BRENNEN

Industries / Transport and vehicles / Trucking and freight / Last-mile delivery

Failed deliveries

Failed deliveries are attempts that do not end with the parcel delivered, because nobody is home, the address is wrong or the building cannot be entered, and the redelivery or return they cause.

6

Milestones achieved, from 2010 to 2026.

3

To watch for: announced, not yet confirmed outside.

3

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 2010

    A one-hour window

    The UK parcel carrier DPD launched Predict in February 2010, sending customers a one-hour delivery window by email or text. Its franchise partner Interlink Express said 12% of first attempts had failed before, and first-time success improved by 8.8% after. These are the company's own figures.

    cep-research.com

    Credit

    DPD; Interlink Express

  2. 2015

    Japan counts the cost

    A study group of Japan's transport ministry, MLIT, found in September 2015 that about 20% of parcels needed redelivery. It put the cost at about 180 million hours of work a year, about 90,000 drivers, and about 420,000 tonnes of carbon dioxide a year.

    mlit.go.jp

    Credit

    Ministry of Land, Infrastructure, Transport and Tourism, Japan

  3. 2023

    Routing around likely failures

    Lim, Wang and Webster built routes that take each stop's predicted chance of failure into account, and tested them in a randomised field trial with a carrier in Singapore. Failed attempts fell about 10% and each route saved about $13.

    ideas.repec.org

    Credit

    Lim, Wang and Webster, Production and Operations Management

  4. 2024

    Japan caps driver overtime

    From April 2024 Japan limited truck drivers' overtime to 960 hours a year. A government study group estimated that, without changes, trucking capacity would fall short by 14.2% in 2024 and 34.1% in 2030, which made cutting redelivery urgent.

    wwwtb.mlit.go.jp

    Credit

    Government of Japan

  5. 2025

    UK delivery problems at record

    In Citizens Advice's 2025 survey of 8,000 UK adults, run by the polling firm Opinium, 37% had a problem with their most recent delivery, about 15 million people. The most common problem, at 29%, was a driver leaving before the person reached the door.

    citizensadvice.org.uk

    Credit

    Citizens Advice

  6. 2026

    Japan's redelivery rate falls

    MLIT has counted redeliveries every April and October since 2017. In April 2026 the rate was 7.6% across six carriers, down from 8.4% a year earlier; on the older three-carrier basis it was 8.8%, against 10.4% in April 2024. The government's target of 6% for the 2024 fiscal year was missed.

    LessonA national target, a public count and a shift to lockers and leaving parcels at the door cut redelivery, but slower than planned.

    mlit.go.jpmlit.go.jp

    Credit

    MLIT; Japanese carriers

Watch for

  1. Leaving parcels by default

    Japan's transport minister said in November 2025 that okihai, leaving a parcel at the door without a signature, would become a standard service through revised standard carriage terms. The 2026–2030 policy outline targets about 50% of parcels received without meeting the driver by 2030.

    Achieved whenthe revised standard terms are published and carriers offer okihai as a default option.

    logi-today.commlit.go.jp

    Claimed by

    MLIT, November 2025 and March 2026

  2. More rules for UK carriers

    Ofcom, the UK communications regulator, reported that 71% of people had a delivery issue in the past six months and said it is considering additional regulation of parcel firms.

    Achieved whenOfcom publishes new binding rules for parcel carriers.

    irishnews.com

    Claimed by

    Ofcom, October 2026

  3. Fake failed attempts measured

    A study of drivers at a large Indian delivery firm found some logged a failed attempt without going to the address, and that this cut next-day successful deliveries by 1.60%.

    Achieved whenthe paper passes peer review or the finding is repeated with another carrier's data.

    arxiv.org

    Claimed by

    Arora, Choudhary and Kireyev, September 2026

Open

  1. 01

    Messy addresses test set

    A public set of real-world-style addresses from several countries, each labelled deliverable, ambiguous or undeliverable, with the reason. Score free address parsers such as libpostal, and AI models, on whether they flag the bad ones.

    The brief
    Credit · unclaimed
  2. 02

    Log false "nobody home" notices

    A public log, from volunteers with doorbell cameras, of each tracking update that says a delivery was attempted, matched against the video to show whether a driver came.

    The brief
    Credit · unclaimed
  3. 03

    Predict failure from open data

    A model that predicts the chance a first delivery attempt fails from public features only, such as building type, number of units and street type, published with its test data.

    The brief
    Credit · unclaimed

Open 01 · the brief

Messy addresses test set

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public set of real-world-style addresses from several countries, each labelled deliverable, ambiguous or undeliverable, with the reason. Score free address parsers such as libpostal, and AI models, on whether they flag the bad ones.

Why it is still open

A 2023 dataset of 239,000 US addresses tests address parsing with 21 error types but has no deliverability labels, and the MESSY STREETS set of 2026 tests geocoders, not deliverability. An Indian payments firm reported nearly 30% of orders returning to the seller and built a completeness model whose data is not public.

arxiv.orgarxiv.orgsigir-ecom.github.io

Built with what exists

Public address lists such as OpenAddresses; deliberately damaged copies of real public addresses; libpostal; an open AI model run on one graphics chip (see the Note "One GPU to fine-tune"); human labellers to settle the labels.

  1. First step

    Write the label rules and label 200 addresses from one country by hand.

  2. Done looks like

    At least 5,000 labelled addresses from at least five countries, released openly, with precision and recall for flagging undeliverable addresses for at least three tools.

Open 02 · the brief

Log false "nobody home" notices

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public log, from volunteers with doorbell cameras, of each tracking update that says a delivery was attempted, matched against the video to show whether a driver came.

Why it is still open

Evidence so far is surveys, such as the 29% in Citizens Advice's 2025 survey who said the driver left before they reached the door, and one firm's private data in the Indian study. No public event-level log turned up.

citizensadvice.org.ukarxiv.org

Built with what exists

Volunteers' doorbell cameras and carrier tracking emails; a shared form; a web agent to collect each volunteer's tracking events with their consent (see the Note "Deleting 80% to start over"); a person to review each matched video.

  1. First step

    Run the form with ten households for a month and publish the anonymised log.

  2. Done looks like

    At least 500 logged attempts from at least 100 households over three months, with the share of "attempted" notices that the video confirms, by carrier.

Open 03 · the brief

Predict failure from open data

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A model that predicts the chance a first delivery attempt fails from public features only, such as building type, number of units and street type, published with its test data.

Why it is still open

Published failure models use private carrier data, such as Lim and colleagues' Singapore trial and a study of about 500,000 Canadian deliveries. The public Amazon routes dataset holds delivery outcomes for each package, and no model built on it with public building features turned up.

ideas.repec.orgarxiv.orgamazon.science

Built with what exists

The Amazon Last Mile dataset; OpenStreetMap building data; Python with scikit-learn; a held-out set of routes for testing.

  1. First step

    Count the failed-attempt labels in the Amazon dataset by city and check how many stops can be linked to an OpenStreetMap building.

  2. Done looks like

    A published model, code and test split, with its accuracy compared with a simple baseline that predicts the average failure rate.