Temporal Knowledge Graph Forecasting

• Temporal Knowledge Graph forecasting is a technique that extends static graphs with time, representing facts as quadruples (s, r, o, t) to predict future events.
• It challenges models to capture both intra-snapshot structural dependencies and inter-snapshot temporal evolution while addressing unseen entities and long-term dependencies.
• Recent approaches use neural, reinforcement learning, and rule-based methods to enhance explainability and quantify uncertainty in predictions.

Temporal Knowledge Graph (TKG) forecasting aims to predict future facts based on given histories. A TKG extends a traditional knowledge graph with time, commonly representing each fact as a quadruple $(s,r,o,t)$ and the full graph as a sequence of timestamped snapshots. In the extrapolative setting, a query such as $(s,r,?,t)$, $(?,r,o,t)$, or, in some formulations, $(s,?,o,t)$, must be answered using only facts that occurred before the query time. The resulting problem couples temporal dynamics, multi-relational structure, and ranking over candidate entities or relations, and it differs fundamentally from interpolation or completion at observed timestamps (Xia et al., 2024, Lee et al., 2023).

1. Formal task and forecasting regimes

A standard formalization treats a TKG as either a set of temporal quadruples or a sequence of graphs $\mathcal{G}=\{\mathcal{G}_1,\mathcal{G}_2,\ldots\}$. Forecasting is then defined as predicting a missing subject or object for a future query from the historical sequence before the query time, for example from $\mathcal{G}_{<t_q}$ for a query at $t_q$. Several works make the extrapolative constraint explicit: the model may use only facts with timestamps strictly earlier than the query time, and the test period is chronologically later than the training period (Xia et al., 2024, Han et al., 2021, Sun et al., 2021).

This extrapolative constraint is the main distinction from temporal knowledge graph completion at observed timestamps. One line of work states that, compared with the completion task, the forecasting task is more difficult because it faces two main challenges: how to effectively model the time information to handle future timestamps, and how to make inductive inference to handle previously unseen entities that emerge over time (Sun et al., 2021). Another line emphasizes that forecasting requires the model to jointly capture structural dependencies within snapshots and temporal evolution across snapshots, while some newer work argues that stateless snapshot-wise encoders cause “episodic amnesia” and rapid decay of long-term dependencies (Li et al., 12 Feb 2026).

The task is most often posed as entity forecasting, but the literature also includes relation forecasting and broader temporal reasoning. In geopolitical forecasting with hyper-relational temporal structures, relation prediction is written as $((s,?,o,t),Q)$, and forecasting means that the model must infer the missing relation label for a future event using only earlier data (Ahrabian et al., 1 Jan 2026). In selective forecasting, the forecasting function itself may abstain and return $\varnothing$ when confidence is low, rather than forcing a prediction for every query (Hou et al., 2024).

A further refinement is the notion of higher-order history. One recent formulation defines first-order histories as all past facts with the same subject as the query, then recursively constructs second-order and higher-order chains by following intermediate entities across time. This formulation makes explicit that forecasting is not only about one-hop evidence, but also about temporal chains of events that may reveal how the answer is reached through intermediate entities and relations (Xia et al., 2024).

2. Benchmarks, metrics, and evaluation protocols

The benchmark landscape is heterogeneous but highly standardized in its core event-forecasting setting. Frequently used datasets include ICEWS14, ICEWS18, ICEWS05-15, WIKI, YAGO, and GDELT; later work also adds ACLED-CD22, POLECAT-derived datasets, and newer TGB 2.0 datasets such as smallpedia, polecat, icews, and wikidata (Han et al., 2021, Lee et al., 2023, Gastinger et al., 11 Sep 2025). Chronological train/validation/test splits are central. Some papers report standard chronological splits such as $8{:}1{:}1$ with no temporal leakage, while others use dataset-specific temporal partitions so that all training timestamps are strictly earlier than test timestamps (Li et al., 12 Feb 2026, Ahrabian et al., 1 Jan 2026).

