Forecasting on Time-MQA
98.4FCRDeepSeek-V3.2
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| DeepSeek-V3.2mode=zero-shot, serialization=text2026.06 | 98.4 | 270,979 | 85,688 | |
| CADEmodel size=0.6B, backbone=encoder-based2026.06 | 59.8 | 296,897 | 82,458 | |
| CADEContrastive alignment=True, Loss=One-Directional SupCon2026.06 | 59.8 | 296,897 | 32,268 | |
| CADE w/o SupConContrastive alignment=False, Loss=Cross-entropy only2026.06 | 59.6 | 312,649 | 36,852 | |
| Frozen Time-MoEEncoder=Frozen pretrained foundation2026.06 | 58 | 311,892 | 40,660 | |
| Frozen Random LinearEncoder=Frozen randomly-initialized linear2026.06 | 55.4 | 315,402 | 29,917 | |
| Time-MQASerialization=BPE tokenizer2026.06 | 46 | 440,626 | 30,757 | |
| Time-MQA (Full FT)Serialization=BPE tokenizer, Fine-tuning=Full-parameter2026.06 | 46 | 1,104,244 | 3,479,036 | |
| ITFormer2026.06 | 0 | 383,107 | — |