Applied Machine Learning / Aviation Operations / Systems Design

When Airline Crew Do Not Return: Forecasting the Staffing Risk

By Samuel Getachew14 min read

In this article

A friend told me that a colleague at Ethiopian Airlines did not return with the crew after a trip to Tel Aviv. The account made me think about the operational problem: a flight still needs its required crew, even when someone expected to work is unavailable.

My first instinct was to treat this as a prediction problem. Could employee records reveal who was likely to remain abroad? Could a machine learning model produce a probability that helped the airline assign people to flights?

That question needed a better definition of success. A missed return, an intention to leave, and an asylum application are different events. Predicting one does not establish the others.

The proposal I would make is narrower: forecast unexpected crew unavailability at the staffing-pool level, then plan feasible reserve coverage. It addresses the disruption an airline can manage directly.

This is a technical proposal. I do not have Ethiopian Airlines attendance records, and I have not trained or validated this system. The Tel Aviv account is secondhand; I could not independently verify the incident or the person’s reason for remaining abroad.

What the public evidence establishes

Ethiopian Airlines: documented cases, incomplete operational data

A 2006 judgment from the U.S. Court of Appeals for the Fourth Circuit records an Ethiopian Airlines flight attendant arriving on a working flight in June 2001, remaining in the United States, and applying for asylum. The court sent the case back for further proceedings. This documents an individual case, not a current airline-wide incident rate. [1] (opens in a new tab)

CNN also reported in January 2022 that several Ethiopian Airlines employees said they had left Ethiopia through their work as flight crew and remained abroad. They described fears connected to the Tigray conflict. The investigation separately covered technicians who stowed away on aircraft. Those incidents should not be combined into a cabin crew nonreturn count. [2] (opens in a new tab)

Training service commitments have historical support. Article 22.2(3) of the airline’s tenth collective agreement, dated 2012, describes commitments to serve after specified training. The public documents reviewed here do not establish a current cabin crew arrangement requiring an employed guarantor to repay training debt. I also could not verify a current rule restricting an entire graduating cohort after one person fails to return. [3] (opens in a new tab)

These distinctions change the framing. There is evidence that nonreturns have occurred. There is not enough public evidence here to call this Ethiopian Airlines’ largest workforce problem, calculate its prevalence, or state the reported guarantor and cohort rules as current policy.

Pakistan International Airlines: a documented comparison

On 2 March 2024, a Pakistan International Airlines spokesperson told Arab News that ten crew members had failed to report for return flights from Canada over the preceding two years. He also said that profiling, age restrictions, and retaining passports had not prevented further incidents. This is the airline’s account of its experience, not a controlled evaluation of those measures. [4] (opens in a new tab)

The same reporting does not establish ten confirmed asylum applications. A later case illustrates why that matters. In November 2025, PIA said it was investigating a crew member who missed a Toronto return duty. According to its spokesperson, the employee reported illness when contacted. The published report did not establish an asylum claim or intentional abandonment. [5] (opens in a new tab)

The reports provide incident counts without the corresponding number of crew layovers. They therefore cannot tell us how likely a nonreturn is. They also provide no defensible estimate of the savings a prediction system would generate.

Define the failure the system should address

For crew operations, the immediate failure is an uncovered duty. The question is whether enough eligible people can report at the required place and time.

The proposed system would estimate the number of assignments that may become unexpectedly unavailable within a defined reporting window. It would combine those estimates with known vacancies and the actual reserve roster.

Observed event: a scheduled crew member is unavailable when required to report.
Operational consequence: a required position may need replacement coverage.
Decision: where and when to make qualified reserve capacity available.
Unknown motive: a separate matter that the forecast does not need to infer.

Even where a valid training bond exists, recovering training costs does not supply a replacement for the next departure. Debt recovery and immediate crew coverage solve different operational problems.

Why I would change the original prediction target

The employee-level idea looks straightforward until the target is examined. A model trained on missed duties learns patterns in missed duties. Calling its output an asylum probability would claim knowledge the labels do not contain.

NIST identifies this as a validity problem: an indicator must actually measure the concept it claims to represent. Its AI Risk Management Framework Playbook also warns about proxies for attributes that cannot be directly observed. Applying that principle here, an attendance record is not a validated measure of loyalty or future immigration intent. [6] (opens in a new tab)

Using personality and socioeconomic inputs would introduce another problem. Personal finances, family ties, and inferred beliefs could turn an operational tool into a system for restricting opportunities based on personal circumstances. Removing explicit sensitive fields would not, by itself, remove the possibility of proxy effects. [6] (opens in a new tab)

There is also a substantive boundary. Article 14 of the Universal Declaration of Human Rights recognizes the right to seek asylum from persecution. An airline’s need to cover a duty does not make a presumed future protection claim an appropriate adverse prediction target. [7] (opens in a new tab)

I would therefore use the model to allocate backup capacity. Individual qualification and duty records would still be needed to determine whether a reserve can legally and practically cover an assignment. They would not become a personal disappearance score.