The dominant ranking metrics are MRR and Hits@$(s,r,?,t)$0, usually under time-aware filtering. Time-aware filtering removes alternative valid answers at the same timestamp before ranking, and several papers present it as the fairest evaluation setting for forecasting because a fact invalid at one time may be valid at another (Han et al., 2021, Lee et al., 2023). Some papers additionally report raw and time-unaware filtered settings, while LLM-based multiple-choice formulations instead report Acc@4, Acc@6, and Acc@10 (Chang et al., 21 Jan 2025). In one generalization-oriented LLM study, MRR is not used because LLMs do not naturally produce a full ranking over all entities (Bai et al., 31 May 2025).

Single-step and multi-step forecasting are also distinguished. In single-step evaluation, a model may use true facts from earlier test timestamps, whereas in multi-step evaluation it must use its own previous predictions instead of ground truth facts, making the setting harder and more realistic (Lee et al., 2023, Sannidhi et al., 2024). Long-horizon forecasting is sometimes studied by varying the future horizon $(s,r,?,t)$1, revealing that performance typically drops as the forecast horizon grows (Han et al., 2021).

Dataset properties materially affect conclusions. A recurrency-oriented baseline study reports that YAGO and WIKI have very high recurrence, whereas ICEWS14 and ICEWS18 have much lower direct recurrency. The same study argues that the importance of simple baselines is often neglected and that rigorous comparison against recurrence heuristics is necessary to distinguish real progress from illusory progress (Gastinger et al., 2024). This suggests that cross-paper comparisons are meaningful only when the protocol, filtering setting, and dataset recurrence profile are aligned.

3. Neural, graph-based, and continuous-time forecasting models

A major family of models forecasts by learning dynamic entity and relation representations from historical snapshots. TANGO extends neural ordinary differential equations to multi-relational graph convolutional networks and models hidden states continuously over time rather than as discrete states. Its formulation combines a multi-relational graph convolution component with a graph transition component so that both temporal evolution and edge formation or dissolution are represented explicitly (Han et al., 2021). The graph transition layer is designed to focus on what changes in the graph, and the model is evaluated with DistMult or TuckER decoders over all candidate entities (Han et al., 2021).

Path-based forecasting provides a different mechanism. TimeTraveler formulates forecasting as reinforcement learning over historical knowledge graph snapshots: an agent starts at the query entity, traverses temporal edges backward in time, and searches for the answer. Its key ingredients are a relative time encoding function, a time-shaped reward based on a Dirichlet distribution, and an Inductive Mean representation for unseen entities (Sun et al., 2021). Because predictions arise from explicit temporal paths, this line of work emphasizes explainability as well as inductive inference.

Other models focus on richer graph structure. LMS, or Learning Multi-graph Structure, combines an evolutional graph over recent snapshots, a query-specific union graph over relevant historical facts, and a temporal graph over timestamps. It then fuses evolutional and union representations with an adaptive gate and uses time-aware decoders constrained by historical statistics (Zhang et al., 2023). EST, or Entity State Tuning, makes a more radical architectural claim: most existing methods are stateless and recompute entity representations at each timestamp from a limited query window, whereas persistent and continuously evolving entity states can be maintained in a global state buffer and updated by a dual-track memory mechanism balancing plasticity against stability (Li et al., 12 Feb 2026).

Uncertainty-aware continuous-time forecasting introduces yet another axis. WGP-NN employs weighted Gaussian processes to jointly model the temporal evolution of event occurrence probability and time-dependent uncertainty. The model combines neighborhood aggregation, pseudo-point generation, and a weighted Gaussian process so that forecasting yields not only a mean score but also predictive variance for future links (Nolting et al., 2023). This directly targets a limitation noted in the paper: prior methods can express uncertainty in learned entity representations, but not the predictive uncertainty of the future link itself (Nolting et al., 2023).

