Time Series Imputation on ETTh1
0.013MAEMSTN-Transformer
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| MSTN-TransformerMask R.=12.5%2025.11 | 0.013 | 0.002 | — | |
| MSTN-BiLSTMMask R.=12.5%2025.11 | 0.022 | 0.007 | — | |
| MSTN-TransformerMask R.=25%2025.11 | 0.025 | 0.004 | — | |
| MSTN-TransformerMask R.=Avg2025.11 | 0.033 | 0.006 | — | |
| MSTN-TransformerMask R.=37.5%2025.11 | 0.038 | 0.007 | — | |
| MSTN-BiLSTMMask R.=25%2025.11 | 0.042 | 0.012 | — | |
| MSTN-BiLSTMMask R.=Avg2025.11 | 0.052 | 0.015 | — | |
| MSTN-TransformerMask R.=50%2025.11 | 0.057 | 0.011 | — | |
| MSTN-BiLSTMMask R.=37.5%2025.11 | 0.062 | 0.017 | — | |
| MSTN-BiLSTMMask R.=50%2025.11 | 0.081 | 0.022 | — | |
| ModernTCN+ALER-TImissing rate (r)=0.125, Average sequence length L={96, 192, 336, 720}2026.07 | 0.127 | — | — | |
| Bridge-TS-2Ratio=12.5%2025.12 | 0.131 | 0.037 | — | |
| Non-stationaryRatio=12.5%2025.12 | 0.131 | 0.037 | — | |
| ModernTCNmissing rate (r)=0.125, Average sequence length L={96, 192, 336, 720}2026.07 | 0.134 | — | — | |
| T1Scenario=Point missing, Missing Ratios=0.1, 0.3, 0.5, 0.7 (avg), Sequence Length=962026.02 | 0.138 | 0.049 | — | |
| GPT2(3)Mask R.=12.5%2025.11 | 0.14 | 0.043 | — | |
| Bridge-TS-2Ratio=25.0%2025.12 | 0.146 | 0.048 | — | |
| Auto*+ALER-TImissing rate (r)=0.125, Average sequence length L={96, 192, 336, 720}2026.07 | 0.146 | — | — | |
| Non-stationaryRatio=25.0%2025.12 | 0.15 | 0.051 | — | |
| GTMAveraging=varying data missing ratios, Protocol=point-wise missing ratios2025.02 | 0.152 | 0.053 | — | |
| GPT2(3)Mask R.=25%2025.11 | 0.156 | 0.054 | — | |
| TimesNetMask R.=12.5%2025.11 | 0.159 | 0.057 | — | |
| Bridge-TS-2Ratio=37.5%2025.12 | 0.165 | 0.061 | — | |
| StationaryMask R.=12.5%2025.11 | 0.165 | 0.06 | — | |
| TimesNetRatio=12.5%2025.12 | 0.168 | 0.061 | — | |
| Cross*+ALER-TImissing rate (r)=0.125, Average sequence length L={96, 192, 336, 720}2026.07 | 0.17 | — | — | |
| TimesNet+ALER-TImissing rate (r)=0.125, Average sequence length L={96, 192, 336, 720}2026.07 | 0.171 | — | — | |
| GPT4TSMask Ratio=Avg2024.11 | 0.173 | 0.069 | — | |
| GPT2(3)Number of Layers=3, Sequence Length=96, Mask Ratios=12.5%, 25%, 37.5%, 50%2023.02 | 0.173 | 0.069 | — | |
| GPT4TSAveraging=varying data missing ratios, Protocol=point-wise missing ratios2025.02 | 0.173 | 0.069 | — | |
| GPT2(3)Mask R.=Avg2025.11 | 0.173 | 0.069 | — | |
| ZEUSProtocol=Zero-shot, Masking Strategy=Random2026.07 | 0.175 | 0.079 | — | |
| Auto*missing rate (r)=0.125, Average sequence length L={96, 192, 336, 720}2026.07 | 0.176 | — | — | |
| Non-stationaryRatio=37.5%2025.12 | 0.177 | 0.068 | — | |
