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

Dmitry Osin; Egor Surkov; Petr Mokrov; Igor Udovichenko; Ivan Karpukhin; Evgeny Burnaev; Egor Shvetsov; Alexander Korotin

cs.LGOct 8, 2026

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

Authors: Dmitry Osin, Egor Surkov, Petr Mokrov, Igor Udovichenko, Ivan Karpukhin, Evgeny Burnaev, Egor Shvetsov, Alexander Korotin

Organizations: Applied AI Institute · ITMO University · Vega Institute Foundation · SB AI Lab ·

Abstract

Event sequences are chains of random events, each described by heterogeneous (numerical and categorical) features. The problem of event sequence forecasting occurs in finance, healthcare, business process intelligence, and other fields. In this work, we study multi-horizon forecasting of event sequences as a conditional generative modeling problem. We evaluate forecasts not only for prediction accuracy, but also for distributional realism. To this end, we present a benchmark comprising metrics for both quality aspects, strong statistical baselines, recent models adapted from adjacent fields, and curated real-world datasets. Careful evaluation reveals that no method excels at both accuracy and realism, highlighting a fundamental trade-off between the two. Finally, we introduce Wasserstein BaryBooster, an algorithm that aggregates samples from a realism-strong diffusion model into a state-of-the-art predictor, demonstrating that distributional realism can be converted into prediction accuracy. Code is available at https://github.com/xaosina/event-sequence-forecasting.

Figures & tables

Figure 1: A toy event sequence. After a bet on black at roulette, the evening follows one of two realistic chains of transactions: the person either lost (51%) or won (49%). The prediction that minimizes MSE and maximizes per-event accuracy takes the categories of the more likely chain and averages the amounts, pairing cheap categories with large sums (e.g., an $18 bus ride). It scores best on predictive metrics, yet matches neither chain. On the other hand, each possible outcome is realistic, but its expected MSE is about twice as large.

Method	Features	Seq.	Forecasting	Diffusion
CDiff ( Zeng et al., 2024 )	mixed ✓	✓	✓	data
TabSyn ( Zhang et al., 2024 )	mixed ✓	✘	✘	latent
TabDiff ( Shi et al., 2025 )	mixed ✓	✘	✘	data
Add and Thin ( Lüdke et al., 2023 )	none ✘	✓	✓	data
PSDiff ( Lüdke et al., 2025 )	num ✘	✓	✓	data
EdiTPP ( Lüdke et al., 2026 )	none ✘	✓	✓	data

Table 1: Comparison of related generative models. “Features” indicates the supported feature type: “none” (time only), “cat” (a single event type, as in marked TPP, MTPP), “num” (numerical features only, as in spatio-temporal TPP, STPP), or “mixed” (both numerical and categorical). “Seq.” indicates whether the method natively handles sequential (as opposed to single-row tabular) data. “Forecasting” indicates whether the method allows history-conditional generation. “Diffusion” indicates whether the method is a latent diffusion model, a data-space diffusion model, or non-diffusion (ND). Methods applicable to event sequence forecasting are highlighted.

Dataset	# seq. (train / test)	avg. seq. len.	# num. (incl. time)	# cat.	time span
Gender	11K / 2.8K	$404.31\pm 292.48$	2	2	15 months
Age	24K / 6K	$881.00\pm 124.75$	2	2	2 years
Alphabattle	100K / 35K	$366.91\pm 212.58$	2	4	2 years

Table 2: Dataset statistics. Number of sequences in each split, average sequence length, number of categorical and numerical features describing each event, and overall time span.

Method	TMS ↑			cDFS ↑
	Alphabattle	Gender	Age	Alphabattle	Gender	Age
Hist.	0.382	0.221	0.366	0.544	0.476	0.234
Mode	0.203	0.047	0.263	0.000	0.000	0.000
Rep.	0.440	0.361	0.454	0.009	0.050	0.065
AR	0.286	0.188	0.354	0.000	0.002	0.000
MT	0.286	0.153	0.360	0.000	0.000	0.000

Table 3: TMS and cDFS at horizon

L=32

. Within the baseline group, we highlight the best , the second best , and the third best methods. BaryBoosted methods are highlighted when they would rank first/second/third after being added to the pool. Red cells mark cDFS

<0.07

(easily detected).

Figure 2: Trade-off between TMS and cDFS for each dataset on horizon

L=32

. The solid line shows the Pareto front over baseline methods.

