Diffusion models have become a widely used framework for probabilistic time series forecasting, modeling the distribution of future values given an observed history. In time series forecasting, however, the future continues the observed history, creating an asymmetry the standard diffusion process leaves unaddressed, with slowly-varying content largely determined by the observed continuity while higher-frequency dynamics carry most of the residual uncertainty. Existing diffusion-based forecasters decouple this asymmetry through an external rule before generation, leaving the corruption trajectory blind to which parts of the target the history can already anchor. We propose DiffDiff, a diffusion framework that embeds this predictability asymmetry into the diffusion trajectory itself, so that a single end-to-end diffusion process becomes aware of which parts of the target the history can already anchor. DiffDiff makes the forward operator step-dependent so that the noisy intermediate state progressively shifts from the target itself toward its second-order differenced structure, while a conditioning pathway supplies the denoiser with both value-domain and differential history information balanced by a stage-adaptive gate at each diffusion step. The terminal distribution approaches a standard Gaussian, preserving compatibility with existing samplers. On seven benchmarks across four prediction horizons, DiffDiff outperforms six diffusion baselines, while our analyses show that its trajectory reallocates modeling effort according to history predictability and improves distributional fidelity.
Figures & tables
Figure 1: Overall architecture of DiffDiff . Top : the differencing forward process gradually shifts the clean target x0 toward its differenced structure under the step-dependent operator At , and the deterministic reverse step recovers xs from the denoiser prediction x^0 . Bottom : the stage-adaptive conditional denoiser encodes both the value-domain history xhist and its first-order differences Δxhist , fuses them through a sigmoid gate conditioned on the diffusion timestep, and outputs the value-domain prediction along with auxiliary statistics for instance denormalization.
Model
H
Electricity
ETTm2
Exchange
Traffic
Weather
Solar
Wind
Wins
MSE
MAE
CRPS
MSE
MAE
CRPS
MSE
MAE
CRPS
MSE
MAE
CRPS
MSE
MAE
CRPS
MSE
MAE
CRPS
MSE
MAE
CRPS
Standard
96
.371
.442
.193
.134
.285
.122
.103
.251
.110
.221
.320
.149
.125
.268
.116
.650
.547
.248
.645
.567
.246
0
192
.403
.467
.195
.153
.303
.123
.210
.365
.157
.208
.305
.142
.188
.343
.144
.642
.564
.246
.892
.696
.301
0
336
.951
.560
.206
.207
.361
.147
.656
.626
.261
.215
.315
.147
.256
.397
.168
.515
.498
.209
.980
.753
.328
0
720
.565
.561
.223
.217
.368
.151
1.024
.859
.407
.351
.431
.180
.500
.593
.254
1.362
.823
.348
1.249
.890
.361
0
Diff-F
96
.281
.369
.152
.067
.189
.078
.098
.239
.099
.157
.233
.107
.093
.222
.095
.490
.447
.197
.637
.563
.234
10
Table 1: Ablation results on seven benchmarks across four prediction horizons. For every column the best value across the four methods is bold and the second-best is underlined . Standard uses an isotropic forward with a plain denoiser. The other three rows replace the forward only (Diff-forward), the denoiser only (Std+G), or both ( DiffDiff ). The rightmost column reports per-row wins.
Model
H
Electricity
ETTm2
Exchange
Traffic
Weather
Solar
Wind
Wins
MSE
MAE
CRPS
MSE
MAE
CRPS
MSE
MAE
CRPS
MSE
MAE
CRPS
MSE
MAE
CRPS
MSE
MAE
CRPS
MSE
MAE
CRPS
CSDI
96
1.18
.853
.332
.444
.535
.257
.377
.505
.223
1.96
1.20
.453
.479
.568
.233
1.38
.987
.343
.968
.786
.296
0
192
7.77
2.24
.892
2.41
1.29
.539
.753
.715
.277
1.99
1.21
.456
3.64
1.59
.639
2.63
1.30
.553
3.50
1.52
.640
0
336
13.0
1.69
2.37
7.27
2.38
1.06
3.84
1.65
.592
1.98
1.21
.457
5.24
2.05
.805
3.30
1.52
.644
9.74
2.61
1.09
0
720
1.03
.814
.327
6.61
2.25
1.12
12.1
2.73
.902
2.00
1.21
.457
14.0
3.42
1.57
3.52
1.48
.633
13.0
3.20
1.43
0
TMDM
96
.315
.412
.168
.072
.200
.081
.111
.255
.102
.205
.272
.117
.123
.257
.105
.359
.346
.136
.663
.560
.231
4
Table 2: Forecasting results on seven benchmarks across four prediction horizons. For every column the best value across methods is bold and the second-best is underlined . The rightmost column reports per-row wins.
