Flexible Intensities Matter: A Comprehensive Re-evaluation of Hawkes Processes and Neural Temporal Point Processes

Hendrik Mehrtens, Florian Buettner, Oliver Stegle

TL;DR

The supposed performance gap between classical Hawkes processes and modern neural models is largely an artifact of comparing "weak" classical baselines against "strong" neural architectures. By equipping Hawkes processes with flexible spline-based components, they become competitive with—and sometimes outperform—neural models on standard benchmarks, revealing that the primary advantage of neural TPPs lies in modeling when events occur rather than which type of event happens.

What Problem Does This Paper Address?

Temporal Point Processes (TPPs) are the mathematical workhorse for modeling event sequences in continuous time, such as social media interactions, seismic activity, or patient journeys in a hospital. For decades, the Hawkes process was the standard, modeling events as a sum of a baseline rate and "triggering" effects from past events.

Recently, the research community shifted toward Neural TPPs, which use deep learning (RNNs, Transformers) to model complex, non-linear dependencies. Recent benchmarks suggested that Neural TPPs significantly outperform Hawkes processes. However, the authors of this paper noticed a flaw in the comparison: researchers were pitting highly flexible neural models against the most restrictive, "vanilla" versions of Hawkes processes (constant baselines and simple exponential kernels). This paper asks: if we give classical models the same level of flexibility, does the neural advantage disappear?

Methods

The authors developed TORCHPP, a modular framework that allows for a "fair fight" by systematically increasing the flexibility of both classical and neural models.

  1. Flexible Hawkes Processes: Instead of fixed exponential decays, they implemented kernels and baselines using learnable linear splines. They also introduced inhibitive variants using a Softplus clamping function, allowing events to suppress (rather than just excite) future occurrences.
  2. Neural TPPs: They evaluated standard architectures like THP and SAHP but introduced NeuralSplineTPP, which uses a Transformer encoder paired with a spline-based intensity head. This allows for closed-form cumulative intensity and exact sampling.
  3. Likelihood Decomposition: A key methodological contribution is decomposing the joint log-likelihood into two terms:
  4. $\log p(t)$: How well the model predicts the timing of the next event.
  5. $\log p(d \mid t)$: How well the model predicts the type (mark) of the next event, given the time.
Figure 1: Conceptual differences between Hawkes processes (1) and neural TPPs (2) on an event sequence (• type A, ▲type B). Top (Hawkes process). Each past event interacts via a kernel response ϕdi→d(t −ti) (colored curves) on top of a (possibly time-varying) baseline µd(t) (thin solid line). The total intensity (upper sub-panel) is the sum of baseline and active kernels. At a query time t, the intensity thus decomposes into the sum of the baseline and the kernel values. Bottom (Neural TPP). The same event history is consumed by a neural encoder, which produces a hidden state at each event location. The head of the neural TPP projects this encoding to an intensity function. Conceptually, this allows the neural TPP to model higher-order non-additive effects between events, in exchange for losing the interpretable decomposition of the Hawkes process.

Datasets

The study spans seven datasets, ranging from small benchmarks to massive medical records:

  • Standard Benchmarks: Taxi (NYC trips), StackOverflow (user badges), Retweet, Amazon (reviews), and Taobao (user behavior).
  • Large-Scale Medical Tasks:
    • UK Biobank (UKB): Disease histories for 500k participants (1,257 event types).
  • MIMIC-IV: ICU records containing procedures and medications (117 event types).
Table 1: Dataset statistics. Small- to medium-sized, low-dimensional standard benchmarks are shown in the top rows, and the medical evaluation tasks, with orders of magnitude more instances and event types, in the bottom rows. “D”: number of event types. “Term.”: whether each type occurs at most once. “Mean Len”: mean number of events per sequence. Sequences are ICU-stay-level (with each patient only in one split) for MIMIC-IV and patient-level for UKB.

Key Insights and Ideas

  • The "Gap" is an Artifact: Once Hawkes processes are given flexible splines and inhibitive kernels, the performance gap reported in previous literature largely vanishes on standard benchmarks. Most of these datasets are dominated by pairwise interactions that Hawkes processes can capture perfectly well.
  • Flexibility = Better Timing, Not Better Typing: The authors found that increasing model flexibility (more spline knots) significantly improves the time prediction term ($\log p(t)$) but has almost zero impact on the type prediction term ($\log p(d \mid t)$). On most benchmarks, models differ far more in when they think an event will happen than what they think it will be.
  • Medical Data is Different: The UK Biobank data, despite its size, is well-explained by additive (Hawkes-like) interactions. However, MIMIC-IV was the only dataset where neural models maintained a clear lead in predicting event types. This suggests that ICU data contains complex, non-additive dependencies that finally justify the overhead of a neural TPP.