Method	R1 (amount) ↑			R1 (time) ↑			F1 (mcc) ↑
	Alphabattle	Gender	Age	Alphabattle	Gender	Age	Alphabattle	Gender	Age
LDM	0.479	0.180	0.216	0.148	0.144	0.074	0.400	0.364	0.323
LDM-WB (ours)	0.583	0.334	0.438	0.337	0.331	0.331	0.459	0.448	0.371

Table 4: BaryBoosting improves prediction metrics beyond the optimized TMS (

L=32

; excerpt of Table 9 ).

Appendix figures & tables21 assets

Supplementary material from the paper’s appendix.

Appendix

Method	GRU 128-1L	GRU 512-1L	GRU 512-2L (default)	GRU 1024-3L	Tr. 128-1L	Tr. 512-1L	Tr. 512-2L	Tr. 1024-3L
GT	1.000	1.000	1.000	0.996	1.000	1.000	1.000	1.000
Hist.	0.198	0.234	0.234	0.225	0.727	0.756	0.727	0.831
Mode	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000
Rep.	0.031	0.037	0.065	0.705	0.008	0.003	0.002	0.334
AR	0.000	0.000	0.000	0.000	0.000	0.000	0.001	0.007
MT	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.069

Appendix

Table 5: Age : cDFS under eight discriminator configurations, horizon

L=32

. Spearman

\rho

compares the model ranking under each configuration with the ranking under GRU-512-2L.

Method	GRU 128-1L	GRU 512-1L	GRU 512-2L (default)	GRU 1024-3L	Tr. 128-1L	Tr. 512-1L	Tr. 512-2L	Tr. 1024-3L
GT	1.000	1.000	1.000	1.000	1.000	1.000	1.000	1.000
Hist.	0.427	0.493	0.477	0.470	0.612	0.673	0.535	0.871
Mode	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000
Rep.	0.177	0.067	0.050	0.054	0.007	0.008	0.001	0.197
AR	0.001	0.002	0.002	0.001	0.001	0.000	0.000	0.001
MT	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000

Appendix

Table 6: Gender : cDFS under eight discriminator configurations, horizon

L=32

. Spearman

\rho

compares the model ranking under each configuration with the ranking under GRU-512-2L.

Method	GRU 128-1L	GRU 512-1L	GRU 512-2L (default)	GRU 1024-3L	Tr. 128-1L	Tr. 512-1L	Tr. 512-2L	Tr. 1024-3L
GT	1.000	1.000	1.000	1.000	1.000	1.000	1.000	1.000
Hist.	0.573	0.556	0.544	0.550	0.338	0.313	0.147	0.874
Mode	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000
Rep.	0.015	0.011	0.009	0.006	0.000	0.002	0.000	0.494
AR	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000
MT	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000

Appendix

Table 7: Alphabattle : cDFS under eight discriminator configurations, horizon

L=32

. Spearman

\rho

compares the model ranking under each configuration with the ranking under GRU-512-2L.

	cDFS $\uparrow$			Shape $\uparrow$			Trend $\uparrow$
Method	Age	Gender	Alph.	Age	Gender	Alph.	Age	Gender	Alph.
Hist.	0.234	0.476	0.544	0.987	0.962	0.984	0.961	0.785	0.949
Mode	0.000	0.000	0.000	0.792	0.806	0.801	0.402	0.700	0.567
Rep.	0.065	0.050	0.009	0.994	0.978	0.991	0.973	0.879	0.955
AR	0.000	0.002	0.000	0.852	0.851	0.840	0.538	0.642	0.620
MT	0.000	0.000	0.000	0.853	0.811	0.855	0.401	0.272	0.712

Appendix

Table 8: Shape and Trend scores alongside cDFS at horizon

L=32

on the three transaction datasets. Bold marks cDFS

>0.07

and the corresponding Shape and Trend cells.

Figure 3: Relationship between cDFS and the auxiliary fidelity scores — Shape score (left) and Trend score (right) — across all three datasets pooled. Horizon $L=32$ . Higher is better on both axes. The figure shows that high Shape and Trend scores do not, by themselves, imply a high cDFS: many baselines achieve competitive Shape or Trend values while collapsing along the cDFS axis, suggesting that cDFS is a more holistic criterion.