Figure 2: Reshaping of the corruption trajectory by the differencing forward process. Top : Per-frequency power spectral density (PSD) ratio of DiffDiff to the standard isotropic forward at three diffusion progresses on electricity , with values above (below) unity marking enhanced (attenuated) bands. Bottom : Retained energy fraction of the window mean and the first- and second-order increments under each forward path along the diffusion progress on solar .
Figure 3: Component-level corruption alignment and gate dynamics. Left: Per-component coefficient of determination R2 (horizontal) versus corruption aggressiveness (vertical) on wind and exchange-rate , with each component reducing a window to a scalar, namely the window mean (level), summed fast Fourier transform (FFT) magnitudes in three frequency bands (low, mid, high), and the mean first-order absolute difference (diff1). Components are traced across five diffusion progresses, with open markers for the standard forward and filled markers for DiffDiff , jittered horizontally for visibility. Right: Gate activation of the differential branch versus diffusion timestep across seven datasets, with shaded bands showing ±1 standard deviation.
Figure 4: Spectral, local-change, and marginal fidelity of generated forecasts. Left: root mean square error (RMSE) between log power spectral densities (PSDs). Middle: normalized second-order-difference Wasserstein distance on a logarithmic horizontal axis. Right: marginal quantile–quantile (QQ) RMSE. Lower is better for all three metrics. Values are medians computed over all seven datasets and four prediction horizons.
Figure 5: Per-window 95% prediction interval of DiffDiff and three diffusion baselines on illustrative windows from electricity (top) and exchange rate (bottom) at H=96 , each selected as a top-CRPS-improvement window of DiffDiff over the strongest baseline on that dataset. The gray line shows the history, the black line the ground truth, the colored line the median forecast, and the shaded band the 95% prediction interval.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Frequency
Train
Val
Test
Electricity
1 hour
18,317
2,633
5,261
ETTm2
15 min
34,465
11,521
11,521
Exchange
1 day
5,120
665
1,422
Traffic
1 hour
12,185
1,757
3,509
Weather
10 min
36,792
5,271
10,540
Solar
10 min
36,601
5,161
10,417
Appendix
Table 3: Dataset statistics.
Model
Electricity
ETTm2
Exchange
Traffic
Weather
Solar
Wind
Overall
TQNet
.376/.436
.120/.256
.552/.493
.151/.227
.189/.307
.881/.603
1.107/.762
.482/.440
Times2D
.362/.428
.123/.259
.496/.492
.171/.258
.190/.308
.936/.623
1.098/.762
.482/.447
PathFormer
.348/.417
.119/.253
.615/.522
.162/.238
.190/.309
1.028/.619
1.074/.744
.505/.443
DiffDiff
.338/.407
.120/.258
.356/.457
.156/.237
.190/.315
.455/.433
.823/.687
.348/.399
Appendix
Table 4: Point-forecasting comparison with deterministic forecasters using four-horizon mean MSE and MAE under the shared L=96 protocol. Lower is better.
Difficulty measure
Mean Spearman ρ
95% CI
Positive settings
Q4/Q1
TQNet absolute error
.1417±.0035
[.0851, .2077]
26/28
1.6427±.0455
Times2D absolute error
.1282±.0038
[.0749, .1919]
26/28
1.5841±.0378
Matched-history dispersion ( q50 )
.1304
[.0393, .2310]
14/24
1.8845
Matched-history dispersion ( q75 )
.1331
[.0504, .2350]
22/24
1.5355
Appendix
Table 5: Association between local curvature and forecasting difficulty. Predictor statistics report mean ± sample standard deviation over five seeds.
Variant
H
Electricity
ETTm2
Exchange
Traffic
Weather
Solar
Wind
Wins
MSE
MAE
CRPS
MSE
MAE
CRPS
MSE
MAE
CRPS
MSE
MAE
CRPS
MSE
MAE
CRPS
MSE
MAE
CRPS
MSE
MAE
CRPS
DiffDiff
96
.269
.362
.150
.068
.190
.079
.101
.246
.103
.156
.234
.108
.093
.222
.093
.463
.435
.191
.599
.547
.226
15
192
.299
.381
.158
.102
.239
.098
.198
.358
.147
.151
.231
.106
.135
.271
.112
.483
.448
.191
.815
.675
.278
11
336
.349
.413
.171
.128
.273
.110
.388
.495
.207
.148
.232
.106
.202
.332
.137
.389
.395
.163
.912
.742
.300
15
720
.435
.474
.196
.182
.330
.133
.735
.728
.327
.168
.252
.114
.328
.434
.178
.485
.455
.189
.966
.785
.317
13
CrossAttn
96
.284
.372
.154
.068
.191
.079
.096
.238
.098
.158
.236
.110
.094
.224
.095
.467
.436
.192
.611
.548
.226
4
Appendix
Table 6: Conditioning ablation. Within each (dataset, metric) column the best value across variants is in bold and the second-best is underlined . The rightmost column reports per-row wins across the 21 (dataset, metric) settings.