Would an ML and LLM hybrid improve the proposal?

I considered a hybrid in which an LLM reads incident notes and personnel documents, while a numerical model estimates risk. That adds a language-processing component, but it does not resolve the missing evidence about intent.

An LLM-generated interpretation of a personnel note is still an interpretation. It cannot convert an unverified explanation into a reliable training label. NIST’s generative AI guidance identifies confidently presented false content as a specific risk. That risk matters when a generated explanation could affect someone’s work. [8] (opens in a new tab)

The core inputs needed here are structured: assignment counts, reporting times, availability, qualifications, locations, and operating constraints. The outputs are also structured: an uncertainty distribution, a reserve plan, and any remaining coverage gaps.

For this task, I would leave the LLM out of the proposed decision system. A language interface is not necessary to calculate those quantities or enforce exact rules. The selected architecture is probabilistic statistical learning plus constrained reserve scheduling. This is an engineering choice based on the task and evidence, not a claim that one model has already outperformed every alternative on airline data.

The proposed architecture

Component	Input	Output
Attendance aggregation	Roster snapshots and observed duty availability	Scheduled exposure and unexpected unavailability counts
Probabilistic forecast	Aggregate counts and operational context known at the forecast cutoff	Distribution of unavailable assignments by pool and window
Mapping to duties	Count forecast and recorded duty and pairing structure	Possible shortages with timing, qualification, and duration
Constraint engine	Qualifications, duty records, rest rules, locations, and positioning times	Feasible reserve assignments
Reserve optimizer	Forecast uncertainty, known gaps, feasible assignments, and costs	Reserve coverage plan and explicit uncovered duties
Operations review	Plan, uncertainty, and binding constraints	A documented staffing decision

A staffing pool would group genuinely interchangeable qualifications at a relevant location and reporting window. Small groups would need broader aggregation where operationally meaningful. The forecast would not identify people as likely asylum applicants.

1. Build a dataset of observable availability

A long company history does not establish the existence of a usable training dataset. The relevant records must use consistent definitions and preserve what was known before each duty.

Scheduled exposure: how many crew assignments were still expected to be available at the forecast cutoff.
Outcome: how many subsequently became unavailable at the required reporting time.
Operational context: the station, reporting window, required qualification, season, and documented disruptions known at that cutoff.
Existing gaps: confirmed absences and known vacancies, handled directly rather than predicted again.
Linked duties: the affected pairing and absence episode, so downstream effects are not mistaken for independent incidents.

One person’s absence can affect several flights in the same pairing. Treating every affected flight as a separate independent absence would misrepresent the event. Research on probabilistic airline absence and recovery explicitly connects crew absence, pairing structure, and feasible reserve coverage. [9] (opens in a new tab)

Personal messages, inferred personality traits, political views, family finances, and suspected immigration plans would not be inputs. The forecast needs reliable attendance outcomes, not a richer dossier.

2. Forecast a bounded count with uncertainty

My proposed starting model is a hierarchical beta-binomial count model. The name is technical, but the function is specific: estimate how many assignments may become unavailable out of the number scheduled.

Bounded count: the prediction cannot be negative or exceed the scheduled exposure. [10] (opens in a new tab)
Variation between windows: the beta-binomial allows more variation than a binomial model with one fixed absence probability. [10] (opens in a new tab)
Partial pooling: operationally comparable groups can share information, while the model retains uncertainty about each group. This can limit unstable estimates from small samples. [11] (opens in a new tab)

The output would include an expected unavailable count and a predictive interval. The scheduler would use the distribution, since planning only for the average can leave the system exposed to larger shortfalls.

For reserve planning, the forecast must be translated into possible shortages attached to actual reporting times, qualifications, and linked duties. That translation would preserve dependencies supported by the operational records. If those records cannot support it, the output must remain an aggregate capacity estimate rather than a detailed coverage plan.

This statistical structure does not establish that the forecasts will be accurate. Pool definitions, prior assumptions, missing records, and changes in operations all matter. A beta-binomial component also does not automatically model a disruption shared across several stations or consecutive duty windows. Those dependencies must be represented where the evidence supports them, and their remaining uncertainty must stay visible.

3. Convert the forecast into feasible reserve coverage

The reserve scheduler would balance coverage against staffing and positioning costs. Its decisions would include reserve location, duty start time, and which required duties available reserves could cover.

The following would be hard constraints:

Required role, aircraft qualification, and other assignment qualifications.
Actual reserve availability and existing commitments.
Travel time, entry permissions where relevant, and the ability to reach the duty.
Applicable duty limits, rest requirements, and labor rules.
Coverage of the complete required duty, not only its first flight.