Method	TMS ↑			R1 (amount) ↑			R1 (time) ↑			F1 (mcc) ↑
	Alphabattle	Gender	Age	Alphabattle	Gender	Age	Alphabattle	Gender	Age	Alphabattle	Gender	Age
Hist.	0.382	0.221	0.366	0.498	0.117	0.245	-0.049	-0.325	-0.179	0.424	0.397	0.331
Mode	0.203	0.047	0.263	0.000	0.000	0.000	0.000	0.000	0.000	0.081	0.088	0.052
Rep.	0.440	0.361	0.454	0.524	0.182	0.236	0.147	0.137	0.128	0.444	0.425	0.374
AR	0.286	0.188	0.354	0.043	-0.018	0.082	0.064	0.096	0.183	0.287	0.324	0.208
MT	0.286	0.153	0.360	-0.000	0.016	0.020	-0.042	-0.043	0.185	0.369	0.343	0.270

Appendix

Table 9: TMS, R1 (amount), R1 (time), F1 (mcc) at horizon

L=32

. Across all shown methods, the best is in bold , the second best is underlined and the third best is dash-underlined .

Figure 4: Relationship between TMS and matched per-feature submetrics at horizon $L=32$ : matched $R_{1}$ for amount (left), matched $R_{1}$ for time (middle), and matched $F_{1}$ for MCC (right). Higher is better on both axes. Across all three views, TMS grows consistently with each matched submetric.

Figure 5: Trade-off between TMS and cDFS for each dataset on horizon $L=4$ . The solid line shows the Pareto front over baseline methods.

Figure 6: Trade-off between TMS and cDFS for each dataset on horizon $L=16$ . The solid line shows the Pareto front over baseline methods.

Figure 7: Trade-off between TMS and cDFS for each dataset on horizon $L=64$ . The solid line shows the Pareto front over baseline methods.

Method	TMS ↑		cDFS ↑
	3-seed	reported	3-seed	reported
LDM	0.430 $\pm$ 0.002	0.426	0.983 $\pm$ 0.006	1.000
DEF	0.505 $\pm$ 0.004	0.508	0.001 $\pm$ 0.000	0.000
LDM-WB	0.523 $\pm$ 0.001	0.522	0.202 $\pm$ 0.005	0.224

Appendix

Table 10: Age, horizon

L=32

: 3-seed mean

\pm

std vs. reported in main results (single reported values shown without uncertainty). Higher is better. Within each column, the best is in bold , second best is underlined

Figure 8: Generated continuations for one Age test client at horizon

L=32

. Each event is a stem at its timestamp; height is the log-transformed amount and color encodes the category (the nine most frequent categories receive their own color, the rest share grey). The vertical rule separates the observed history, identical across all panels, from the 32 generated events. Original is the ground-truth continuation.

Method	TMS ↑	cDFS ↑
ADiff4TPP	0.402	0.997
LDM	0.426	1.000
ADiff4TPP-WB	0.512	0.108
LDM-WB	0.522	0.224

Appendix

Table 11: ADiff4TPP and LDM at horizon

L=32

on Age dataset, with and without Wasserstein BaryBooster.

	Age		Gender		Alphabattle
$N$	TMS $\uparrow$	cDFS $\uparrow$	TMS $\uparrow$	cDFS $\uparrow$	TMS $\uparrow$	cDFS $\uparrow$
1	0.428	0.966	0.319	0.874	0.404	0.855
5	0.499	0.257	0.407	0.137	0.472	0.133
10	0.510	0.271	0.425	0.300	0.485	0.210
15	0.517	0.229	0.430	0.271	0.491	0.188
30	0.522	0.224	0.436	0.258	0.495	0.177

Appendix

Table 12: Wasserstein BaryBooster on LDM as a function of the sample size

N

at horizon

L=32

;

N{=}1

is the base LDM without boosting.

Figure 9: TMS change at horizon

L=32

after applying Wasserstein BaryBooster to different base models. The largest improvement is observed for diffusion-based models, while DEF shows no visible gain and the remaining methods improve only slightly.

Aspect	ADiff4TPP [ Mukherjee et al., 2026 ]	Our LDM
Diffusion formulation	Flow matching	EDM
Backbone	DiT	DiT
Data–latent mapping	$\beta$ -VAE	$\beta$ -VAE

Appendix

Table 13: High-level comparison between ADiff4TPP and our latent diffusion model (LDM).

Panel of Figure 10 — Figure 10: Schematic illustration of the considered methods.

