gitadmin/foxhunt
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
979 Commits 13 Branches 0 Tags
ee9f8f6fec181e080e4e91ef65ab9309e81672e9
Commit Graph

6 Commits

Author SHA1 Message Date
jgrusewski
199287f78e fix(ml): warmup assertions, hyperopt architecture alignment, greedy eval 
- Add debug_assert_eq! guards in 4 train_baseline functions to catch
  bar/feature length misalignment at debug time (#4)
- Remove "last sample targets itself" block in hyperopt PPO adapter
  that created ~0 return sample biasing toward HOLD (#5)
- Align hyperopt state_dim 54→51 and num_actions 45→3 to match
  train_baseline architecture, making tuned hyperparams transferable (#6)
- Use greedy_action() in evaluate_baseline PPO eval for deterministic results

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
 2026-02-24 01:28:50 +01:00
jgrusewski
aaaaef7f48 fix(ml): address code review — greedy PPO eval, gamma alignment, Sharpe fix 
Fixes from code review of DQN/PPO validation pipeline:

1. PPO validation: use greedy_action() (argmax) instead of stochastic
   act() — deterministic early-stopping signal, matching DQN's eps=0.

2. DQN eval gamma: align to 0.95 (was 0.99) matching train_baseline.
   Gamma doesn't affect greedy inference but configs should match.

3. Sharpe annualization: use 1380 bars/day (23h futures session) not
   390 (6.5h equities). Fixes ~1.8x underestimate for ES futures.

4. compute_reward: accept f64 total_cost_bps (was f32) to match
   shared spread_cost_bps() from baseline_common.rs.

5. Default max_steps_per_epoch: 2000 (was 0/unlimited) for OOM safety.

6. Hyperopt PPO: add timestamp dedup to decode_ohlcv_bars for .FUT
   parent symbols with overlapping contracts.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
 2026-02-24 01:02:51 +01:00
jgrusewski
24e72b370e fix(ml/ppo): use real policy actions and cap trajectory length to prevent OOM 
Three critical PPO fixes:

1. Add PPO::act_with_log_prob() returning (action, log_prob, value).
   The existing act() discarded the policy log-probability, making
   PPO importance sampling use wrong ratios during training.

2. Cap trajectory length with --max-steps-per-epoch in train_baseline
   PPO path. DQN already had this limit; PPO iterated all features
   (~500K per fold), causing OOM on 4GB GPU.

3. Replace random actions and fake log_prob/value in hyperopt PPO
   adapter with real agent.act_with_log_prob() calls. Trajectories
   now reflect actual policy behavior for meaningful hyperopt.

Also adds warmup offset alignment to PPO trajectory collection
(matching the DQN fix) and fixes .unwrap() in test.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
 2026-02-24 00:52:02 +01:00
jgrusewski
26b51a4f99 feat(ml): real validation, transaction costs, and data fixes for DQN/PPO pipeline 
Replace stub validation functions with real model inference (DQN greedy,
PPO act()) so early stopping optimizes actual trading performance instead
of market volatility. Add transaction costs (commission + bid-ask spread)
to reward computation across train/hyperopt/evaluate examples.

Key changes:
- Symbol filtering (--symbol ES.FUT) prevents mixing futures contracts
- BTreeMap timestamp dedup handles overlapping .FUT contract bars
- Return clamping (--max-bar-return) filters contract roll boundaries
- Warmup offset alignment fixes feature-to-bar index mismatch
- Kelly sizing: 3 stubs replaced with real data-driven implementations
- Adam optimizer: BUG #14 diagnostic logging demoted to trace
- TFT: varmap_mut() accessor for checkpoint loading
- PPO hyperopt: with_costs() builder for tx cost configuration

DQN eval (ES.FUT, 2 folds): Sharpe=11.36, MaxDD=7.42%, WinRate=33.2%

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
 2026-02-24 00:43:11 +01:00
jgrusewski
e3f32742fa feat(ml): walk-forward training pipeline with real Databento data 
Fix zstd decoder in train_baseline.rs and evaluate_baseline.rs (same
pattern as hyperopt adapters — branch on .dbn.zst extension). Add CLI
flags for walk-forward config (train/val/test/step months), learning
rate, and max-steps-per-epoch to make pipeline validation feasible.

Pipeline validated end-to-end: hyperopt (5 trials, best Sharpe 2.37) →
walk-forward training (4 folds, 6/1/1 month windows on ES.FUT) →
evaluation (4 fold test sets, checkpoints + norm stats saved).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
 2026-02-23 21:48:10 +01:00
jgrusewski
7d9f1c6e17 feat(ml): add walk-forward training binary for DQN/PPO 
Add ml/examples/train_baseline.rs that trains DQN and PPO models using
expanding walk-forward windows on real Databento OHLCV data.

Features:
- CLI args via clap (--model, --epochs, --batch-size, --data-dir, etc.)
- Recursive .dbn.zst file discovery and OHLCV bar loading
- 51-dim feature extraction via extract_ml_features()
- Walk-forward window generation with NormStats per fold
- DQN training loop with epsilon-greedy, experience replay, early stopping
- PPO training loop with GAE, trajectory collection, early stopping
- PnL-based reward (BUY/SELL/HOLD)
- Safetensors checkpoint saving per fold
- NormStats JSON export for evaluation reproducibility

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
 2026-02-23 18:59:56 +01:00