Next Event Prediction (NEP) Overview

• NEP is a forecasting task family that estimates the next event from historical sequences across various domains such as healthcare, business processes, and video reasoning.
• It leverages multiple history representations like timestamped event chains, semantic narratives, and evolutionary state graphs to capture temporal dynamics.
• NEP models utilize diverse objectives including autoregressive prediction and classification to achieve measurable gains in metrics like AUROC and C-index.

Next Event Prediction (NEP) denotes a family of forecasting tasks that estimate what happens next from an observed history. Across current literature, the “next event” may be a diagnosis, procedure, medication, or lab in a patient trajectory; the next activity in a business-process trace; the more plausible future event given a video clip and dialogue; a missing entity in a future temporal knowledge-graph fact; a timestamp-label pair in a marked temporal point process; or a future interaction inside a continuous-time dynamic graph. Accordingly, NEP appears both as single-step classification and as joint prediction of event type, timestamp, existence, or multiple future events within a horizon (Chen et al., 29 Sep 2025, Oved et al., 2024, Lei et al., 2020, Yu et al., 2024, Karpukhin et al., 2024, Wang et al., 2022).

1. Formal scope of the task

A defining feature of NEP is that the prediction target is always conditioned on a prefix, but the formal object being predicted varies substantially by domain. In clinical foundation modeling, an electronic health record is written as a chronological chain of timestamped events, $\mathcal{P}=\{e_1,e_2,\ldots,e_n\}$, and the core objective is

$$p(e_{t+1}\mid e_1,e_2,\ldots,e_t)=\mathrm{LLM}_\theta(e_1,e_2,\ldots,e_t),$$

with each event containing an event type, an event value, and a timestamp (Chen et al., 29 Sep 2025).

In predictive business process monitoring, the standard formulation is a prefix-to-next-activity map. If a trace prefix is $p_k=\langle e_1,e_2,\dots,e_k\rangle$, the target is the next activity $a_{k+1}$, written as $f(p_k)=a_{k+1}$. Related business-process formulations define an event as a tuple with activity, timestamp, and optional attributes, and train on prefix-label pairs generated from sliding windows or from all prefixes of a trace (Oved et al., 2024, Engelhard et al., 2020, Tama et al., 2020).

Marked temporal point process work makes the distinction between next-event prediction and long-horizon forecasting explicit. There the history is a sequence of timestamp-label pairs $(t_i,l_i)$, and standard evaluation predicts the next event’s label and time, whereas long-horizon settings predict $K$ future candidates inside a horizon $(t,t+\mathrm{H})$ (Karpukhin et al., 2024). Temporal knowledge graph forecasting further shifts the target from event labels to future entities in quadruples $(s,r,o,t)$, typically predicting $(s,r,?,t+k)$ or $$p(e_{t+1}\mid e_1,e_2,\ldots,e_t)=\mathrm{LLM}_\theta(e_1,e_2,\ldots,e_t),$$0 from past graph snapshots (Yu et al., 2024). In multimodal video-language work, NEP can also be cast as binary multiple-choice forecasting: given a premise clip with aligned dialogue and two candidate future events, choose which one is more likely to happen next (Lei et al., 2020).

This suggests that NEP is best understood as a task family rather than a single canonical problem specification.

2. Representing history and temporal context

A major axis of variation in NEP research is how the observed past is represented before prediction. One line of work argues that compressing a sequence into a current-state summary obscures order and timing. In the EHR setting, the proposed shift is from visit-level modeling to timestamped event chains, preserving the order and timing of individual diagnoses, procedures, medications, labs, and vital signs (Chen et al., 29 Sep 2025).

Other domains build richer semantic or structural state descriptions from the same prefix.