Order
H
Electricity
ETTm2
Exchange
Traffic
Weather
Solar
Wind
Wins
MSE
MAE
CRPS
MSE
MAE
CRPS
MSE
MAE
CRPS
MSE
MAE
CRPS
MSE
MAE
CRPS
MSE
MAE
CRPS
MSE
MAE
CRPS
D1
96
.283
.370
.154
.067
.189
.079
.101
.246
.103
.157
.235
.108
.094
.224
.095
.465
.437
.191
.612
.552
.228
7
192
.319
.391
.161
.100
.237
.097
.197
.348
.142
.150
.229
.105
.137
.275
.114
.460
.438
.186
.864
.687
.280
4
336
.358
.419
.172
.128
.273
.111
.388
.495
.207
.147
.231
.105
.210
.339
.140
.418
.413
.170
.926
.745
.302
6
720
.456
.488
.203
.182
.329
.133
.735
.728
.327
.167
.252
.114
.339
.443
.182
.492
.454
.188
.978
.794
.321
7
D2 (Ours)
96
.269
.362
.150
.068
.190
.079
.101
.246
.103
.156
.234
.108
.093
.222
.093
.463
.435
.191
.599
.547
.226
18
Appendix
Table 7: Differencing order ablation. Dk denotes the order of the differencing operator in the forward process, with D2 as the default used by DiffDiff . Within each column the best value across orders is in bold . The rightmost column reports per-row wins.
Dataset
H
MSE ↓
CRPS ↓
0.1
0.3
0.5
0.7
0.9
0.1
0.3
0.5
0.7
0.9
ETTm2
96
.071
.070
.068
.069
.068
.079
.079
.079
.081
.081
192
.100
.102
.102
.101
.100
.096
.098
.098
.098
.098
336
.131
.133
.128
.131
.132
.111
.112
.110
.112
.113
720
.201
.201
.182
.181
.182
.140
.140
.133
.132
.132
Exchange
96
.108
.104
.101
.111
.105
.105
.101
.103
.107
.107
Appendix
Table 8: Sensitivity to λmax .
Dataset
H
MSE ↓
CRPS ↓
2
12
24
48
72
96
2
12
24
48
72
96
ETTm2
96
.068
.067
.068
.068
.067
.067
.082
.080
.080
.079
.077
.077
192
.103
.102
.099
.102
.100
.099
.099
.099
.097
.098
.097
.097
336
.146
.142
.144
.128
.139
.143
.120
.117
.119
.110
.115
.117
720
.190
.186
.190
.182
.186
.193
.137
.137
.137
.133
.134
.137
Exchange
96
.100
.113
.091
.101
.093
.101
.107
.112
.099
.103
.098
.100
Appendix
Table 9: Sensitivity to label window length Ll .
Dataset
H
MSE ↓
CRPS ↓
1
3
5
10
1
3
5
10
ETTm2
96
.068
.068
.068
.068
.079
.079
.079
.079
192
.102
.102
.102
.103
.098
.098
.098
.098
336
.128
.131
.128
.133
.110
.112
.110
.112
720
.183
.184
.182
.184
.133
.134
.133
.134
Exchange
96
.112
.108
.101
.103
.106
.104
.103
.103
Appendix
Table 10: Sensitivity to future loss weight wfut .
Dataset
H
MSE ↓
CRPS ↓
5
10
20
50
100
5
10
20
50
100
ETTm2
96
.068
.068
.068
.069
.069
.080
.079
.079
.079
.079
192
.102
.102
.101
.102
.102
.099
.098
.097
.097
.097
336
.128
.128
.128
.129
.130
.112
.110
.110
.110
.110
720
.182
.182
.182
.183
.183
.134
.133
.132
.132
.132
Exchange
96
.102
.101
.108
.111
.112
.101
.103
.104
.105
.106
Appendix
Table 11: Sensitivity to DDIM sampling steps S .
Dataset
H
p (vs. MA-TSD)
p (vs. NsDiff)
p (vs. TMDM)
Electricity
96
1.2e−18
3.0e−05
4.0e−03
192
5.7e−14
1.7e−08
4.8e−18
336
2.9e−50
2.3e−20
1.6e−18
720
1.5e−86
6.9e−01
7.3e−01
ETTm2
96
<10−300
1.3e−15
6.8e−02
192
3.2e−284
3.5e−15
1.7e−04
Appendix
Table 12: Paired Wilcoxon signed-rank test on per-window CRPS.
