The Problem

Introduction

“The chain is only as strong as its weakest link” is an old idiom that has proven applicable across a multitude of circumstances.

It is now entirely appropriate to apply it to Artificial Intelligence (AI), Large Language Models (LLMs) and Natural Language Processors (NLPs), and the data that they are trained upon and use.

Much has been written and spoken about the need to examine AI training data:

• for outliers and anomalies;
• to define and refine algorithms to reduce aberrations and hallucinations;
• critically examine query/response patterns for incorporated biases;
• ensure that the neural training and weighing results in rigorous pruning of defunct and adjunctive pathways; and
• that reverse and recursive feedback is appropriate.

These are all valid elements to be considered.

These elements tend to strengthen the operational and functional behaviours and heuristics of the AI tools, while enhancing trust across the AI landscape in the results that they provide.

The first generations of data sets used to train LLMs and NLPs and the neural networks that underpinned them, acted essentially as expansive data hoovers. They ingested vast quantities of data based on the available information feeds that could be sourced and were represented as informative based on the number of records that they could acquire. The definition of the number of variables across each data line acquired was impressive – tens, hundreds and sometimes thousands of attributes were collected on a myriad of contextual areas and feed into the data maul of the training network.

AI systems inherit the strengths and weaknesses of the data they learn from. Early generations of LLMs were optimising for breadth, correlation and parameter scaling. While this raised performance, it also exposed fragilities—hallucinations, outdated answers and brittle behaviour—especially when facts or relationships change over time.

Deep Data adds the temporal dimension explicitly: when something happened, how fast it changed, and how those changes relate to outcomes. This white paper translates those ideas into practical guidance for organisations deploying AI in time‑sensitive domains, and where the rate of change of data and relationships are crucial for inference generation.

The Evolution

From Big Data to Deep Data

Deep Data augments traditional datasets with temporal context. In many cases the introduction of temporal elements considerably adds to the data volumes that are required to ensure completeness of the subject data domain.

These key temporal concepts include:

• timestamps and acquisition provenance;
• temporal modelling (e.g., Slowly Changing Data [SCD], temporal tables);
• temporal skew (rate of change as a signal);
• time‑varying attributes;
• temporal granularity; and
• temporal analysis techniques (time‑series, forecasting, event‑sequence analysis).

These ideas underpin models that must reason about what was true and when—and what is changing now.

A few key aspects of understanding the temporal dimension include:

Temporal data component: This is any data that includes some form of timestamp, indicating the specific point or period in time when the data record was recorded or was valid.

Examples include user activity logs/timestamps, system activity metrics, event occurrence recording, or date-time marked sensor readings.

Temporal modelling: This ensures that data structures and databases are designed to effectively store and manage temporal data. This can be as simple as s SQL ’Datetime’ attribute on a row or as complex as a separate custom ‘temporal object’. It may be an absolute reference from a known point, a series of variables based on timezone and locality, associated with a state of another object/record, or simply a capture of ‘wall clock time’.

This often involves using specific data types for dates, timestamps, and intervals, and employing techniques like Slowly Changing Dimensions (SCDs) or temporal tables in databases to maintain historical records.

Temporal skew: The “rate of change” of associated variables can be of greater importance than the specific values of the dependent variables themselves. Facilitating the recording of how frequently or rapidly a time record changes may itself provide valuable data to interpret the dependent attributes within the record.

A known repetitive timed ‘snapshot’ of a series of attributes from a designated point can track variants and changes without having to interrogate the timestamp of every record.

Time-varying dependent attributes: Defining data attributes that change over time. The temporal dimension allows tracking these changes, with the rate of change consisting of known defined values or mapped/matched to unplanned event driven situations.

Temporal granularity: Data can be recorded at different levels of time precision, from milliseconds to years. Choosing the appropriate granularity is vital for analysis and depends on the specific use case.

Temporal analysis: This involves examining data over time to identify trends, seasonality, anomalies, and relationships between events.

Techniques include time series analysis, forecasting, and event sequence analysis.

Where current LLMs struggle with time

Peer‑reviewed studies have begun to examine where LLMs have been shown to mishandle time‑limited facts, and struggle with event ordering and temporal entailment. These results have been observed when exposed to post-training data set and have measurable performance and accuracy degradation.

Time‑conditioning (by incorporating explicit dates associated with the records) and timestamping help, but continual updates risk catastrophic forgetting. Retrieval‑augmented approaches that consult up‑to‑date sources reduce some errors but are not a universal cure.

