Manaz Afzal

Logistics · Freight operations

Which arrivals will run late, ranked for a freight desk two to four hours out

Two scores for flight-linked freight: one the evening before from the timetable, one on the day once the aircraft lands on its previous leg. Tested on three months of US domestic flights the models never saw.

73%of top-10% day-of alerts arrived late; inbound rule 62%
0.80AUC on the day; route history 0.61
40%of delay minutes caused by a late inbound aircraft
0.65AUC a day ahead: useful for booking choices, not promises

Executive summary

Scoring each arrival when its aircraft lands on the previous leg makes 73% of a freight desk’s alerts real, against 62% for the inbound-delay rule desks use today.

73%of top-10% day-of alerts arrived late, against 62% for the inbound-late rule
0.80AUC on the day (95% CI 0.798 to 0.802); route history 0.61
1.77mlater flights in the out-of-time test, October to December 2024

Situation

Freight desks learn a flight is late when the status page changes, often after the truck is waiting or the customer has a delivery window. A list where most flagged flights land on time gets ignored.

Approach

Two LightGBM scores on 7.08 million US domestic flights: a plan score the evening before from the timetable, and a day-of score once the aircraft lands on its previous leg. Trained January to September, tested October to December.

Result

At a list of 10% of flights, 73% of day-of alerts are real and the list catches 45% of all late arrivals. The day-ahead score is weaker (AUC 0.654) and suits booking choices, not customer promises.

The business problem

A freight forwarder’s control tower watches the flights its shipments are booked on. When a flight runs late, a connection can be missed and a customer has to be told, but the team often learns of it only when the aircraft fails to land on time. Forwarders and airlines that combine schedules, previous-leg delays and airport conditions say the value is in hours of warning, which leaves time to rebook or set expectations.

I built two scores: one the evening before from the timetable, and one on the day once the aircraft lands on its previous leg, and measured how many late arrivals each puts in the top of the alert list two to four hours out.

Team
An ops desk of 2 to 5 per shift, plus drivers or ramp crews on rosters
Scale
200 to 3,000 flight-linked shipments a day
Systems
CargoWise or a forwarding system, airline and airport flight status feeds, email or Teams alerts
Constraint
Alerts only matter if few enough to act on; a desk ignores a list where most flagged flights land on time

What was built

Two scores. A plan score runs the evening before from the timetable: route, hour, airline, which leg of the day the aircraft is on, and its planned turnaround. A day-of score reruns as each aircraft lands on its previous leg and adds how late it was. The desk sees a ranked list, sized to what it can act on.

Timetable and tailsRoute, hour, aircraftrotation, turnaroundPlan scoreEvening before:rebook or widen windowPrevious leg landsInbound delayfrom status feedDay-of score2 to 4 hours out:ranked alert listDesk actsWarn customer, movetruck, re-roster crew
Daily workflow. The desk decides what to do with each flagged flight.

Results

0%10%20%30%40%JanMarMayJulSepNov19 July: 58% late in a day(global IT outage)Mid-Januarywinter storms
In 2024, 21% of 6,965,247 flown US domestic flights arrived 15 or more minutes late. Summer and holiday peaks are worse, and single days can be far worse. Data: US Bureau of Transportation Statistics, On-Time Performance, 7,079,061 flights, Jan to Dec 2024

On the day: most late arrivals are already late on the previous leg

Inbound early/on time3,436,346 flights11%Inbound 1–15 min late827,560 flights21%Inbound 16–30 min late333,255 flights43%Inbound 31–60 min late276,485 flights74%Inbound 61–120 min late181,616 flights88%Inbound 120+ min late104,234 flights86%
If the aircraft arrives 31 to 60 minutes late on its previous leg, this flight arrives late 74% of the time; over an hour, 86 to 88%.
Inbound on time or < 30 min lateInbound 30+ min late≤30 min turn98%31–45 min turn97%46–60 min turn92%61–90 min turn79%91–120 min turn53%121–240 min turn41%240+ min turn40%
A short planned turnaround cannot absorb an inbound delay. With a turn of 45 minutes or less, a 30-minute inbound delay makes the next arrival late 97 to 98% of the time; with two hours or more, about 40%.