Key Results

On the five standard benchmarks, the "souped-up" Hawkes variants matched or beat less flexible neural models (like THP or NHP). For example, on the Taxi dataset, the best Hawkes variant reached a joint log-likelihood nearly identical to the best neural model (a difference of only 0.09 nats).

Table 2: Main results: per-event test log-likelihood on the standard benchmark datasets (nats, higher is better; mean over 5 seeds; EasyTPP convention, i.e. conditioned on the first event and without the final survival term, Section 3.3). Each cell is log p(t) / log p(d | t) (time / type), and the joint log p(t, d) = log p(t) + log p(d | t) is their sum (Table 8; ±std in Appendix H). Models differ chiefly in the time term log p(t), the axis that flexibility improves, while the type term log p(d | t) is comparatively tight across history-aware models (near-flat on StackOverflow and Retweet). Row labels: baseline/kernel (C: constant, S: spline, E: exponential), I.: inhibitive, N.: neural. Bold: best per dataset and component; underline: best joint log-likelihood; ties at the displayed precision are all marked.

In the medical domain:

  • UK Biobank: Flexible Hawkes processes nearly matched the best neural TPP, trailing by only 0.07 nats.
  • MIMIC-IV: NeuralSplineTPP opened a significant gap (0.45 nats) over the best Hawkes variant, particularly in predicting events that occur within the same minute as the previous event.

Limitations and Future Work

  • Scope of Models: The study focused on intensity-based models and one intensity-free baseline. It did not cover newer flow-based, diffusion-based, or state-space TPPs.
  • Task Limitation: The evaluation was strictly on autoregressive next-event prediction. The authors did not test long-horizon simulation or planning.
  • Knot Placement: The NeuralSplineTPP places knots based on time elapsed since the last event, whereas Hawkes splines are anchored to every past trigger event. This structural difference might favor neural models for very short-range dependencies.
  • Future Direction: The authors suggest that future TPP research should move away from "saturated" benchmarks like Retweet and focus on complex, high-dimensional datasets like MIMIC-IV to truly test the limits of neural architectures.

Summary updated

Abstract

Temporal point processes (TPPs) are the canonical framework for event sequence data, with applications spanning seismology, social media, and electronic health records. Recently, neural TPPs have been reported to substantially outperform classical TPPs, such as Hawkes processes, but these comparisons pitted flexible neural models against Hawkes processes in their most restricted form, and did not consider large-scale medical data. Here, we revisit this comparison through a systematic evaluation of Poisson and Hawkes processes of increasing flexibility against neural TPPs, on five established benchmark datasets and on two large-scale medical evaluation tasks derived from UK Biobank disease histories and MIMIC-IV ICU records, under autoregressive next-event prediction. First, the previously reported gap between Hawkes processes and neural TPPs largely disappears once Hawkes processes are equipped with flexible, learnable baselines and kernels. Second, the flexibility of the intensity parametrization also limits prior neural TPPs: a Transformer with a spline-based intensity head outperforms previous intensity-based neural TPPs and is on par with the intensity-free model. Decomposing the log-likelihood into a time term $\log p(t)$ and a type term $\log p(d \mid t)$, we find that this flexibility advantage lies almost entirely in the time density: more flexibility does not improve the type term. On these benchmarks, the gains of neural TPPs thus lie mostly in modeling when events occur, not which. Third, our results underscore the value of complex datasets for evaluating TPPs: only on the MIMIC-IV ICU records do neural TPPs retain an advantage in the type term. Our results inform both future method development and model choice. We release our framework, TorchPP, as an open-source package with consistent implementations of classical and neural TPPs.
BibTeX
@misc{mehrtens2026flexible,
  author = {Hendrik Mehrtens and Florian Buettner and Oliver Stegle},
  title = {{Flexible Intensities Matter: A Comprehensive Re-evaluation of Hawkes Processes and Neural Temporal Point Processes}},
  year = {2026},
  eprint = {2611.06348},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.06348},
}
Figure