Summary

The JeTech world model is not an offline RL algorithm. It is a market path generator.

past 224 days of OHLC
  -> world model
  -> next 96 days of OHLC path distribution

The offline RL algorithm comes later:

synthetic OHLC path
  -> action or position labeler
  -> reward and environment simulator
  -> transition dataset
  -> IQL, ReBRAC, CQL, TD3+BC, DT

The correct research label is world-model-assisted offline RL or model-based offline RL with synthetic market rollouts. The world model creates the environment and data used for policy learning. It does not learn the policy itself.

A useful trading world model must do more than draw realistic-looking charts. It must preserve volatility clustering, heavy tails, drawdowns, regime duration, gaps, correlations, OHLC validity, and strategy rankings after transaction costs.

What the World Model Learns

In model-based RL, a world model often learns:

p(next_state, reward | state, action)

JeTech's current price world model learns:

p(future_ohlc_path | past_ohlc_window)

That distinction matters.

For JeTech's medium-horizon OHLC agents, we assume our actions do not move the market. Execution RL and order-book simulation would need action-conditioned dynamics. The current research target is a price-path generator that feeds offline RL.

Link to Model-Based Offline RL

MOPO, MOReL, and COMBO learn dynamics from offline data and use synthetic rollouts. Their warning applies to JeTech: model rollouts become dangerous outside dataset support.

Generated paths should not all receive the same trust. The model is weakest in sparse regimes such as crashes, volatility spikes, and liquidity gaps. If the generator under-samples tails, the agent learns too much leverage. If it over-samples tails, the agent may collapse to flat exposure.

What Makes a Good Market Generator

The generator should match these properties:

The last row matters most. A path generator for policy learning should preserve strategy ranking. If trend following wins in synthetic paths while mean reversion wins in real validation, the world model is a dangerous training environment.

Recent financial diffusion work points in the same direction. Generation of synthetic financial time series by diffusion models applies DDPMs to financial time series and stylized facts. Beyond Visual Realism emphasizes that visual realism does not guarantee strategy-level usefulness. CoFinDiff adds controllable trend and volatility conditions to synthetic financial generation.

Strength of the Current JeTech Structure

The current pipeline separates responsibilities:

1. Train world model on real OHLC
2. Generate synthetic OHLC paths
3. Add exposure through synthetic action labelers
4. Simulate reward and transitions
5. Train offline RL policies
6. Evaluate with real 5Y gate

This separation helps debugging.

The team can ask whether the offline RL algorithm failed or the market generator failed.

Risks in the Current Structure

Action and Price Are Independent

The current model generates price first, then action labelers attach positions. This is simple, but it weakens the joint relationship between regime and action.

In a crash regime, random actions should have wide loss dispersion. Trend following may enter late. Mean reversion may catch rebounds or catch falling knives. A price-first pipeline can miss these joint patterns.

Short term, labelers should become regime-aware. Medium term, JeTech should test joint transition generation.

Tail Calibration Is Hard

Financial policy survival depends on rare days. If tails are too thin, the policy learns too much risk. If tails are too thick, the policy may stop trading.

Minimum tail checks:

1d / 5d / 20d return quantiles
realized volatility percentiles
gap size distribution
max drawdown distribution
drawdown recovery time
volatility after loss events

Regime Duration Matters

A model can predict the next candle well and still fail over 96 days. The generated path must preserve trend persistence, range behavior, crash recovery, and volatility decay.

One-step MSE is not enough. JeTech should add multi-horizon return, realized volatility, drawdown shape, and tail quantile metrics.

Likelihood Can Pick the Wrong Model

Likelihood favors common market days. Trading policies need rare regimes too. World model selection should include stylized facts, strategy ranking, tail calibration, and real-gate transfer.

Improvement Plan

A. Path Quality Report

After each world model run, generate samples and compare them with validation data.

world_model_quality.json
  return_distribution
  volatility_clustering
  drawdown_distribution
  tail_event_frequency
  ohlc_validity
  regime_duration
  strategy_ranking

Store distances, not only pass or fail flags: KS distance, Wasserstein distance, quantile error, autocorrelation error.

B. Regime-Conditioned Generation

Add explicit generation conditions:

This lets the training pipeline oversample rare but important states.

C. Joint Transition Generator

The medium-term target is complete transition generation:

current:
  past OHLC -> future OHLC
  future OHLC + labeler -> action and reward transitions

target:
  past trajectory + policy archetype
    -> OHLC, action, reward, next observation, done

DAWM shows one recent direction: generate state-reward trajectories with a diffusion world model and infer missing actions with inverse dynamics. JeTech does not need to copy it, but the final system should not keep price and action fully isolated.

D. Synthetic-to-Real Gate

World models should not be promoted by standalone metrics alone.

Only world models that pass these gates should generate training datasets.

Research Position

The world model is not the trading policy. It creates the market distribution used to train the trading policy.

A useful world model does not maximize synthetic Sharpe. It increases the rate at which trained policies survive the real gate.