What the desk gets

Tested on 1,765,279 flights from October to December 2024, after training on January to September. 16.3% of test flights arrived late.

Flights flaggedRoute × hour historyPlan model (day ahead)Inbound-late ruleDay-of model
Top 5%27%46%64%96%
Top 10%26%36%62%73%
Top 20%24%29%40%48%
Top 30%23%26%31%37%

Share of flagged flights that arrived late, at each list size. Higher is better: it is the share of alerts that are worth acting on.

At a list of the top 10% of flights, 73% of day-of alerts are real, against 62% for a rule that flags any flight whose aircraft is 15+ minutes late inbound. That list catches 45% of all late arrivals. At the top 5%, 96% of alerts are real.

MethodWhen it runsAUC (95% CI)PR-AUC
Route × hour historyDay ahead0.612 (0.610–0.615)0.223
Plan model (a day ahead)Day ahead0.654 (0.652–0.656)0.315
Inbound late > 15 min ruleOn the day0.720 (0.718–0.723)0.423
Day-of model (+ inbound delay)On the day, 2 to 4 hours out0.800 (0.798–0.802)0.627

The day ahead is harder

From the timetable alone the plan score reaches AUC 0.654, against 0.612 for each route and hour's own history (gain 0.040 to 0.043, 95% CI). At the top 10% it flags 36% real late arrivals, against 26% from history. It is useful for choosing which bookings get a wider window, not for promising a customer anything.

Leg 1 of the day1,709,401 flights14%Leg 2 of the day1,584,073 flights19%Leg 3 of the day1,358,816 flights22%Leg 4 of the day1,088,857 flights25%Leg 5 of the day664,777 flights28%Leg 6 of the day392,544 flights29%Leg 7 or later of the day166,779 flights31%
Delays stack through the day: 14% of first legs arrive late, 31% of seventh and later legs. A booking on an aircraft’s first rotation of the day is a cheap way to cut risk.
0%10%20%30%05:0008:0011:0014:0017:0020:0023:0005:00: 9%20:00: 30%All flights 21%
By departure hour: 9% late at 05:00, 30% by 19:00 and 20:00.

Why flights run late

Late inbound aircraft40%Airline (crew, maintenance, fuelling)35%Air traffic system and airport19%Severe weather6%Security0%
Late inbound aircraft is the largest reported cause of delay minutes (40%), ahead of the airline's own causes (35%). Weather is reported for 6%, though some weather delay is booked under air traffic.
Signal (day-of model)Share of model gain
Inbound delay on the previous leg60.7%
Planned turnaround (minutes)15.7%
Scheduled departure hour5.2%
Leg of the day4.3%
Origin airport4.2%
Destination airport3.8%
Airline2.4%
Month2.0%

Data and method

Source
US Bureau of Transportation Statistics, Reporting Carrier On-Time Performance, monthly files for 2024 (public domain)
Size
7,079,061 flights, 15 airlines, 348 airports; 96,315 cancelled and 17,499 diverted, excluded from scoring
Split
Out of time: train January to September (1,500,000 of 5,199,968 flights used), settings chosen on August to September, test October to December
Audit
20,152 flights with no aircraft registration (no rotation features); 836 duplicate rows; 73% of flights linked to the aircraft's previous leg

Four methods: each route and departure hour's late rate from training months; a rule that flags a flight when its aircraft is more than 15 minutes late inbound; a LightGBM plan model on timetable fields; and a LightGBM day-of model adding the inbound delay. Confidence intervals from 200 bootstrap redraws of 400,000 test flights.

A random split, which mixes days from the test months into training, puts the plan model at AUC 0.719 instead of 0.654. It learns which days were bad. The figures above use the out-of-time split, which is what a live desk would see.