| TimesNetmissing rate (r)=0.125, Average sequence length L={96, 192, 336, 720}2026.07 | 0.177 | — | — | |
| SAITSScenario=Point missing, Missing Ratios=0.1, 0.3, 0.5, 0.7 (avg), Sequence Length=962026.02 | 0.178 | 0.092 | — | |
| CSDIScenario=Point missing, Missing Ratios=0.1, 0.3, 0.5, 0.7 (avg), Sequence Length=962026.02 | 0.178 | 0.083 | — | |
| TimesNetMask R.=25%2025.11 | 0.178 | 0.069 | — | |
| FEDformerRatio=12.5%2025.12 | 0.18 | 0.062 | — | |
| GPT2(3)Mask R.=37.5%2025.11 | 0.18 | 0.072 | — | |
| PatchTST+ALER-TImissing rate (r)=0.125, Average sequence length L={96, 192, 336, 720}2026.07 | 0.182 | — | — | |
| DLinear+ALER-TImissing rate (r)=0.125, Average sequence length L={96, 192, 336, 720}2026.07 | 0.184 | — | — | |
| PatchTSTScenario=Point missing, Missing Ratios=0.1, 0.3, 0.5, 0.7 (avg), Sequence Length=962026.02 | 0.185 | 0.082 | — | |
| TimesNetSequence Length=96, Mask Ratios=12.5%, 25%, 37.5%, 50%2023.02 | 0.187 | 0.078 | — | |
| TimesNetAveraging=varying data missing ratios, Protocol=point-wise missing ratios2025.02 | 0.187 | 0.078 | — | |
| TimesNetMask R.=Avg2025.11 | 0.187 | 0.078 | — | |
| Bridge-TS-2Ratio=50.0%2025.12 | 0.189 | 0.08 | — | |
| ModernTCNScenario=Point missing, Missing Ratios=0.1, 0.3, 0.5, 0.7 (avg), Sequence Length=962026.02 | 0.189 | 0.083 | — | |
| StationaryMask R.=25%2025.11 | 0.189 | 0.08 | — | |
| FEDformerMask R.=12.5%2025.11 | 0.19 | 0.07 | — | |
| Peri-midFormerMask Ratio=Avg2024.11 | 0.191 | 0.083 | — | |
| HGTS-Former2025.08 | 0.192 | 0.085 | — | |
| RLinear+ALER-TImissing rate (r)=0.125, Average sequence length L={96, 192, 336, 720}2026.07 | 0.193 | — | — | |
| TimesNetMask R.=37.5%2025.11 | 0.196 | 0.084 | — | |
| TimeMixer++2025.08 | 0.198 | 0.091 | — | |
| TimesNetProtocol=Supervised, Masking Strategy=Random2026.07 | 0.198 | 0.089 | — | |
| Non-stationary TransformerMask Ratio=Avg2024.11 | 0.201 | 0.094 | — | |
| StationarySequence Length=96, Mask Ratios=12.5%, 25%, 37.5%, 50%2023.02 | 0.201 | 0.094 | — | |
| PatchTSTMask R.=12.5%2025.11 | 0.201 | 0.093 | — | |
| StationaryMask R.=Avg2025.11 | 0.201 | 0.094 | — | |
| ZEUSProtocol=Zero-shot, Masking Strategy=Block2026.07 | 0.202 | 0.115 | — | |
| ModernTCNProtocol=Supervised, Masking Strategy=Random2026.07 | 0.202 | 0.086 | — | |
| Non-stationaryRatio=50.0%2025.12 | 0.205 | 0.093 | — | |
| Helix+ALER-TImissing rate (r)=0.125, Average sequence length L={96, 192, 336, 720}2026.07 | 0.209 | — | — | |
| StationaryMask R.=37.5%2025.11 | 0.212 | 0.102 | — | |
| TimesNetRatio=37.5%2025.12 | 0.213 | 0.102 | — | |
| iTransformer2025.08 | 0.213 | 0.13 | — | |
