Repository navigation
π‘ RFC: JEPA World Models for Order Flow β Self-Supervised Representation Learning on L3/MBO Market Data #186
NeoNix-Lab
started this conversation in
Ideas
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Context & Motivation
In market microstructure research, datasets are vast but labeled signal is scarce. A single trading session on BTC-USDT Market-by-Order (MBO/L3) or Level 2 order books generates millions of events: additions, modifications, cancellations, and aggressive trades.
Traditional supervised learning asks:
$$X_t \longrightarrow \text{Price}(t + \Delta t) \quad \text{or} \quad X_t \longrightarrow {\text{LONG}, \text{SHORT}, \text{HOLD}}$$
This setup suffers from two fundamental flaws:
The Paradigm: JEPA (Joint-Embedding Predictive Architecture)
Pioneered by Yann LeCun and recently adapted to financial time series (e.g., Fin-JEPA):
Why JEPA Fits Financial Order Books
Integration with Quant Platform
JEPA acts not as a trading bot, but as a Causal Market Perception Layer / World Model:
Research Synergy with Companion Repositories
Our sister research repository Rnn_V0_1 focuses on recurrent neural networks and sequence modeling. Combining self-supervised sequence encoders (JEPA / RNN) with Quant Platform's deterministic replay runtime and purged/embargoed cross-validation (ADR-0031) provides an institutional laboratory for representation learning.
Open Questions for Discussion
Join the discussion and share your perspectives on self-supervised financial learning!
All reactions