Drift, Freshness, Refresh, and Dataset Retirement

Refreshing a table every hour does not make it relevant. The source population, product policy, instrumentation, label process, or relationship between features and outcomes may have changed. Drift is a family of phenomena rather than one metric. Feature distributions can shift while model performance holds steady. Performance can decline without any large shift in the features.

Google Flu Trends settles the point, because nothing about it was stale. It rebuilt its estimates continuously from live search queries. Four researchers examined it in Science on 14 March 2014 and found that GFT had overshot the actual 2011-2012 level by more than 50%. The caption to their graph records the run of failures: “From 21 August 2011 to 1 September 2013, GFT reported overly high flu prevalence 100 out of 108 weeks”. The mean absolute errors they print are 0.486 for GFT, 0.311 for the lagged CDC series and 0.232 for the two combined. The freshest of the three inputs was the worst of the three.

The cause was not influenza. The authors attribute the error to "algorithm dynamics": changes Google itself made to search. That is a source change dressed up as a signal about the world. What the model had partly learned was the calendar. Lazer and his co-authors put it in one line: “In short, the initial version of GFT was part flu detector, part winter detector.”

Kandula and Shaman reappraised the system independently, and report the same pattern: GFT estimates were often overestimates of ILI, with large over-estimates at the 2012/13 peak. Google stopped publishing current estimates on its own site in August 2015. The Flu Trends Team wrote that “Instead of maintaining our own website going forward, we're now going to empower institutions who specialize in infectious disease research to use the data to build their own models”.

Data engineering owns much of the evidence needed to interpret change: source versions, coverage, missingness, schema, categories, label maturity, and historical snapshots. Without that evidence, a team looking at GFT in 2012 could see only that the number was wrong. It could not see that the search engine underneath it had moved. A refresh policy should specify what is appended, corrected, recomputed, relabeled, expired, and communicated. The goal is to adapt in a controlled way, without destroying comparability or rewriting history invisibly.

Calling everything drift can lead to the wrong remedy.

Concept drift has a definition, and it is narrower than everyday use. It “refers to an online supervised learning scenario”, wrote Gama and four co-authors in a 2014 survey in ACM Computing Surveys. It is the relation “between the input data and the target variable” that “changes over time”. A shift in the inputs alone does not meet that definition.

Run Google Flu Trends through this taxonomy and it lands in two boxes, neither of them concept change. It was a source or pipeline change, because Google altered its own search behaviour under the model. It was also expected seasonal variation, because winter terms tracked the calendar rather than the illness. Retraining on fresher queries would have addressed neither. The remedy for a source change is to characterise the source. The remedy for seasonality is to compare against a seasonal reference. Diagnosis comes before the refresh, not after it.

A mature policy combines triggers instead of relying on one clock.

The fourth column has a documented instance. On 2 November 2021 Zillow Group announced it would wind down Zillow Offers, because its home-price forecasting model had failed in a changed market. The Q3 2021 shareholder letter it filed with the SEC states the arithmetic: “We have been unable to accurately forecast future home prices at different times in both directions by much more than we modeled as possible, with Zillow Offers unit economics swinging approximately 1,200 basis points from Q2 to an expected -500 to -700 basis points in Q4 2021.” The same filing records “a $304 million write-down on inventory owned at the end of Q3 as a result of purchasing homes at higher prices than our current estimates of future selling prices”, a further $240-265 million of losses expected in Q4, and “our decision to wind down our Zillow Offers operations, which will unfortunately involve a reduction in our workforce of approximately 25% over the next few quarters”. CEO Rich Barton gave the diagnosis in the accompanying press release: “We've determined the unpredictability in forecasting home prices far exceeds what we anticipated and continuing to scale Zillow Offers would result in too much earnings and balance-sheet volatility”. No data-quality alert fired. Outcome degradation was the trigger, and the answer was not another refresh. It was retirement of the product.

The opposite discipline is to settle the trigger before the first retrain. In October 2023 the US FDA, Health Canada and the UK's MHRA jointly published five guiding principles for predetermined change control plans. A PCCP is a manufacturer-proposed plan: the modifications it intends to make, the protocol for implementing and controlling them, and the assessment of their impacts. The first principle is “Focused and Bounded”. The document also ties itself to existing practice: “These principles draw upon the overarching GMLP guiding principles, in particular principle 10, which states that deployed models are monitored for performance and re-training risks are managed.” FDA's final guidance makes the payoff explicit: “The FDA reviews the PCCP as part of a marketing submission for an AI-enabled device to ensure the continued safety and effectiveness of the device without necessitating additional marketing submissions for implementing each modification described in the PCCP”. Written down in advance, a refresh trigger stops being an operational habit. It becomes an approved protocol.