CSDI
TMDM
TimeDiff
NsDiff
MA-TSD
DiffDiff
Mem.Train (MB)
181.33
4015.88
227.30
779.38
31.96
38.57
Mem.Inference (MB)
59.38
293.15
88.99
269.07
150.69
212.81
Tim.Train (ms)
43.12
46.02
22.08
48.82
10.19
12.95
Tim.Inference (ms)
21284.91
10543.89
12671.42
3908.74
1705.27
1462.65
CRPS
1.12
.150
.182
.159
.145
.133
Appendix
Table 13: Computation efficiency comparison. Bold face indicates the best result per row.
Time series forecasting predicts future values from ordered observations using statistical, neural, and probabilistic models to support decisions under uncertainty.
Diffusion models offer a natural way to model uncertainty in time series forecasting, yet their iterative sampling process is often treated as a uniformly beneficial refinement procedure. Our study challenges this view by examining how forecast quality evolves throughout reverse diffusion. We find that general temporal structure is often recovered at relatively high noise levels, whereas continued low-noise refinement can introduce statistical drift and degrade the final forecast. Our analysis further suggests that this behavior explains why prior methods often favor relatively narrow diffusion architecture and schedule design. Building on this observation, we propose a label-free global stopping criterion that detects the optimal termination point, eventually speeding up inference and improving predictive accuracy. Additionally, since early stopping terminates inference in high-noise regions, we propose a Bernoulli timestep sampler that concentrates training on this region while preserving coverage of the full diffusion process. Extensive experiments conducted across eight real-world datasets demonstrate the superior performance of our method compared to existing approaches.
Dat Nguyen-Cong, Luong Tran, Tung Kieu
FPT Software AI Center, FPT Corporation · Department of Computer Science, Aalborg University, Denmark
Probabilistic time series forecasting requires modeling and predicting complex and time-varying distributions. Recently, Denoising Diffusion Probabilistic Model (DDPM)-based approaches have shown promise by equipping the dif- fusion process with pretrained mean and variance estimators to accommodate distributional shift. However, these methods typically follow the standard DDPM framework and consider only partial components of the evidence lower bound (ELBO), treating the training of estimators as designed regression tasks separate from the variational inference framework. To address this, we rethink the ELBO under the Location-Scale Noise Model (LSNM) and find that it naturally induces a Gaussian negative log likelihood objective for the estimators and inherently defines a joint training objective that unifies recent diffusion paradigms for probabilistic forecasting. Building on this principled ELBO reformulation, we propose Diff- PTS, a general framework that enables end-to-end optimization of all components within the ELBO. Across multiple benchmarks, DiffPTS consistently outperforms recent models, achieving state-of-the-art performance with an average CRPS/MSE reduction of over 14.53%/16.55% compared to existing diffusion-based methods. The code is available at https://github.com/wwy155/DiffPTS.
Weiwei Ye, Dongyuan Li, Hangchen Liu +3
The University of Tokyo · Harbin Institute of Technology, Shenzhen
Diffusion models have recently shown strong potential for probabilistic multivariate time-series forecasting by modeling complex conditional distributions. Recent decoupled diffusion frameworks further separate forecasting into deterministic prediction and stochastic residual generation, making it natural to derive dependency graphs from deterministic representations and use them to guide residual diffusion. However, we show that this direct structural transfer is unreliable. Although deterministic-derived graphs encode useful global dependency priors, they exhibit substantial edge-level misalignment with residual dependency structures, introducing inaccurate or redundant conditions during residual generation. This reveals a previously overlooked deterministic-to-residual structural alignment problem in decoupled diffusion forecasting. To address this problem, we propose GARDiff, a Graph-Aligned Residual Diffusion framework for probabilistic multivariate time-series forecasting. Instead of treating deterministic-derived graphs as fixed diffusion conditions, GARDiff progressively adapts them to residual generation. Specifically, GARDiff estimates residual uncertainty to distinguish high- and low-uncertainty regions, enabling uncertainty-aware structural refinement, and further performs timestep-aware edge sparsification during reverse diffusion to evolve graph conditions from broad dependency aggregation to localized residual refinement. Extensive experiments on six real-world benchmarks demonstrate that GARDiff consistently improves probabilistic forecasting performance and uncertainty calibration over strong baselines.
Rui Han, Min Yang, Xu Zhang +3
Shandong University, Jinan, China · Macquarie University, Sydney, Australia · China University of Petroleum (East China), Qingdao, China +1