elmonstrodelbrest/chaosai

How elmonstrodelbrest/chaosai works

ChaosAI is a research pipeline for learning financial time-series representations and evaluating cross-sectional portfolio policies, not a live trading system. The principal workflow tokenizes market data, pretrains a JEPA foundation model with a CFM auxiliary objective, precomputes embeddings and an RL buffer, trains a cross-sectional DQN, then evaluates it on held-out data, optionally applying Multiverse Crossing confidence filtering. The graph preserves the README’s tokenizer, macro/on-chain context, and TD-MPC2 capabilities. Several links follow documented scripts and method stages; unsampled implementations are not inferred beyond those documented relationships.

Main entry points: jepa.py JEPA pretraining · dqn_agent.py Cross-sectional DQN · tokenizer.py FSQ tokenizer · env_cross_sectional.py Portfolio environment

Components 23

Data and context

Market data
FSQ tokenizer src/jax_v6/strate_i/tokenizer.py
Token sequences
Grain loader src/jax_v6/data/grain_loader.py
Macro context src/jax_v6/data/macro_loader.py

Representation learning

JEPA pretraining src/jax_v6/jepa.py
Mamba encoder src/jax_v6/encoders/mamba2_encoder.py
JEPA predictors src/jax_v6/predictors/predictor.py
Flow predictor src/jax_v6/predictors/flow_predictor.py
Representation losses src/jax_v6/losses/vicreg.py
Training runtime src/jax_v6/training/train_step.py
JEPA embeddings

Portfolio learning

RL buffer
Portfolio environment src/jax_v6/strate_iv/env_cross_sectional.py
Cross-sectional DQN src/jax_v6/strate_iv/dqn_agent.py
Replay buffer src/jax_v6/strate_iv/replay_buffer.py

Evaluation and robustness

Multiverse Crossing src/jax_v6/strate_iv/multiverse_crossing.py
Held-out evaluation scripts/eval_oos_temporal.py
MVX evaluation scripts/eval_oos_mvx.py

Additional modeling

On-chain graphs src/strate_v/graph_builder.py
On-chain GNN src/strate_v/gnn_model.py
TD-MPC2 agent src/jax_v6/strate_iv/tdmpc2.py

Other components

Researcher

Connections 27

Researcher → FSQ tokenizer: runs tokenizer
Market data → FSQ tokenizer: tokenizes OHLCV
FSQ tokenizer → Token sequences: produces codes
Token sequences → Grain loader: reads shards
Grain loader → JEPA pretraining: provides batches
Macro context → JEPA pretraining: provides context
On-chain graphs → On-chain GNN: builds graph input
On-chain GNN → JEPA pretraining: provides embeddings
JEPA pretraining → Mamba encoder: invokes encoder
JEPA pretraining → JEPA predictors: invokes predictor
JEPA pretraining → Flow predictor: uses CFM objective
JEPA pretraining → Representation losses: computes losses
Training runtime → JEPA pretraining: updates model
JEPA pretraining → JEPA embeddings: encodes sequences
JEPA embeddings → RL buffer: precomputes transitions
RL buffer → Replay buffer: supplies experience
Replay buffer → Cross-sectional DQN: samples batches
Cross-sectional DQN → Portfolio environment: selects portfolio actions
Portfolio environment → Cross-sectional DQN: returns rewards
JEPA embeddings → Multiverse Crossing: perturbs representations
Cross-sectional DQN → Held-out evaluation: evaluates policy
RL buffer → Held-out evaluation: provides held-out data
Held-out evaluation → Researcher: reports results
Cross-sectional DQN → MVX evaluation: evaluates with filter
Multiverse Crossing → MVX evaluation: provides confidence
RL buffer → MVX evaluation: provides held-out data
TD-MPC2 agent → Portfolio environment: uses portfolio environment

More Python diagrams: microsoft/markitdown · anthropics/skills · significant-gravitas/autogpt · 521xueweihan/hellogithub · huggingface/transformers

Drawn by GitDiagram on September 29, 2026 from the repository's file tree, README and source files. Read as Markdown ·