4.5Mdelivery records in LaDe-D
Unpack the
estimate.
Data. Models. Cleaning.
See what’s
inside.
Data. Models. Cleaning.
Start with
the data.
Every LaDe-D delivery record, across five cities.
View the dataset on Hugging FaceClean
inputs.
Bad records are checked and repaired before training.
Order first.
Then time.
Fixed models. Fresh context.
Hold the
results.
Simulated bad data, before and after the controls.
Data and evaluation
Study design
LaDe-D delivery records from Yantai are used to predict remaining time for pending parcels at each recorded delivery checkpoint.
Training uses events and outcomes before September 1. September 1–15 selects the model; September 16–30 evaluates it.
Evaluation limits
The evaluation includes 155,296 queries across 14,234 parcels. September informed development; October was not tested. Repeated queries are dependent, and no repeat-run uncertainty is available.
Event times approximate data availability. Actual ingestion delays are unknown; recorded completion is not independently verified receipt.
Models
Order and ETA
A LightGBM model predicts stop order. Its output feeds the standard ETA model alongside current context and historical patterns. A separate delay-aware ETA model handles stale feeds.
Training uses out-of-fold order predictions and excludes the current courier-day from transition statistics. Model weights stay fixed during replay.
Input limits
Pending work may be incomplete. The last completed stop approximates courier location; straight-line distance is not road travel time.
The line fan illustrates boosting, not fitted trees or individual predictions. Stop-order prediction is not road-route optimization.
Cleaning and monitoring
Record cleaning
The loader rejects missing IDs, all copies of repeated IDs, invalid timestamps and negative durations. Zero-duration and long deliveries are kept.
The guard removes re-sent copies, repairs large GPS offsets using historical area medians, and excludes bulk-scanned courier-days from training. Repairs are estimates, not verified locations.
Prediction and monitoring
During prediction, bulk scans are retained and flagged. Feed age or upload lag can select a delay-aware ETA model. Data-contract and drift checks report input changes; a sampled comparison gates model replacement.
In a separate simulation, the full set of controls reduced MAE under combined faults from 84.0 to 66.9 minutes. This does not isolate cleaning or establish live reliability. Missing events remain unknown.
Results and sources
Comparable results
The saved matched rerun scores M2G4RTP at 77.11 minutes MAE and Hermes at 58.64 on the same 149,263 predictions. Only M2G4RTP has this matched comparison.
Published LaDe scores (Yantai MAE: LightGBM 23.32, FDNET 22.41, RankETPA 21.36, M2G4RTP 17.21) use different inputs, populations and evaluation, so they cannot be ranked against Hermes.
Current pipeline
The current cleaning-enabled bundle scores 58.48 minutes MAE on 155,296 development queries. This is a separate run, not a replacement for the matched score.
For queries with 3–24 hours remaining, its MAE is 107.06 minutes. These are development results, not deployment evidence. Diagrams and photographs are illustrative.
Dataset card: Apache-2.0. Paper: CC BY-NC 4.0. Confirm applicable terms before redistribution. No raw records or trained models are included. AI assisted coding and presentation; predictive models ran locally.