Views
No views yet
| Component | Source | Key Innovation |
|---|---|---|
| Time Series Encoder | TS-JEPA (Sennadir 2025) | 1D-CNN patch tokenizer + Transformer |
| Action Conditioning | JEPA-WMs (Terver 2025) | AdaLN + RoPE in predictor |
| Collapse Prevention | EB-JEPA (Terver 2026) | SIGReg + Inverse Dynamics Model |
| Multi-step Rollout | EB-JEPA | K-step autoregressive training |
| Planner | JEPA-WMs + EB-JEPA | CEM L2 cost / MPPI cumulative cost in latent space |
| TD Branch | TD-JEPA (Bagatella 2025) | Optional separate task encoder for zero-shot RL |
Financial Time Series (T, F)
│
▼
[TimeSeriesTokenizer] ── 1D-CNN patches + position encoding
│
├───► [Context Encoder] (student)
│ │
│ ▼
│ [Predictor] ◄─── Action embedding (weights + signals)
│ (AdaLN + RoPE) │
│ │ │
│ ▼ ▼
│ Predicted target [ActionEmbedder]
│ embeddings ├── weights (continuous)
│ ├── signals (discrete)
│ └── hedge (binary)
│
└───► [Target Encoder] (teacher, EMA frozen)
│
▼
Ground truth target embeddingspython finjepa/run_training_fast.py1python finjepa/train.py --data_source hf \
2 --dataset_name paperswithbacktest/Stocks-Daily-Price \
3 --n_assets 5 --batch_size 128 --epochs 50 --push_to_hub