Common failure modes and proven mitigations:

• Distribution shift across time → use time‑split evaluation and drift monitoring
• Catastrophic or spurious forgetting during updates → apply rehearsal/replay or regularisation
• Temporal factuality errors → combine date‑conditioning with timestamped retrieval (Retrieval-Augmented Generation or RAG)

Business value across domains

In a number of industries where the occurrence of series of data is critical to understanding patterns, inference and making predictions the existing AI models have been subjected to queries that are exposing systemic weaknesses.

Studies show LLMs often fail on facts that are only true within specific time windows and can hallucinate or give outdated answers if not trained or updated with time-aware methods.

New benchmarks (e.g., MenatQA[1] and several recent temporal-reasoning datasets/benchmarks) show LLMs struggle with event ordering, temporal entailment, and timestamp-sensitive QA; synthetic temporal benchmarks are also being used to control for data contamination.

AI functional development in biomedical NLP and legal text classification demonstrate measurable performance degradation when models are applied to data from times later than the training corpus, stressing the need for time-aware updating strategies in high-stakes domains. Currency of data for updated training, especially where conditions are changing that influence the relationships between attributes, is essential to ensure that the reflected models are aware of altered data conditions.

When you continually update models with new temporal data, they can forget older (still-relevant) knowledge or degrade due to distributional shifts. The research literature describes methods (rehearsal, regularisation, incremental curricula) to mitigate this behaviour, but no perfect solution has been published yet. This disregard for older data and altered distributional association has been labelled ‘catastrophic forgetting.’

Several recent works report that jointly modelling text with explicit timestamps (conditioning the model on a date) improves performance on time-sensitive queries and reduces temporal errors. This is a step in reducing catastrophic forgetting and hallucinations, but is not a panacea for all the problems that temporally sensitive data encounter.

Separately, work shows LLMs can be used for forecasting and time-series applications by encoding numeric/time-series data as text or prompts. Once again, these studies are promising, but distinct from the problems of temporal factuality and dataset timestamping.

Adding the time dimension improves signal‑to‑noise in forecasting and decision support. Allowing for co‑dependent trend analysis in Business Intelligence (BI) reporting and forecasting, and better market and risk dynamics in the finance sector.

Financial management

From tracking stock market fluctuations and the related events that influence buy/sell behaviours, to predicting financial trends based on market conditions (and their changes). Understanding the patterns behind risk evaluations and managing the risks and residual elements resulting from a treatment plan. Using scaled predictions to facilitate financial hedging in areas such as aviation fuel purchasing, futures markets, commodity markets and energy markets.

Research

Applying temporal attribute analysis facilities across primary research activities is emerging in a number of disparate areas. Studying environmental changes, including but limited to:

• climate change;
• population and demographic shifts;
• agricultural production and food diversity;
• water security and use;
• changes in consumer purchasing criteria;
• geophysical/geological changes and prediction of events; and
• other time-dependent phenomena,

is becoming increasingly prominent.

Operational systems

Realtime tracking systems that manage performance and operation allow for greater predictability and treatment of adverse events. Identifying bottlenecks, and predicting failures in operational systems where a range of time windows can be assessed simultaneously is critical.

From high performance systems such as F1 engine management and ‘on track’ control, to monitoring points containing thousands of sensors managing moisture, wind, sunshine and soil nutrient levels for highly valuable crops.

Healthcare

In medical and clinical settings using temporally aware AI to monitor patient vital signs, tracking disease progression, and analysing effectiveness of treatment outcomes and any side-effects.

In medical research to investigate new treatment and therapeutic scenarios. In medication development using protein folding and recombination visualisations of time dependent drug creation to understand the mechanisms underpinning the medication.

Effectively incorporating the temporal dimension enables deeper insights, more accurate predictions, and better decision-making based on the dynamic nature of data.

An example – Medical research and public health

A vitally important area for the use of Deep Data and the techniques surrounding it are in the medical research and public health policy areas.

The recent spate of measles outbreaks in Western societies reflects a changed relationship between the disease and parents responsible for childhood vaccination. The reduction in vaccination rates in Organisation for Economic Co-operation and Development (OECD) countries reflects a suspicion of vaccines and a misinformation rhetoric that links vaccines to specific syndromes and conditions such as autism. Leaving aside the ethical and moral debate of the merits of vaccination, this new set of temporally related data allows for strong causation links to be derived from emerging data sets.