Setting	History representation	Preserved signal
EHR foundation modeling	Chronological chain of timestamped events	Order and timing of individual clinical events
Business-process NAP	Coherent natural-language story from a prefix trace	Semantic content, temporal attributes, free text
Petri-net-based NEP	$$p(e_{t+1}\mid e_1,e_2,\ldots,e_t)=\mathrm{LLM}_\theta(e_1,e_2,\ldots,e_t),$$1	Decay response, token counters, marking, resource counts
Time-series event prediction	Evolutionary state graph $$p(e_{t+1}\mid e_1,e_2,\ldots,e_t)=\mathrm{LLM}_\theta(e_1,e_2,\ldots,e_t),$$2	Changing relations among states

In semantic business-process modeling, SNAP converts each trace prefix into a coherent natural-language story and fine-tunes a language foundation model for multiclass next-activity prediction. The story may include the current activity, turn number, temporal features, role or resource attributes, prior activity sequence, and free-text fields such as user utterances and bot responses (Oved et al., 2024). In contrast, DREAM-NAP begins from a discovered Petri net, augments each place with a time decay function, replays the log, and concatenates decay response, token movement counters, markings, and optional resource counts into a timed state sample $$p(e_{t+1}\mid e_1,e_2,\ldots,e_t)=\mathrm{LLM}_\theta(e_1,e_2,\ldots,e_t),$$3 (Theis et al., 2019).

Irregular timing has motivated dedicated time representations. Rather than concatenating raw time, one approach contextualizes event embeddings with duration via Time Mask or combines event embeddings with a learned soft categorical time embedding via Event-Time Joint Embedding. The same work adds next-event-duration regularization to encourage hidden states to retain temporal structure useful for event prediction (Li et al., 2017). In time-series event prediction, the representation becomes explicitly graph-structured: segments are mapped to representative states, transitions between state mixtures define edge weights, and the resulting evolutionary state graph is updated over time (Hu et al., 2019).

The recurring theme is that NEP performance is tightly coupled to what counts as “history”: raw sequence, structured state, semantic narrative, or evolving graph.

3. Learning objectives and model families

Autoregressive prediction is one prominent paradigm, but not the only one. In clinical NEP, the objective is framed as causal autoregressive language modeling under a causal mask, explicitly contrasted with bidirectional masked modeling because allowing the model to “peek” at future events would violate the temporal logic of clinical forecasting. Implementation uses decoder-only LLMs, structured instruction-response pairs, and cross-entropy over the target tokens of the next event (Chen et al., 29 Sep 2025).

Classification-style objectives remain common elsewhere. VLEP uses a transformer-based multimodal classifier over video, dialogue, and commonsense knowledge, with cross-entropy loss to select the correct future event candidate (Lei et al., 2020). SNAP, classical business-process next-activity prediction, and several benchmark studies also formulate NEP as multiclass classification over activity labels, often with softmax outputs and cross-entropy loss (Oved et al., 2024, Weinzierl et al., 2020, Tama et al., 2020).

A different family replaces positional supervision with set prediction. DeTPP predicts $$p(e_{t+1}\mid e_1,e_2,\ldots,e_t)=\mathrm{LLM}_\theta(e_1,e_2,\ldots,e_t),$$4 future event candidates in parallel, each with a presence probability $$p(e_{t+1}\mid e_1,e_2,\ldots,e_t)=\mathrm{LLM}_\theta(e_1,e_2,\ldots,e_t),$$5, a predicted time shift $$p(e_{t+1}\mid e_1,e_2,\ldots,e_t)=\mathrm{LLM}_\theta(e_1,e_2,\ldots,e_t),$$6, and a label distribution $$p(e_{t+1}\mid e_1,e_2,\ldots,e_t)=\mathrm{LLM}_\theta(e_1,e_2,\ldots,e_t),$$7. Ground truth events within a horizon are matched to predictions with the Hungarian algorithm, and optimization uses a matching-based loss that combines time error, label likelihood, and a binary cross-entropy presence term. This design is meant to avoid error accumulation, repetitive outputs, and incorrect fixed-position alignment (Karpukhin et al., 2024).

Adversarial training has also been used to improve generalization. One business-process framework defines a generator $$p(e_{t+1}\mid e_1,e_2,\ldots,e_t)=\mathrm{LLM}_\theta(e_1,e_2,\ldots,e_t),$$8 and discriminator $$p(e_{t+1}\mid e_1,e_2,\ldots,e_t)=\mathrm{LLM}_\theta(e_1,e_2,\ldots,e_t),$$9, both implemented as LSTMs, and combines adversarial feedback with the conventional supervised next-step loss $p_k=\langle e_1,e_2,\dots,e_k\rangle$0. A related encoder-decoder model for suffix and remaining-time prediction uses open-loop training, Gumbel-Softmax for differentiable discrete generation, and a discriminator over generated versus real suffixes (Taymouri et al., 2020, Taymouri et al., 2021).