Assumptions and limitations

  • US domestic flights: the Australian regulator (BITRE) publishes on-time figures monthly by route, not per flight, so an Australian desk needs its own status-feed history to refit
  • The day-of score assumes the inbound delay is known when the aircraft lands on its previous leg; a status feed with a few minutes’ lag is enough
  • Freight moved and what an alert is worth are not in the data; list sizes should be set from the desk’s own cost of a missed truck against a wasted call
  • Cancellations are excluded; they need a separate alert
  • Late is the airline definition, 15+ minutes; a freight cut-off may need a different threshold

What a freight desk would do next

  • Weeks 1 to 2: collect six months of flight status history for the routes in use from the status feed
  • Weeks 3 to 4: refit, agree a list size with the desk lead, connect alerts to the TMS
  • Next two months: run day-of alerts on half the shipments and compare missed truck slots and customer complaints
  • Then: add the plan score to booking rules for the most time-critical freight

Technical depth

Each aircraft is followed through its day by registration: leg number, planned turnaround from the previous scheduled arrival, and the actual delay on that previous leg. The day-of model is trained with the inbound delay as known; the plan model never sees it. Rounds are chosen by early stopping on August and September, then refit on all training months.

Analysis code (Python, LightGBM, pandas)

Financial impact

About A$149k a year in truck and desk time for a 500-shipment desk

The comparison is with the inbound-late rule at the same list size, so the gain is the extra real alerts and the fewer wasted ones, not alerts versus none.

InputValueSource
Flight-linked shipments a day500 within the 200 to 3,000 range in the brief
List size the desk acts on10% of flights agreed with the desk lead
Real alerts at that list: model73%Measured
Real alerts at that list: inbound rule62%Measured
Truck time recovered per real alert acted on45 min driver re-planned instead of waiting
Driver and vehicle cost per hourA$95
Desk time wasted per false alert5 min call or re-plan that was not needed
Desk cost per hourA$50
Operating days a year360

Calculation

  1. Alerts a day50
  2. Extra real alerts a day vs rule5.5
  3. Fewer false alerts a day5.5
  4. Value a dayA$415
  5. Value a yearA$149,325

Annual value against the inbound-late ruleA$149,325

Fewer customer complaints and missed delivery windows are not counted. Cancellations are excluded from the model and need their own alert.

Implementation and rollout

A refit on the desk’s own routes in four weeks, then a two-month split test

  1. Weeks 1 to 2

    Status history

    Six months of flight status for the routes in use from the feed (FlightAware or airline APIs); link each flight to its aircraft’s previous leg.

    Ops systems lead, Manaz

  2. Weeks 3 to 4

    Refit and connect

    Refit both scores, agree the list size with the desk lead, push alerts into the TMS and SMS templates.

    Manaz, desk lead

  3. Months 2 to 3

    Split test

    Day-of alerts on half the shipments; compare missed truck slots, wasted calls and complaints with the other half.

    Ops desk, Manaz on monitoring

  4. Month 4 onward

    Booking rules

    Add the plan score to booking rules for the most time-critical freight; quarterly refit.

    Desk lead, sales

Tools

FlightAware AeroAPI or airline status feeds, CargoWise or the existing TMS, Python, LightGBM and pandas, SMS and email notices through the TMS or Twilio

Team

  • Ops desk lead owns list size and the alert playbook
  • Ops systems lead status feed and TMS access
  • Manaz data pipeline, models, integration and monitoring

The timeline is a plan for a typical engagement; it depends on data access and how many systems are involved.

Risks and mitigations

What could go wrong in production, and the control for each

RiskMitigation
Status feed lag hides the inbound delayScore on actual landing messages; fall back to the plan score when the inbound leg is missing (27% of flights had no link in the US data).
Desk ignores alerts if the list growsFixed list size per shift, set from the desk’s capacity; weekly report of real-alert rate.
Australian routes behave differently from US dataRefit on the desk’s own feed history before go-live; BITRE publishes only monthly route totals, so per-flight history is needed.
Cancellations treated as late arrivalsSeparate cancellation alert from the status feed; excluded from this score.

Key takeaways

Key takeaways

  • Most late arrivals are already late on the previous leg; the inbound delay carries 61% of the day-of model’s gain.
  • The model beats the rule desks use today at every list size, most at small lists: 96% real alerts at the top 5%.
  • A day ahead the timetable alone is weak (AUC 0.654). Use it to choose first-rotation flights, not to promise windows.
  • A random split would have overstated the day-ahead score (0.719 against 0.654); the out-of-time test is what a desk would see.