List your company
ResearchFormerly WizWand

Multivariate Long-term Time Series Forecasting on Traffic (MSE, MAE)

0.353MSES_Mamba
0.343240.409120.4750.54088Oct 2, 2024Jan 16, 2025May 2, 2025Aug 16, 2025Nov 30, 2025Mar 16, 2026Jun 30, 2026
Updated 4mo ago

Evaluation results

MethodLinks
S_MambaPrediction Length=96, Backbone=S_Mamba2026.030.3530.253
S_MambaPrediction Length=192, Backbone=S_Mamba2026.030.3540.261
S_Mamba + TimeAPNPrediction Length=96, Backbone=S_Mamba2026.030.360.254
PatchTST + TimeAPNPrediction Length=96, Backbone=PatchTST2026.030.3620.258
TA-SparseMGForecast horizon (T)=96, Look-back length (L)=7202026.060.3620.243
PatchTSTHorizon=96, Parameter(720)=6.31M2024.100.366—
S_MambaPrediction Length=336, Backbone=S_Mamba2026.030.3680.264
Time TrackerModeling Type=Channel-Mixing (CM), Input Length (L)=96, Predicting Length (F)=962025.050.370.249
S_Mamba + TimeAPNPrediction Length=192, Backbone=S_Mamba2026.030.370.264
PatchTSTPrediction Length=96, Backbone=PatchTST2026.030.3720.259
TALONEvaluation Protocol=one-for-all, Lookback Window (L)=672, Training Horizon=96-step2026.060.3730.253
DLinear + TimeAPNPrediction Length=96, Backbone=DLinear2026.030.3740.262
PatchTST + TimeAPNPrediction Length=192, Backbone=PatchTST2026.030.3750.265
Timer XLEvaluation Protocol=one-for-all, Lookback Window (L)=672, Training Horizon=96-step2026.060.3780.256
AutoTimesEvaluation Protocol=one-for-all, Lookback Window (L)=672, Training Horizon=96-step2026.060.3790.265
InA-ProbeEvaluation Protocol=one-for-all, Lookback Window (L)=672, Training Horizon=96-step2026.060.3830.262
InA-ProbeEvaluation Protocol=One-for-all, Prediction Length/Horizon Training Strategy=Single 96-step model serves all horizons2026.060.3830.262
PatchTSTPrediction Length=192, Backbone=PatchTST2026.030.3840.263
iTransformerEvaluation Protocol=one-for-all, Lookback Window (L)=672, Training Horizon=96-step2026.060.3840.272
CALFEvaluation Protocol=One-for-one, Prediction Length/Horizon Training Strategy=Trained respectively for each prediction length2026.060.3860.265
iTransformerEvaluation Protocol=One-for-one, Prediction Length/Horizon Training Strategy=Trained respectively for each prediction length2026.060.3860.275
PatchTSTHorizon=192, Parameter(720)=6.31M2024.100.388—
PENGUINLook-back length (L)=336, Prediction horizons (H)={96, 192, 336, 720}2025.080.3880.262
MixLinearHorizon=96, Parameter(720)=0.176K2024.100.389—
SparseTSFHorizon=96, Parameter(720)=0.925K2024.100.389—
S_Mamba + TimeAPNPrediction Length=336, Backbone=S_Mamba2026.030.390.278
iTransformerLook-back length (L)=336, Prediction horizons (H)={96, 192, 336, 720}2025.080.390.275
iTransformerModeling Type=Channel-Mixing (CM), Input Length (L)=96, Predicting Length (F)=962025.050.3920.263
DLinear + TimeAPNPrediction Length=192, Backbone=DLinear2026.030.3930.272
iTransformerHorizon=96, Parameter(720)=7.04M2024.100.395—
PatchTST + TimeAPNPrediction Length=336, Backbone=PatchTST2026.030.3950.278
PatchTSTEvaluation Protocol=One-for-one, Prediction Length/Horizon Training Strategy=Trained respectively for each prediction length2026.060.3970.279
FITSHorizon=96, Parameter(720)=41.76K2024.100.398—
SparseTSFHorizon=192, Parameter(720)=0.925K2024.100.398—
PatchTSTHorizon=336, Parameter(720)=6.31M2024.100.398—
PatchTSTPrediction Length=336, Backbone=PatchTST2026.030.3980.27
TimeLLMEvaluation Protocol=One-for-one, Prediction Length/Horizon Training Strategy=Trained respectively for each prediction length2026.060.4020.284
MixLinearHorizon=192, Parameter(720)=0.176K2024.100.403—
SEEDInput length (L)=96, Forecasting horizons (T)={96, 192, 336, 720}2025.120.4040.265
TimeMixerEvaluation Protocol=One-for-one, Prediction Length/Horizon Training Strategy=Trained respectively for each prediction length2026.060.4040.285
PatchTSTLook-back length (L)=336, Prediction horizons (H)={96, 192, 336, 720}2025.080.4050.282
TimerModeling Type=Channel-Independent (CI), Input Length (L)=96, Predicting Length (F)=962025.050.4070.26
PatchTSTEvaluation Protocol=one-for-all, Lookback Window (L)=672, Training Horizon=96-step2026.060.4080.298
FPTEvaluation Protocol=One-for-one, Prediction Length/Horizon Training Strategy=Trained respectively for each prediction length2026.060.4080.288
