Research map / Probabilistic and scientific modeling / Molecules and proteins

Realism VS Accuracy: Event Sequence Forecasting from a Generative Modeling Perspective

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Neural point-process models

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Benchmarks event sequence forecasting with heterogeneous features on both prediction accuracy (TMS) and distributional realism (cDFS), finds an accuracy-realism trade-off, and introduces Wasserstein BaryBooster, which aggregates diffusion samples into a top-accuracy predictor.

Problem

Long-horizon forecasting of event sequences with mixed numerical and categorical features is evaluated almost only on prediction accuracy, ignoring whether generated continuations are distributionally realistic.

Approach

Frames event-sequence forecasting as conditional generation and evaluates it on two axes: a new matching-based accuracy metric (TMS) and a history-conditioned classifier two-sample realism metric (cDFS). It then aggregates multiple diffusion samples around their TMS barycenter (Wasserstein BaryBooster) to convert realism into accuracy.

Key results

• LDM-WB reaches the highest TMS on all three datasets (0.495 Alphabattle, 0.436 Gender, 0.522 Age at L=32), above Repeat (0.440, 0.361, 0.454) and DEF (0.424, 0.413, 0.508).
• LDM is the only generative model with non-trivial cDFS (0.856 Alphabattle, 0.907 Gender, 1.000 Age), while DEF and most learned models score near 0.
• Repeat lies on the Pareto front of accuracy versus realism across all datasets among baselines.

Surprising findings

• A one-line Repeat baseline (copy last L events) is the strongest accuracy baseline on Alphabattle and lies on the baseline Pareto front.
• No method excels at both accuracy and realism; the strongest accuracy models have near-zero cDFS, revealing a trade-off.

Limitations

• Only three card-transaction datasets; no non-financial corpus met the criteria.
• Wasserstein BaryBooster adds substantial inference-time overhead from sampling N candidates and barycentric aggregation.
• cDFS absolute values depend on the discriminator; ranking stability checked only on a limited set of configurations.

Experimental setup

Datasets: Gender Age Alphabattle

Benchmarks: Gender Age Alphabattle

Metrics: TMS cDFS Shape Trend R1 F1

Baselines: Repeat HistSampler Mode AR GPT-2 MT GPT-2 DDM (CDiff) DEF

Builds on: TabSyn CDiff DEF HoTPP T-mAP Optimal Transport Distance

Closest papers by meaning

• Interacting Diffusion Processes for Event Sequence Forecasting (ICML 2024 · similarity 0.86)
• EasyTPP: Towards Open Benchmarking Temporal Point Processes (ICLR 2024 · similarity 0.85)
• Time Series Generation Under Data Scarcity: A Unified Generative Modeling Approach (NeurIPS 2025 · similarity 0.85)
• LLM Processes: Numerical Predictive Distributions Conditioned on Natural Language (NeurIPS 2024 · similarity 0.84)
• HDT: Hierarchical Discrete Transformer for Multivariate Time Series Forecasting (AAAI 2025 · similarity 0.84)
• TSGM: A Flexible Framework for Generative Modeling of Synthetic Time Series (NeurIPS 2024 · similarity 0.84)