World Embedding v0.1.0

World Embedding v0.1.0 - Initial Release

Pre-computed data (in the repository)

• Daily embedding vectors (data/world_embedding_daily.csv): 9,520 business days x 64 dimensions (1985-01-02 to 2021-06-30)
• Unsupervised regime labels (data/world_embedding_regime_labels.csv): 16 discrete codes from k-means clustering

Pre-trained model weights (attached below)

Download pretrained_weights.zip (8.6 MB) containing:

Directory	Description	Training period
ew_w1/	Expanding window 1	Train 1985-2000, OOS test 2001-2005
ew_w2/	Expanding window 2	Train 1985-2005, OOS test 2006-2011
ew_w3/	Expanding window 3	Train 1985-2011, OOS test 2012-2017
reference_model/	Full-sample reference model	1985-2017

Each directory contains:

• best_model.pt - PyTorch model state dict
• config.json - Hyperparameters used for training
• norm_stats.npz - Feature normalization statistics (mean/std)

Loading weights

import torch
from worldembedding.model import DSSDE

# Load config and weights
state = torch.load("ew_w3/best_model.pt", map_location="cpu")
model = DSSDE(config)  # see config.json for parameters
model.load_state_dict(state)

Python package

pip install git+https://github.com/elittb/world-embedding.git

from worldembedding import load_embedding, get_principal_components
emb = load_embedding()
epc = get_principal_components(n_components=5)

Paper

World Embedding: The Daily Economic State and Bond Risk Premia
Elham Tabatabaei (2026)
SSRN: https://papers.ssrn.com/abstract=6503446