Taken together, these models show that TKG forecasting has not converged on a single representation strategy. Continuous-time dynamics, temporal path search, multi-graph structure, persistent state memory, and predictive uncertainty have all been argued to capture aspects that snapshot-wise transductive scoring misses. A plausible implication is that the main architectural disagreements concern where temporal signal should live: in embeddings, in paths, in timestamp semantics, in persistent memory, or in predictive distributions.

4. Rules, recurrence, and explainable forecasting

Explainable forecasting has a strong rule-based tradition. TLogic is presented as an explainable framework for link forecasting on temporal knowledge graphs based on temporal logical rules extracted via temporal random walks. Its stated goals are to provide explanations that preserve time consistency and to work well in the inductive setting where already learned rules are transferred to related datasets with a common vocabulary (Liu et al., 2021). In this strand of work, forecasting is not primarily a latent representation problem but a problem of extracting and applying temporally constrained symbolic dependencies.

A striking counterpoint to increasingly complex models is the recurrency baseline literature. One baseline paper defines strict recurrency, relaxed recurrency, and a combined recurrency model, all motivated by the idea that what happened before may happen again in the future. Its empirical finding is that, compared to 11 methods on five datasets, the baseline ranks first or third in three of them (Gastinger et al., 2024). The same paper reports that none of the analyzed methods pass the “sanity check” of consistently modeling simple recurrency across all relations, and that on YAGO and WIKI recurrence is such a dominant signal that a simple heuristic nearly matches or exceeds many learned methods (Gastinger et al., 2024).

More recent rule learning refines this symbolic approach rather than abandoning it. CountTRuCoLa learns four simple types of rules—$(s,r,?,t)$2-rules, $(s,r,?,t)$3-rules, $(s,r,?,t)$4-rules, and $(s,r,?,t)$5-rules—and uses a confidence function that combines recency and frequency. The method is explicitly fully explainable: each prediction is generated by explicit symbolic rules, each rule has a simple semantic interpretation, the score decomposes into recency and frequency terms, and the prediction can be traced back to exact supporting facts and their time distances (Gastinger et al., 11 Sep 2025). Evaluated on nine datasets, it is reported to achieve the highest test MRR on 7 of 9 datasets and second-best on the remaining two (Gastinger et al., 11 Sep 2025).

These results support two recurring themes. First, explainability in TKG forecasting is often operationalized as path or rule traceability rather than post hoc feature attribution. Second, performance gains do not necessarily require greater architectural complexity; in recurrence-dominated datasets, carefully designed symbolic models can match or surpass neural systems. This directly addresses a common misconception that more expressive latent models are uniformly superior across benchmarks.

5. LLM-based forecasting and graph–language hybrids

LLMs introduced a new formulation in which forecasting is cast as prompting, generation, or hybrid graph–language reasoning. One early result shows that TKG forecasting can be reframed as in-context learning: relevant historical facts are converted into prompts and the model generates ranked predictions from token probabilities. The central empirical claim is that LLMs, out-of-the-box, perform on par with state-of-the-art TKG models carefully designed and trained for TKG forecasting, and that using numerical indices instead of entity or relation names does not significantly affect the performance, showing that prior semantic knowledge is unnecessary in that setup (Lee et al., 2023). This finding has often been interpreted as evidence that LLMs can leverage patterns in the prompt structure rather than lexical semantics.

Subsequent work tries to repair weaknesses of naive prompting. GenTKG combines temporal logical rule-based retrieval with few-shot parameter-efficient instruction tuning, so that only a small set of historically relevant facts is placed in context and the LLM is aligned to generate the missing entity directly (Liao et al., 2023). Chain-of-History reasoning argues that first-order history alone is insufficient and instead explores high-order histories step-by-step, selecting important histories or history chains before final prediction; it can also act as a plug-and-play module that fuses LLM scores with graph-based backbones such as RE-NET, RE-GCN, and TiRGN (Xia et al., 2024). G2S, or General-to-Specific learning, further claims that forecasting entangles general patterns and scenario information, and therefore first trains on anonymous temporal structures and only later injects scenario mappings through in-context learning or fine-tuning (Bai et al., 31 May 2025).