SOFTSInput length (L)=96, Forecasting horizons (T)={96, 192, 336, 720}2025.120.4090.267
FITSHorizon=192, Parameter(720)=41.76K2024.100.409—
DLinear + TimeAPNPrediction Length=336, Backbone=DLinear2026.030.4090.28
SparseTSFHorizon=336, Parameter(720)=0.925K2024.100.411—
CATSLook-back length (L)=336, Prediction horizons (H)={96, 192, 336, 720}2025.080.4110.276
DLinearHorizon=96, Parameter(720)=485.3K2024.100.413—
DLinearPrediction Length=96, Backbone=DLinear2026.030.4130.287
TimeLLMEvaluation Protocol=one-for-all, Lookback Window (L)=672, Training Horizon=96-step2026.060.4140.305
MixLinearHorizon=336, Parameter(720)=0.176K2024.100.416—
iTransformerHorizon=192, Parameter(720)=7.04M2024.100.417—
S_MambaPrediction Length=720, Backbone=S_Mamba2026.030.420.288
FITSHorizon=336, Parameter(720)=41.76K2024.100.421—
DLinearHorizon=192, Parameter(720)=485.3K2024.100.423—
CycleNetLook-back length (L)=336, Prediction horizons (H)={96, 192, 336, 720}2025.080.4230.288
DLinearPrediction Length=192, Backbone=DLinear2026.030.4240.29
iTransformerInput length (L)=96, Forecasting horizons (T)={96, 192, 336, 720}2025.120.4280.282
FPTEvaluation Protocol=one-for-all, Lookback Window (L)=672, Training Horizon=96-step2026.060.4280.312
iTrans.Horizon=Avg, Input length=962026.060.4280.282
SimpleTMForecast horizon (T)=720, Look-back length (L)=7202026.060.430.263
PatchTST + TimeAPNPrediction Length=720, Backbone=PatchTST2026.030.4320.298
iTransformerHorizon=336, Parameter(720)=7.04M2024.100.433—
DLinearLook-back length (L)=336, Prediction horizons (H)={96, 192, 336, 720}2025.080.4340.295
MoLELook-back length (L)=336, Prediction horizons (H)={96, 192, 336, 720}2025.080.4340.295
MTLinearLook-back length (L)=336, Prediction horizons (H)={96, 192, 336, 720}2025.080.4340.295
PatchTSTPrediction Length=720, Backbone=PatchTST2026.030.4350.29
SparseTSFLook-back length (L)=336, Prediction horizons (H)={96, 192, 336, 720}2025.080.4360.29
SimpleTMEvaluation Protocol=one-for-all, Lookback Window (L)=672, Training Horizon=96-step2026.060.4360.317
DLinearHorizon=336, Parameter(720)=485.3K2024.100.437—
DLinearPrediction Length=336, Backbone=DLinear2026.030.4380.299
S_Mamba + TimeAPNPrediction Length=720, Backbone=S_Mamba2026.030.4380.302
TimeMixerEvaluation Protocol=one-for-all, Lookback Window (L)=672, Training Horizon=96-step2026.060.4420.321
PatchTSTModeling Type=Channel-Independent (CI), Input Length (L)=96, Predicting Length (F)=962025.050.4440.283
TQNetInput length (L)=96, Forecasting horizons (T)={96, 192, 336, 720}2025.120.4450.276
DLinear + TimeAPNPrediction Length=720, Backbone=DLinear2026.030.4470.3
SparseTSFHorizon=720, Parameter(720)=0.925K2024.100.448—
FEDformer + TimeAPNPrediction Length=96, Backbone=FEDformer2026.030.4490.283
DUETInput length (L)=96, Forecasting horizons (T)={96, 192, 336, 720}2025.120.4510.269
MixLinearHorizon=720, Parameter(720)=0.176K2024.100.452—
SimpleTMEvaluation Protocol=One-for-one, Prediction Length/Horizon Training Strategy=Trained respectively for each prediction length2026.060.4530.331
FITSHorizon=720, Parameter(720)=41.76K2024.100.457—
PatchTSTHorizon=720, Parameter(720)=6.31M2024.100.457—
DLinearHorizon=720, Parameter(720)=485.3K2024.100.466—
DLinearPrediction Length=720, Backbone=DLinear2026.030.4660.316
iTransformerHorizon=720, Parameter(720)=7.04M2024.100.467—
FEDformer + TimeAPNPrediction Length=192, Backbone=FEDformer2026.030.470.298
FEDformer + TimeAPNPrediction Length=336, Backbone=FEDformer2026.030.4760.316
PatchTSTHorizon=Avg, Input length=962026.060.4810.304
FEDformer + TimeAPNPrediction Length=720, Backbone=FEDformer2026.030.520.32
SCINetHorizon=192, Parameter(720)=23.62M2024.100.535—
SCINetHorizon=336, Parameter(720)=23.62M2024.100.54—
StateFlowHorizon=Avg, Input length=962026.060.5470.348
FEDformerHorizon=96, Parameter(720)=17.98M2024.100.573—
FEDformerPrediction Length=96, Backbone=FEDformer2026.030.5750.357
TimesNetHorizon=96, Parameter(720)=301.7M2024.100.593—
TimesNetModeling Type=Channel-Independent (CI), Input Length (L)=96, Predicting Length (F)=962025.050.5930.321
AutoformerModeling Type=Channel-Mixing (CM), Input Length (L)=96, Predicting Length (F)=962025.050.5970.371