Several recent methods target failure modes beyond ordinary supervised training. CaseQ argues that maximum likelihood estimation fails under temporal distribution shift because of a latent context confounder, and approximates $p_k=\langle e_1,e_2,\dots,e_k\rangle$1 through backdoor adjustment and variational inference with hierarchical branching structures (Yang et al., 2022). WGP-LN and FD-Dir move from point predictions to time-dependent distributions on the simplex, capturing how the entire next-event distribution evolves with future time and how uncertainty grows in poorly supported regions (Biloš et al., 2019). CET separates event occurrence from event timing with a binary stochastic layer $p_k=\langle e_1,e_2,\dots,e_k\rangle$2, learning whether an event will ever occur and, if so, when (Engelhard et al., 2020).

4. Major application domains

Healthcare has recently treated NEP as a foundation-modeling strategy rather than only a downstream task. In oncology-heavy EHR modeling, records are reformulated as timestamped event chains, the model is fine-tuned to predict the next clinical event, and patient embeddings extracted from final hidden states are then used for survival prediction and diagnosis tasks. The paper explicitly positions this as learning the next step in a patient’s medical journey rather than a static representation of current state (Chen et al., 29 Sep 2025).

Business process monitoring provides an older and more heterogeneous NEP tradition. Here the next event is usually the next activity label, though some systems also predict timestamps or full suffixes. Methods range from semantic stories with BERT, DeBERTa, and GPT-3; to GAN-like training for next activity and next timestamp; to Petri-net-guided state encoding; to broad empirical comparisons of classifiers, encoding windows, and ensembles (Oved et al., 2024, Taymouri et al., 2020, Theis et al., 2019, Weinzierl et al., 2020, Tama et al., 2020).

Video and multimodal reasoning have produced two distinct formulations. VLEP asks which of two future events is more likely after a short video clip with dialogue, emphasizing commonsense, intention, reaction, and causal reasoning (Lei et al., 2020). A later line treats NEP as a self-supervised objective for multimodal LLMs: the model sees only past video frames and generates a summary of future events derived from future frames, with V1-33K as training data and FutureBench as an evaluation set (Wang et al., 28 May 2025). A further extension, Video-Next-Event Prediction, shifts the answer modality itself from text to generated video by coupling a vision-LLM and a video diffusion model through Joint-GRPO on VANS-Data-100K (Cheng et al., 20 Nov 2025).

Graph-structured forecasting supplies several additional variants. ONSEP performs temporal knowledge graph forecasting by combining Dynamic Causal Rule Mining with Dual History Augmented Generation, using short-term history and long-term causal history at test time (Yu et al., 2024). CEP3 models community event forecasting on continuous-time dynamic graphs, jointly predicting timestamp, source node, and destination node through a hierarchical factorization and autoregressive graph updates (Wang et al., 2022). EvoNet represents segmented time series as evolutionary state graphs and predicts binary future events from node-level and graph-level propagation (Hu et al., 2019). DeTPP addresses transactional, medical, social, and financial event streams within marked temporal point processes, emphasizing robust long-horizon event forecasting with next-event extensions (Karpukhin et al., 2024).

5. Benchmarks, metrics, and empirical behavior

Empirical evaluation in NEP is highly domain-specific, and metrics vary accordingly. In clinical foundation modeling, training used more than 1.2 million patients and about 200 million clinical events, with public evaluation on MSK-CHORD, which has 24,950 patients, and on EHRSHOT, where the authors focus on the new-diagnosis task. Reported headline gains state that NEP improves temporal reasoning tasks by 7.2% C-index over general-purpose LLMs and surpasses specialized EHR models by 4.6% AUROC; in oncology survival, NEP-8B improves 7 of 10 cancer-stage subgroups and shows particularly strong gains in Stage IV metastatic settings (Chen et al., 29 Sep 2025).

VLEP established a benchmark of 28,726 future event prediction examples from 10,234 video clips, split into 20,142 train, 4,392 dev, and 4,192 test examples. On the test set, chance is 50.00%, the full video + dialogue + future baseline reaches 67.46%, humans achieve 76.25% with dialogue and future candidates only, and 90.50% with full video plus dialogue. The resulting model-human gap is a central empirical finding (Lei et al., 2020). In the later video-NEP line, FutureBench contains 1056 QA pairs, and even a strong text-only reasoning model gets only 32.0% accuracy, which is presented as evidence that the benchmark requires visual and temporal understanding (Wang et al., 28 May 2025).