Traditional data that links vaccination to diminished instances of measles is a great example of strong positive correlation, but not causation. However, the emerging circumstances of lower vaccination rates and increasing disease outbreak provides stronger evidence for a reflective causation relationship. Of course, there are many other factors that contribute to the increase in measles proliferation, but the use of temporally linked data for monitoring and research provides a robust platform upon which to build effective public health and education initiatives.

Research programs such as the Murdoch Children’s Research Institute’s (MCRI) Generation Victoria (GenV) is gathering deep data from a state-wide cohort of children born between October 2021 and October 2023. The longitudinal data set collects situational, environmental, genomic and phenotype data about the children (and some parental data) from immediately prior to birth until the children are eighteen (18) years old.

Much of the data collection is non-invasive and utilises several unique and novel data collection facilities that supplies episodic data whilst also benefiting the parents. A range of bio-samples collected from birth are maintained in state-of-the-art bio-freezes for analysis as data insights are uncovered.

The GenV research is focused on 6 principal areas: Mental health & Wellbeing, Obesity & Diabetes, Allergy, immunity & infection, Development & learning, Organ health, and Healthy environments. However, the dataset that is being built will allow for other researchers to investigate a range of different topics based on the multiplicity of the data elements being collected across time.

The GenV data asset is provisioning researchers within Australia and around the world with a rich, temporally validated set of longitudinal data attributes – linked to physical samples – that is enabling valuable insights into childhood diseases and conditions. The development of appropriate clinical treatments, education of parents and medical staff on likely environmental and health related precursors, and effective diagnostic regimes are just some of the benefits from the Institute’s work.

These are some of the distinct advantages of rich Deep Data sets.

There are a number of other medical research and data collecting programs around the world that are being leveraged for their breadth and depth of data.

These specific programs facilitate data sharing and co-operation though collaborative research platforms. These include the United Kingdom’s Biobank, The U.S. National Institute of Health’s All of Us Program, the Michael J. Fox Foundation, Global Health Data Exchange (GHDx) among a growing list of deep data set providers. There are other data sets that are held by corporations that do not facilitate sharing as they are based on for-profit research, these are typically multi-national pharmaceutical companies.

Interestingly the World Health Organisation (WHO) has been actively encouraging cross-border medical research for many years. The advent of Deep Data sets that collect longitudinal, temporally-rich data can critically advantage the eight-hundred (800) odd WHO collaborating centres[2].

These centres “…are institutions such as research institutes, parts of universities or academies, which are designated by the Director-General to carry out activities in support of the Organisation’s programmes. Currently there are over 800 WHO collaborating centres in over 80 Member States working with WHO on areas such as nursing, occupational health, communicable diseases, nutrition, mental health, chronic diseases and health technologies.” [WHO https://www.who.int/about/collaboration/collaborating-centres ]

Longitudinal data across cohorts in medical research settings are allowing for real‑time monitoring in Operational Systems and better patient trajectories and treatment effects in Healthcare.

In each case the question is not only what correlates, but when correlations hold, how quickly they drift, and what cadence of refresh preserves utility.

The Solution and How

Designing for temporality: A practical playbook

So, now that we have examined the basis of temporal data within the AI/LLM/NLP domain, the question is:

How do we effectively incorporate this into our planning for building the data sets that are required?

The following are some of the methods that are recommended to be considered:

Practical implications for dataset creators & model trainers

Always include timestamps / provenance metadata: for training samples (and keep the original collection dates). Timestamping enables time-conditioning, temporal filtering and contamination checks.

Evaluate models on time-split test sets: train on older data, and test on newer data to measure temporal generalisation. Use temporal benchmarks where available.

Use continual-learning strategies: rehearsal buffers, selective replay, regularisers) when updating models to avoid catastrophic forgetting; monitor for “spurious forgetting” by rerunning the same prompts as the data is acquired.

Combine retrieval/up-to-date knowledge sources: use Retrieval-Augmented generation (RAG). RAG conditions models based on curated external data that has been specifically designed for the target domain. Rather than rely only on static pretraining snapshots for time-sensitive facts. Many empirical studies are resulting in the static/RAG hybrid approach.

Domain-specific refresh cadence: high-change domains (news, biomedical, law) require more frequent updates and stricter temporal validation. Consider applying a continuous feed model of validated (i.e. trusted source) data into the AI model’s matrix.

Open problems & active research directions

By no means is the area of temporal data definitions in AI model complete. There are several active research projects looking at optimising the ways in which longitudinal data sets can be efficiently and effectively used.

The AI vendors are continuously refining their existing capabilities and looking to incorporate new features into their respective offerings. Improving the way in which temporal data attributes are used within the AI/NLP products is an active endeavour for all of the vendors for which publicly available information exists. It can only be surmised that the more private companies are also engaged in similar developments.