ICAO’s fatigue-management provisions cover both flight and cabin crew. The applicable limits depend on the operating rules; a model should not substitute a universal number for those requirements. [12] (opens in a new tab)

If no eligible reserve can arrive and complete the duty, the system must report the gap. Lowering a rest or qualification requirement to produce a complete-looking roster would make the output misleading.

There is relevant research precedent. Schrotenboer and colleagues’ 2023 study used data from a Dutch airline to evaluate reserve scheduling that accounts for expected reserve shortfall. Its research comparisons reported fewer cancellations and last-minute alterations than the benchmark. The paper’s scope and assumptions include a principal focus on one execution day and independent reserve demand across periods. [13] (opens in a new tab)

That study supports attention to reserve flexibility. It does not validate this exact forecast, establish results for Ethiopian Airlines, or show that the system prevents crew nonreturns.

Hospitals face a useful operational parallel. They must cover work with appropriately trained staff despite uncertain attendance. A spare person is useful only if that person can perform the required work.

Ryu and Jiang’s 2025 nurse-staffing study models uncertain demand, absenteeism, and cross-trained float pools. Its numerical case studies use data from a collaborating hospital. The authors report better staffing decisions outside the fitting sample when the model accounts for the relationship between staffing levels and absenteeism. [14] (opens in a new tab)

Planning question	Airline crew	Hospital nursing
What needs coverage?	Required positions by duty, qualification, and station	Required nursing coverage by unit and shift
What supplies flexibility?	Qualified reserve crew	Cross-trained float staff
What limits substitution?	Qualification, positioning, availability, duty, and rest	Unit competence, availability, working time, and recovery
What must remain uncertain?	Future unavailability and disruption	Future attendance and patient demand

The transferable principle is to forecast coverage needs and allocate qualified backup capacity. The hospital study uses a different optimization formulation. It does not validate the proposed airline model or explain why airline employees remain abroad.

What impact could the airline reasonably expect?

The direct opportunity is better preparation for uncovered duties. If the forecasts are useful and qualified reserves exist, the system could improve the placement and timing of that capacity. It could also make the cost of additional protection against shortfalls explicit.

I would evaluate that claim with operational measures:

Uncovered duties: the share of required crew positions still without feasible coverage.
Disruption: delay minutes and cancellations attributable to crew coverage, with attribution checked against operational records.
Schedule stability: last-minute assignment changes and replacement journeys.
Total cost: reserve staffing, positioning, accommodation, and the recorded cost of remaining disruption.
Forecast reliability: whether predicted ranges and shortfall probabilities match later observed outcomes.
Constraint compliance: whether every recommended assignment satisfies the applicable qualification, duty, and rest requirements.

The forecast would need evaluation on later historical periods using only information available at each earlier cutoff. Its reserve plans would need comparison with a documented existing planning rule under the same staffing and cost constraints. A retrospective comparison would remain an estimate under its assumptions, not proof of live operational savings.

No defensible percentage improvement follows from the public reports. The airline’s incident frequency, available reserves, replacement times, and disruption costs are unknown here. A model cannot create a qualified replacement where none exists, and it cannot claim credit for preventing an event merely because the schedule remained covered.

The costs and harms the proposed solution could still introduce

False precision from limited data

Few relevant observations can leave estimates highly dependent on modeling assumptions. Partial pooling can stabilize an estimate while still making it wrong. A narrow-looking probability should not conceal a weak evidence base.

Poor estimates during shared disruptions

A crisis or transport disruption can affect several people and stations together. Independent forecasts may understate that joint risk. Historical coverage also reflects past assignment decisions. If some groups were routinely excluded from routes, the resulting data would contain limited evidence about their availability on those routes.

Reserve costs and workload displacement

More reserve coverage costs money even when nobody is absent. Positioning staff can also move a shortage elsewhere. Frequent callouts may increase workload and reduce recovery opportunities. A plan that improves one departure while weakening the rest of the roster is not an operational improvement.

Aggregate predictions becoming employee surveillance

Aggregation alone does not guarantee privacy. A small station or narrowly defined qualification group may identify particular people. Managers could also turn a staffing forecast into a reason to restrict a cohort’s assignments. The proposed output must remain a capacity-planning tool, with limited access to underlying records and a way to correct inaccurate operational data.

Optimizing the wrong objective

A scheduler follows the costs and constraints it is given. If fatigue, schedule disruption, or employee impacts are omitted, it may recommend a cheap plan that transfers those costs to staff. Human review needs the assumptions and binding constraints, not just a recommended roster.

A narrow solution to a larger human problem

This proposal could help an airline prepare for unexpected unavailability. It cannot determine why a person leaves, resolve a protection claim, or repair the underlying conditions that lead people to leave. If it adds surveillance and reserve expense without improving feasible coverage, it has failed the operational problem it was meant to solve.