The policy should state not only when to run, but what historical truth is allowed to change.

For high-risk systems, parts of that policy are already fixed in EU law. Article 72 of Regulation (EU) 2024/1689 requires providers to establish and document a post-market monitoring system that actively and systematically collects and analyses performance data throughout the system's lifetime. Paragraph 3 turns monitoring into a document: “The post-market monitoring system shall be based on a post-market monitoring plan.” That plan forms part of the Annex IV technical documentation, so it is filed where the rest of the system's evidence lives, not in a team wiki. The Commission is obliged to adopt a template for it by 2 February 2026. Step 2 of the process below is therefore a written artefact with a home, and step 5 — coordinating downstream state — is the lifetime obligation the same article imposes.

Cause and affected evidence should determine the response. The source-bug case shows how far the consequences travel.

On 4 October 2020 Public Health England reported that 15,841 positive COVID-19 cases were missing from daily reporting and from contact tracing. They should have been reported between 25 September and 2 October. Source files had exceeded a maximum file size: case information was truncated from an Excel spreadsheet once the row limit was reached. Over 75% of them, 11,968 cases, were due to be reported in the final three days. The agency transferred all outstanding cases to contact tracing by 1am on 3 October, put in a rapid mitigation splitting large files, and began an end-to-end review of all systems.

The episode then became a natural experiment. Fetzer and Graeber's study of it states: “Conservative estimates suggest that the failure of timely contact tracing due to the data glitch is associated with more than 125,000 additional infections and over 1,500 additional COVID-19-related deaths.” Fetzer works at the University of Warwick, Graeber at Harvard Business School. The version of record, in PNAS, reports 63% fewer subsequent infections and 66% fewer subsequent deaths for properly traced cases across the six weeks following the glitch. No drift detector would have caught this. The distribution of the cases that arrived was fine. What was wrong was the ones that did not arrive.

Keep a stable benchmark cohort when possible, so model and dataset changes can be compared against common evidence. Add new forward cohorts to test relevance rather than overwriting the old benchmark. When labels or entities are restated, retain original and revised versions with effective dates. Analysts may need both “as known then” and “best known now” views.

ImageNet shows what documenting an edition change actually costs. Its maintainers announced the plan on image-net.org on 17 September 2019: “So far out of 2,832 synsets within the person subtree we've identified 1,593 unsafe synsets”. Five authors published the full arithmetic at FAT* in 2020. Of the 2,832 categories in the person subtree, 1,593 were annotated as unsafe, offensive or sensitive, leaving 1,239 provisionally safe, of which only 158 were judged visual. The paper gives the consumer-facing number: “The unsafe synsets are associated with 600,040 images in ImageNet. Removing them would leave 577,244 images in the safe synsets of the person subtree of ImageNet.” Roughly half the subtree leaves. That is what "document changes in population, coverage, and split membership" means when a benchmark restates itself: anyone comparing a result across the two editions has to be told which one they measured on. A metric shift without that context is ambiguous.

Retirement is the harder half, because withdrawal is not removal. DukeMTMC and MS-Celeb-1M were retracted in April 2019. Three researchers at Princeton University then went looking for what happened next. They analysed 946 unique papers, drawn from corpora of 1,393 citing DukeMTMC, 1,404 citing MS-Celeb-1M and 7,732 citing LFW. In their 20% samples: “In our sample of papers, we found that DukeMTMC and its derivatives had been used 73 times and MS-Celeb-1M and its derivatives had been used 54 times in 2020.” Scaled to the full corpora that is “hundreds of uses in total”, a year after retraction. The authors name the mechanism “runaway data”: derivatives outlive the original. Exposing.ai documents the same retirement from outside — “In response to this report and an investigation by the Financial Times, Duke University has terminated the Duke MTMC dataset” — and the copies that survived it. A retirement condition without notification, licence clarity and derivative tracking is a decision the field never receives.

Eventually the problem itself may change enough that continuity becomes misleading. Create a new dataset product and a bridge study instead of forcing incompatible eras into one version sequence.