A second branch strengthens graph–language alignment directly. TGL-LLM introduces temporal graph learning, hybrid graph tokenization, and a two-stage training paradigm to integrate temporal graph embeddings into an LLM-based TKG model (Chang et al., 21 Jan 2025). SPARK reframes TKG forecasting as a top-$(s,r,?,t)$6 sequence-level generation problem, uses beam search for single-pass candidate generation, and then refines LLM outputs with a trainable TKG adapter derived from a traditional forecasting model such as xERTE or TLogic (Yin et al., 27 Mar 2025). RECIPE-TKG focuses on sparse-history settings and combines rule-based multi-hop retrieval, contrastive LoRA fine-tuning, and test-time semantic filtering to improve semantic coherence and filtered Hits@10 (Akgül et al., 23 May 2025). A more radical variant, sLA-tKGF, avoids a pre-trained LLM as predictor and instead trains a small transformer from scratch under a retrieval-augmented, tabula rasa design that merges historical TKG facts, web search results, and PLLM-generated descriptions into knowledge-infused prompts (Sannidhi et al., 2024).

These LLM-based results suggest a division of labor rather than a clean replacement of graph models. Pure prompting emphasizes symbolic pattern induction from text-like serialized history; retrieval-augmented generation emphasizes selecting the right temporal evidence; plug-and-play hybrids use graph models to supply global structure or re-ranking; and graph tokenization approaches try to make temporal graph states legible to sequence models. A plausible implication is that the main open question is no longer whether LLMs can participate in TKG forecasting, but which parts of the forecasting pipeline should remain graph-native.

6. Reliability, richer structures, and adjacent forecasting tasks

Reliability has become a distinct concern. Selective Temporal Knowledge Graph Reasoning argues that existing TKG reasoning models are unable to abstain from predictions they are uncertain about, which creates risks in real-world applications. Its CEHis module combines a certainty score from the current prediction with a history-based score derived from the model’s own past accuracy on related queries, then abstains when the fused confidence is too low (Hou et al., 2024). The same general concern appears in uncertainty-aware forecasting: WGP-NN models predictive variance for future links, and some recent stateful work introduces Counterfactual Consistency Learning to reduce visibility bias or spurious correlations (Nolting et al., 2023, Li et al., 12 Feb 2026).

Forecasting has also expanded beyond entity ranking. ForecastTKGQuestions defines forecasting temporal knowledge graph question answering, where future questions must be answered using only facts before the question timestamp. The associated ForecastTKGQA model couples a forecasting TKG module with a language understanding module so that entity prediction, yes-unknown questions, and fact reasoning questions can all be handled under the forecasting constraint (Ding et al., 2022). This extension shows that TKG forecasting can serve as an internal inference component for broader temporal QA rather than as an end in itself.

Richer event structure is another emerging direction. One geopolitical forecasting study argues that hyper-relational temporal knowledge graphs still lack support for more than two primary entities in temporal facts, and therefore introduces Hyper-Relational Temporal Knowledge Generalized Hypergraphs (HTKGHs). In that formulation, a temporal fact may contain a group of actors, a relation, a group of recipients, a timestamp, and qualifiers, and the forecasting task becomes relation prediction over these higher-order event structures (Ahrabian et al., 1 Jan 2026). This suggests that some limitations attributed to forecasting models may actually stem from the representational bottleneck of pairwise temporal edges.

Several practical limitations recur across the literature: long-horizon performance degrades as the forecast horizon increases; LLM methods remain constrained by context length or multi-token ID generation; stateful methods incur memory overhead $(s,r,?,t)$7; and many methods assume a stable entity universe, making newly emerging entities a persistent challenge (Han et al., 2021, Bai et al., 31 May 2025, Li et al., 12 Feb 2026). When combined with the recurrence-baseline results, these observations motivate a cautious reading of benchmark progress. This suggests that future work will need to improve not only model architecture, but also forecasting protocols, confidence estimation, structural expressiveness, and the treatment of sparse, inductive, and long-horizon settings.