It can therefore be expected that the resulting products will begin to feature capabilities that are aligned with understanding and exploiting temporal data attributes. The feature race that is likely to eventuate will be interesting to observe.

Notwithstanding the improvements within the products themselves to record, analyse and report on temporal data elements, there are a number of aspects of preparing and using the data sets that will remain with the users of the products.

To operationalise Deep Data, treat time as a first‑class feature across the lifecycle:

1. Always capture timestamps and provenance for training and evaluation data; retain original collection dates.
2. Adopt time‑split evaluation: train on past, test on future; benchmark with temporal datasets where available.
3. Use continual‑learning strategies (replay/rehearsal, selective refresh, regularisers) to limit forgetting.
4. Combine models with timestamped retrieval (RAG) for time‑sensitive facts and policies.
5. Establish domain‑specific refresh cadences for high‑change areas (news, biomedical, legal).
6. Instrument drift detection and re‑run a stable probe set to catch spurious forgetting as data evolves.
7. Govern temporal bias: audit for over‑ or under‑representation of periods and demographics.

Emerging methods: Time‑conditioning, RAG hybrids and KANs

Conditioning models on explicit dates improves temporal factuality; pairing with retrieval provides a current knowledge substrate without discarding history.

Separately, Kolmogorov‑Arnold Networks (KANs/MFKANs) offer interpretability and efficiency for scientific deep‑data settings by learning functions on edges rather than fixed node activations—promising for complex physical or biomedical relationships, though current training speed is a limitation. In combination with Multi-Fidelity KANs using RAG data has shown significant improvements in models that were prone to catastrophic forgetting and hallucinations.

Implementation checklist – Used by DB Results in our AI and temporal models

• Data: ensure every record has collection and validity timestamps; define granularity and retention.
• Architecture: support temporal tables/SCDs and lineage; separate fresh from historical stores.
• Machine Learning: maintain a replay buffer; schedule time‑split evaluations; log prompts/answers with dates.
• RAG: index sources with timestamps; prefer time‑bounded retrieval; surface citation dates in outputs.
• Ops: monitor drift and freshness SLAs; automate refresh pipelines by domain cadence.
• Governance: document temporal assumptions; assess temporal bias and fairness impacts.

Conclusion

Time is the missing variable in many underperforming AI deployments.

By elevating temporal data to a first‑class concern—from collection and storage to training, evaluation and retrieval—you can reduce brittleness, improve factuality and sustain model value as conditions change.

DB Results partners with organisations to design these temporal foundations, modernise data platforms and implement responsible, resilient AI workflows.

If your organisation is exploring AI, data modernisation, or temporal data capabilities, we’d welcome a conversation.

Reference list:

1. MenatQA (Findings EMNLP 2023): https://aclanthology.org/2023.findings-emnlp.100.pdf
2. WHO Collaborating Centres directory: https://who.my.site.com/ecc/s/database-search
3. On the Robustness of Temporal Factual Knowledge in Language Models (arXiv): https://arxiv.org/html/2502.01220v1
4. Time-Aware Language Models as Temporal Knowledge Bases (TACL/ACL): https://aclanthology.org/anthology-files/anthology-files/pdf/tacl/2022.tacl-1.15.pdf
5. Continual Learning of Large Language Models: A Comprehensive Survey (ACM DL): https://dl.acm.org/doi/10.1145/3735633
6. TimE Benchmark (arXiv): https://arxiv.org/html/2505.12891v4
7. Time Matters: Temporal Effects on Biomedical Language Models (PMC): https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12099427/
8. LLMs as few-shot multivariate time series classifiers (SpringerLink): https://link.springer.com/article/10.1007/s10618-025-01145-z
9. How Do Large Language Models Acquire Factual Knowledge? (NeurIPS 2024): https://proceedings.neurips.cc/paper_files/paper/2024/file/6fdf57c71bc1f1ee29014b8dc52e723f-Paper-Conference.pdf
10. ChronosLex: Time-aware Incremental Training for Temporal Generalization (ACL 2024): https://aclanthology.org/2024.acl-long.166.pdf
11. A Study into Investigating Temporal Robustness of LLMs (Findings ACL 2025): https://aclanthology.org/2025.findings-acl.810.pdf
12. Disease trajectory browser for exploring temporal, population-wide disease progression patterns in 7.2 million Danish patients.  Nature Communication:  https://www.nature.com/articles/s41467-020-18682-4