For point-process forecasting, DeTPP is evaluated on HoTPP across StackOverflow, Amazon, Retweet, MIMIC-IV, and Transactions. The paper reports state-of-the-art results on 9 out of 10 comparisons, up to a 77% relative improvement on long-horizon forecasting, and up to a 2.7% relative improvement on the large Transactions dataset for next-event prediction with DeTPP+ (Karpukhin et al., 2024). ONSEP reports Hit@1 improvements over ICL of 9.63% on ICEWS14, 9.35% on ICEWS05-15, and 16.28% on ICEWS18 with InternLM2-7B and history length $p_k=\langle e_1,e_2,\dots,e_k\rangle$3 (Yu et al., 2024).

Business-process NEP remains benchmark-driven. SNAP is evaluated on six datasets and is reported to significantly outperform prior state of the art on all datasets by Wilcoxon signed-rank test, with especially notable gains on the semantically rich MIP dataset (Oved et al., 2024). The GAN-based predictive monitoring framework reports weighted average next-activity accuracies of 0.9518 on Helpdesk, 0.9158 on BPI12(W), 0.9401 on BPI12, and 0.9256 on BPI17, with at least 98% accuracy at the longest prefix lengths for all logs (Taymouri et al., 2020). DREAM-NAPr reports statistically significant gains with $p_k=\langle e_1,e_2,\dots,e_k\rangle$4 under a sign test with Dunn-Sidak correction (Theis et al., 2019).

6. Interpretability, uncertainty, generalization, and recurring methodological issues

Interpretability is an explicit design goal in several NEP lines. In EHR modeling, improved interpretability is attributed to attention weights that align with clinical pathways (Chen et al., 29 Sep 2025). DREAM-NAP retains an explicit Petri net process model with markings, decay functions, and token movement counters (Theis et al., 2019). EvoNet exposes representative states, state transitions, and temporal attention scores, allowing case studies that relate anomalous predictions to specific state paths (Hu et al., 2019). ONSEP stores an explicit causal rule base $p_k=\langle e_1,e_2,\dots,e_k\rangle$5, making its neural-symbolic memory more inspectable than static retrieval alone (Yu et al., 2024).

Uncertainty enters NEP in at least two distinct senses: uncertainty about which event is next, and uncertainty about whether the event occurs at all. WGP-LN and FD-Dir model a distribution over class-probability vectors $p_k=\langle e_1,e_2,\dots,e_k\rangle$6 that evolves with future time, rather than a single categorical output, and use uncertainty-aware objectives to discourage overconfident extrapolation (Biloš et al., 2019). CET separates $p_k=\langle e_1,e_2,\dots,e_k\rangle$7 from the distribution of event time conditioned on occurrence, which is critical when “no event” cannot be represented adequately as merely “very late event” (Engelhard et al., 2020).

Generalization under shift is another recurrent concern. CaseQ attributes out-of-distribution sequential event prediction failure to a latent context confounder and targets an interventional estimand $p_k=\langle e_1,e_2,\dots,e_k\rangle$8 (Yang et al., 2022). ONSEP emphasizes online adaptation by mining and updating causal rules during testing, and reports that dynamic updating beats relying on stale preloaded rules (Yu et al., 2024).

Several recurring methodological issues are treated explicitly in benchmark papers. Choosing an optimal number of events for feature encoding is described as challenging, and larger windows can sharply reduce the number of available training samples (Tama et al., 2020). Ensemble schemes improve the performance of low performing classifiers such as SVM, whereas high performing classifiers such as tree-based classifiers are not better off when ensemble schemes are considered (Tama et al., 2020). In context-enriched DNN benchmarks, encoding strategy is reported to matter more than architecture on average, with hash encoding strongest overall and ordinal encoding consistently worst (Weinzierl et al., 2020). In video NEP, increasing dataset size does not always improve performance, and GRPO reinforcement learning can improve FutureBench while also hurting general benchmark performance and encouraging reward hacking or shortcut behavior (Wang et al., 28 May 2025). A plausible implication is that NEP remains methodologically plural: gains often come not from a universal backbone, but from aligning representation, objective, and evaluation with the causal and temporal structure of the underlying event stream.