| AutoformerSequence Length=96, Mask Ratios=12.5%, 25%, 37.5%, 50%2023.02 | 0.214 | 0.103 | — | |
| TimesNetMask R.=50%2025.11 | 0.215 | 0.102 | — | |
| GPT4TSProtocol=Supervised, Masking Strategy=Random2026.07 | 0.215 | 0.103 | — | |
| GPT2(3)Mask R.=50%2025.11 | 0.216 | 0.107 | — | |
| FEDformerRatio=25.0%2025.12 | 0.217 | 0.089 | — | |
| PatchTSTMask R.=25%2025.11 | 0.217 | 0.107 | — | |
| TimesNetProtocol=Supervised, Masking Strategy=Block2026.07 | 0.218 | 0.111 | — | |
| Helixmissing rate (r)=0.125, Average sequence length L={96, 192, 336, 720}2026.07 | 0.22 | — | — | |
| AutoFormerRatio=12.5%2025.12 | 0.222 | 0.092 | — | |
| ModernTCNProtocol=Supervised, Masking Strategy=Block2026.07 | 0.222 | 0.104 | — | |
| BRITSScenario=Point missing, Missing Ratios=0.1, 0.3, 0.5, 0.7 (avg), Sequence Length=962026.02 | 0.223 | 0.121 | — | |
| PatchTSTSequence Length=96, Mask Ratios=12.5%, 25%, 37.5%, 50%2023.02 | 0.224 | 0.115 | — | |
| PatchTSTAveraging=varying data missing ratios, Protocol=point-wise missing ratios2025.02 | 0.224 | 0.115 | — | |
| PatchTSTMask R.=Avg2025.11 | 0.224 | 0.115 | — | |
| PatchTSTmissing rate (r)=0.125, Average sequence length L={96, 192, 336, 720}2026.07 | 0.224 | — | — | |
| FEDformer2025.08 | 0.225 | 0.099 | — | |
| DLinearmissing rate (r)=0.125, Average sequence length L={96, 192, 336, 720}2026.07 | 0.227 | — | — | |
| RLinearmissing rate (r)=0.125, Average sequence length L={96, 192, 336, 720}2026.07 | 0.228 | — | — | |
| TimesNetRatio=25.0%2025.12 | 0.229 | 0.086 | — | |
| TimesNetRatio=50.0%2025.12 | 0.23 | 0.121 | — | |
| PatchTSTMask R.=37.5%2025.11 | 0.23 | 0.12 | — | |
| PSW-IScenario=Point missing, Missing Ratios=0.1, 0.3, 0.5, 0.7 (avg), Sequence Length=962026.02 | 0.231 | 0.126 | — | |
| PatchTST2025.08 | 0.231 | 0.178 | — | |
| TimeMixer++Scenario=Point missing, Missing Ratios=0.1, 0.3, 0.5, 0.7 (avg), Sequence Length=962026.02 | 0.232 | 0.132 | — | |
| DlinearRatio=12.5%2025.12 | 0.233 | 0.113 | — | |
| MAS4TSMasking Ratio=Average (12.5%, 25%, 37.5%, 50%), Publication Year=20242026.02 | 0.235 | 0.129 | — | |
| iTransformerScenario=Point missing, Missing Ratios=0.1, 0.3, 0.5, 0.7 (avg), Sequence Length=962026.02 | 0.236 | 0.129 | — | |
| FEDformerMask R.=25%2025.11 | 0.236 | 0.106 | — | |
| Cross*missing rate (r)=0.125, Average sequence length L={96, 192, 336, 720}2026.07 | 0.236 | — | — | |
| TimesNetScenario=Point missing, Missing Ratios=0.1, 0.3, 0.5, 0.7 (avg), Sequence Length=962026.02 | 0.237 | 0.13 | — | |
| PatchTSTProtocol=Supervised, Masking Strategy=Random2026.07 | 0.239 | 0.131 | — | |
| StationaryMask R.=50%2025.11 | 0.24 | 0.133 | — | |
| TimeMixer2025.08 | 0.242 | 0.152 | — |