A pickup time the customer can plan the day around
An ETA step that runs when the driver accepts the job, tested on 338,524 pickups over 56 later days.
Executive summary
An estimate made when the driver accepts the job is off by 46 minutes on average, against 77 for the middle of the customer’s window, and lands within 30 minutes 44% of the time.
Situation
Customers are given a half-day window and ring the office to ask when the driver will arrive; dispatch answers from memory.
Approach
A gradient-boosted ETA on what is known at acceptance, trained on 755,195 pickups and tested on 338,524 over 56 later days, with a calibrated 80% range for dispatch.
Result
Average error falls from 77 to 46 minutes. The 80% range holds 79% of the time, so dispatch is alerted when a pickup slips outside it.
The problem
A parcel or courier network that gives the customer a pickup window of a whole day gets failed pickups and calls asking where the driver is. Carriers that send a narrower window, updated as the run progresses, report more first-attempt success and fewer of those calls. A window helps only when the driver arrives inside it.
I measured how accurately a pickup time estimated when the driver accepts the job lands against the actual arrival, and how narrow a window can be while still holding most of the time.
- Team
- Dispatch, customer service, operations manager
- Systems
- Descartes, Bringg, Onfleet or an in-house dispatch system; Salesforce or Zendesk for customer contacts; SMS through MessageMedia or Twilio
- Measure
- Pickups within 30 minutes of the time given; where-is-my-driver contacts per hundred jobs
What I built
An ETA step that runs when a driver accepts a pickup. It reads the customer’s requested window, the time of day, the driver’s distance, how many pickups the driver already has open, and how long pickups usually take for that driver and in that area. It returns a single time for the customer message and an 80% range for dispatch.
Results
Over 56 test days and 338,524 pickups, the time from the model is off by 46 min on average. The middle of the customer’s requested window is off by 77 min, and each driver’s usual time by 82 min. 44% of pickups happen within 30 minutes of the model’s time, against 35% for the window midpoint.
The 80% range contains the actual pickup 78.7% of the time, with a median width of 120 minutes. That is narrow enough for dispatch to spot a pickup going wrong, and wide enough that I would not send it to customers. The customer message carries a single time, updated if the range slips.
What drives the estimate
| Open pickups at acceptance | Test pickups | ETA model | Courier’s usual time |
|---|---|---|---|
| No other open pickups | 34,967 | 51 min | 78 min |
| 1 to 2 | 50,657 | 44 min | 68 min |
| 3 to 5 | 73,536 | 43 min | 68 min |
| 6 or more | 179,364 | 46 min | 93 min |
Data and method
- Data
- LaDe last-mile dataset (Cainiao; Wu et al., KDD 2023; Apache-2.0): 1,146,781 pickup records from Yantai, China, with acceptance time, the customer’s requested window, courier and area identifiers and the actual pickup time. 1,093,719 kept after removing records with missing times or pickups more than 12 hours after acceptance. No customer details are used
- Split
- By day. The first 70% of 184 days train (755,195 pickups), the last 56 days test (338,524)
- Model
- Gradient-boosted trees (scikit-learn HistGradientBoosting) on log minutes to pickup, fitted on 400,000 training pickups
- Range
- Split-conformal 80% range from the last 15% of training days (106,814 pickups), set by predicted-time band
- Baselines
- Middle of the requested window; each courier’s median time in training
- Inputs
- Only what is known at acceptance: window, hour, distance, open pickups, courier and area medians from training, area type and region
Limitations
- The data comes from one Chinese city with dense parcel traffic. An Australian courier’s routes, traffic and customer windows differ, so the model is retrained on the business’s own job history.
- Pickups were almost never later than the customer’s window (0.25% in the test period), so the problem here is precision within the window, not lateness.
- Driver distance is missing for some records; the model handles the gap, but GPS on acceptance improves it.
- The estimate is made once at acceptance. Updating it from live driver location narrows the error further.
How I would take it live
- Export three months of jobs with acceptance, window and completion times from the dispatch system
- Retrain on the business’s own drivers and suburbs and agree the message wording with customer service
- Run the estimate silently for two weeks and compare it with actual pickups
- Then send it in the customer SMS and alert dispatch when a pickup runs outside its range