorded with missing endpoints. If follow-up ends before an observed spell finishes, right censoring creates an additional issue.

Choose analysis appropriate to the intended quantity and available information. Describing the current stock is a different task from estimating the duration distribution of entrants. Survival methods may need to account for delayed entry, censoring, and assumptions about the arrival process.

Under a stable arrival process with suitable independence assumptions, snapshot representation is proportional to spell length. That simple relationship illustrates length bias, but changing entry rates or duration patterns can require more elaborate modeling. More snapshot records do not remove a selection mechanism that the analysis ignores.

Three practical examples

1. A support team samples currently open tickets to understand today's workload, while using an entry cohort for estimates of typical completion time.

2. A subscription business surveys current members and recognizes that long-lived subscriptions are more visible than short subscriptions that already ended.

3. A labor researcher distinguishes people unemployed on a survey date from people who newly enter unemployment during a specified period.

Think of it this way

Think of photographing a waiting room at a random moment. People who wait a long time have more opportunities to appear in the photograph. The image accurately shows who was present, but it does not give every arrival an equal chance of being represented.

Worked example

Consider an original idealized service with stable arrivals. Half of entering cases last exactly one day and half last exactly nine days. Assume the system has operated long enough to ignore start-up and boundary effects, and that arrival timing does not depend on duration type.

Among entrants, mean completed duration is 0.5 × 1 + 0.5 × 9 = 5 days. For every equal number of short and long cases entering, the short group contributes one unit of case-time per case and the long group contributes nine.

Consequently, a random snapshot of active cases has expected type proportions proportional to 0.5 × 1 and 0.5 × 9. Normalizing gives 10% short cases and 90% long cases. If their complete durations are subsequently learned, the stock-weighted mean of those complete durations is 0.1 × 1 + 0.9 × 9 = 8.2 days.

The 8.2-day figure differs from the entrant mean of five days because of selection, not arithmetic error. It is not the mean elapsed time already observed at the snapshot. Under this simplified model, weighting completed durations inversely by their inclusion-length factor would recover equal type shares, but real correction requires justified design and duration assumptions; it is not a universal recipe.

Illustrative case study

In a fictional company, Rowan Support reports average case duration from tickets open on Monday. Managers interpret the result as the expected experience of every new customer. An analyst explains that quickly resolved tickets are less likely to appear in the Monday list. The team retains the snapshot report for backlog management and creates a separate cohort report following new tickets, with appropriate treatment of unfinished cases. Both reports remain useful once their populations and time measures are clearly distinguished.

Common misconceptions and mistakes

Stock sampling is not inherently wrong. It can directly answer questions about the current population in a state. The problem arises when its results are interpreted as if they came from a different sampling design.

Length bias and right censoring are not interchangeable. The former concerns which spells enter the sample; the latter concerns incomplete observation of endings.

Elapsed, residual, and completed durations are distinct. A long elapsed duration does not directly reveal the remaining time without assumptions about the process. Nor does a large stock sample automatically represent short spells or new entrants fairly for duration estimation.

Business use

Businesses use stock samples for backlog audits, active-customer research, inventory reviews, and ongoing-case monitoring. These snapshots can accurately characterize operational conditions at a moment in time.

When evaluating onboarding, retention, or resolution time, ask whether the decision concerns current cases or entering cohorts. Record start and exit information consistently and design follow-up before analysis. An AI model trained only on currently active cases may also face a different population when predicting outcomes for new arrivals. Evaluate that distinction instead of assuming one dataset serves both purposes.

Check your understanding

Which is more likely to appear in a random snapshot under stable arrivals: a one-hour case or a ten-hour case?

Answer: The ten-hour case has more exposure time and therefore more opportunity for inclusion. This does not mean ten-hour cases are more common among entrants.

Related terms

Flow sampling
Survival analysis
Cohort
length-biased sampling
delayed entry
left truncation
right censoring
elapsed duration

Learn next

Study entry-cohort designs and survival analysis with delayed entry. Then compare current-stock summaries with estimates for new arrivals.

Sources

Babington and Cano-Urbina, Stock Sampling with Interval-Censored Elapsed Duration, author-university paper: https://myweb.fsu.edu/jcanourbina/BabingtonCanoUrbina2018.pdf . The introduction explains selection, follow-up, and terminology variants. The stable-arrival calculation and fictional support case are original illustrations, not empirical findings.

Previous: Stochastic weight averaging Gaussian approximation Next: Stock series