Dataset	Model	Train (min)	Inference (min)	Memory (GB)
Age	LDM	525.6	1.2	1.71
Age	DDM	776.9	8.0	1.62
Age	DEF	281.0	0.13	71.12
Age	MT	111.9	0.08	3.94
Age	AR	83.1	2.4	7.43
Age	LDM-WB ( $N{=}30$ )	–	52.4	1.71

Appendix

Table 14: Runtime and memory summary for Age, Gender, and Alphabattle at horizon

L=64

Dataset	Horizon	Total (s)	Sampling	Medoid	Matching	Aggregation
Age	32	310.0	264.9 (85%)	17.8	26.2	0.3
Gender	32	307.2	261.8 (85%)	18.0	26.4	0.3
Alphabattle	32	349.9	301.6 (86%)	18.2	29.0	0.3
Age	64	788.7	484.8 (61%)	108.0	194.8	0.2
Gender	64	786.3	480.6 (61%)	107.5	197.3	0.2
Alphabattle	64	861.5	547.2 (64%)	105.6	207.7	0.3

Appendix

Table 15: Decomposition of the BaryBooster wall time into base-model sampling and the stages of the aggregation algorithm. Seconds per 10 test batches,

N=30

Dataset	Reported by	# seq.	Mean len.	Reason
Our benchmark (Table 2 )
Gender	ours	14k (11k/2.8k)	404	we use
Age	ours	30k (24k/6k)	881	we use
Alphabattle	ours (Alfa-Bank, official test split)	135k (100k/35k)	367	we use
Generative-modeling papers for event sequences
Retweet	[ Xue et al., 2024 , Karpukhin et al., 2026 , Zeng et al., 2024 , Mukherjee et al., 2026 ]	24k	41–56	L

Appendix

Table 16: Public event-sequence datasets considered and reasons for exclusion.

Method	TMS $\uparrow$			cDFS $\uparrow$
	Age	Retweet	Amazon	Age	Retweet	Amazon
Hist.	0.366	0.391	0.343	0.234	0.035	0.082
Mode	0.263	0.156	0.036	0.000	0.009	0.002
Rep.	0.454	0.435	0.415	0.065	0.053	0.070
AR	0.354	0.530	0.151	0.000	0.041	0.110
MT	0.360	0.604	0.196	0.000	0.441	0.140

Appendix

Table 17: TMS and cDFS at horizon

L=32

on the two TPP corpora, with Age for reference. Within each column: best , second , third .

Conditional Generative Modeling

Conditional generative modeling learns to produce data distributions steered by labels, inputs, text, or physical and structural constraints.

See all 395 papers

Probabilistic Forecasting of Business Process Executions with Neural Temporal Point Processes
Sep 18, 2026Jiaxin Yuan, Daniela Grigori, Han van der AaProbabilistic Forecasting Predictive Process Monitoring
Paris Dauphine University-PSL, Pl. du Maréchal de Lattre de Tassigny, 75016 Paris, France · University of Vienna, Universitätsring 1, 1010 Wien, Austria
SurF: A Generative Model for Multivariate Irregular Time Series Forecasting
May 13, 2026Mohammad R. Rezaei, Tejas Balaji, Rahul G. KrishnanIrregular Time-Series Modeling Multivariate Time Series Forecasting
Department of Computer Science University of Toronto Vector Institute Toronto, ON, Canada
LEAF: A Living Benchmark for Event-Augmented Forecasting
May 9, 2026Mingtian Tan, Mihir Parmar, Palash Goyal +5Financial Forecasting Time Series Forecasting
Google · University of Virginia

Sep 18, 2026J/K move · Enter open · S save

Probabilistic Forecasting of Business Process Executions with Neural Temporal Point Processes

Jiaxin Yuan, Daniela Grigori, Han van der Aa

Paris Dauphine University-PSL, Pl. du Maréchal de Lattre de Tassigny, 75016 Paris, France · University of Vienna, Universitätsring 1, 1010 Wien, Austria

Probabilistic Forecasting Predictive Process Monitoring Next-Event Prediction Temporal Point Processes Process Mining

Open paper AskPDF arXiv

Operators of service-based systems act on forecasts of how a running execution will continue, and such a forecast is actionable only if its reliability is known. Mainstream deep-learning models for this task are discriminative and deterministic: they emit a single next activity and a single remaining-time estimate, without a distribution to reason over. We instead cast the problem as generative sequence modelling with marked temporal point processes, which define a joint density over the next mark and its inter-event time and therefore deliver predictive distributions by construction. Real event logs violate the simple-point-process assumption these models rest on, since consecutive events frequently carry identical timestamps; we handle such ties explicitly and combine a transformer encoder with a mixture decoder over inter-event times, trained by exact log-likelihood. On ten public logs, the resulting model matches discriminative baselines on point accuracy, dominates them on the calibration and sharpness of remaining-time distributions, and is the cheapest at inference, since a full predictive distribution is obtained in a single forward pass without sampling.

Abstract

Conditional Generative Modeling

Explore similar work

Probabilistic Forecasting of Business Process Executions with Neural Temporal Point Processes

SurF: A Generative Model for Multivariate Irregular Time Series Forecasting

LEAF: A Living Benchmark for Event-Augmented Forecasting