May 2, 20267 min readJeTech Lab

Agent Model Version Changelog: PPO iTransformer v1

Runtime preset and registry gate notes for the first JeTech iTransformer v1 candidate model family.

Agent Model

iTransformer

PPO

Registry Gate

Summary

ppo_itransformer_v1 is the first iTransformer runtime preset used to produce JeTech agent registry candidates. The run name intentionally carries only the minimum identity fields. Full training parameters and gate results are stored in the run directory config.json, W&B config, and uploaded model sidecar metadata.

Run identity

Field	Value
Canonical run name	`ppo_itransformer_v1_{symbol}_{YYYYMMDD_HHMMSS}`
Algorithm	PPO, Stable-Baselines3
Backbone	iTransformer feature extractor
Version	`v1`
Current batch symbols	`BTC`, `ETH`, `SOL`, `XRP`, `DOGE`
Registry artifact scope	Only gate-passing models are uploaded as S3 registry candidates

Runtime preset

Parameter	Value
`window_size`	`128`
`encoder_hidden_size`	`64`
`encoder_layers`	`2`
`encoder_dropout`	`0.1`
`transformer_num_heads`	`4`
`transformer_ffn_dim`	`128`
`features_dim`	`64`
`policy_network`	`mlp`
`value_network`	`mlp`
`policy_hidden_dims`	`[64, 64]`
`value_hidden_dims`	`[64, 64]`

Training defaults

Field	Value
Synthetic reference period	`5y`
Reference interval	`1d`
Episode length	`1000`
GAN innovation blend	`0.75`
Target rollout batch	`1024`
Eval frequency	Same as rollout batch
Total timesteps	`102400`
Current batch script	`trading/agents/scripts/train_crypto_batch.sh`
Batch repeat count	`30` runs per symbol

n_envs and n_steps are resolved automatically from the runtime CPU count. The current batch logs show n_envs=4, n_steps=256, and rollout_batch=1024.

Risk and action controls

Parameter	Value
`target_exposure_deadband`	`0.05`
`rebalance_deadband`	`0.05`
`cooldown_bars`	`2`
`drawdown_stop`	`0.40`
`reward_turnover_penalty`	`0.0005`
`reward_downside_penalty`	`0.05`
`reward_running_drawdown_penalty`	`0.005`
`reward_drawdown_increment_penalty`	`0.50`

Real evaluation and gate

Real evaluation uses recent 5y, 1d data during training. The system still records returns, MDD, and Sharpe across 20 three-month diagnostic windows, but registry promotion is decided by one continuous metric.

Gate field	Value
Gate mode	`5y_30d_avg_sharpe`
Pass condition	`5Y 30d Avg Sharpe >= 1.0`
Rolling window	`30` daily returns
Annualization	`365` periods/year
Diagnostic windows	`20 x 3M`, report-only

For passing models, the uploaded .meta.json includes gate.backtest_5y_rolling_sharpe_avg_30d, gate.min_backtest_5y_rolling_sharpe_avg_30d, and gate.research_metrics. JeTech Lab reads that sidecar metadata to show the candidate model table's 5Y 30d Avg Sharpe column.

W&B logging

W&B receives real-evaluation metrics rather than the full training scalar set.

real_eval/num_timesteps
real_eval/backtest_5y_rolling_sharpe_avg_30d
real_eval/backtest_5y_return_count
real_eval/mean_strategy_window_return
real_eval/mean_market_window_return
real_eval/mean_strategy_window_mdd
real_eval/mean_market_window_mdd
real_eval/mean_strategy_window_sharpe
real_eval/sharpe_window_count
real_eval/sharpe_window_ratio

Change log

Date	Change
2026-05-02	Documented the `ppo_itransformer_v1` batch preset
2026-05-02	Switched the registry gate from three-month window count to `5Y 30d Avg Sharpe >= 0.5`
2026-05-04	Raised the registry gate pass threshold to `5Y 30d Avg Sharpe >= 1.0`
2026-05-02	Standardized candidate table display around `5Y 30d Avg Sharpe`