gitadmin/foxhunt
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

perf(dqn): restore best-epoch checkpoint before walk-forward evaluation

Browse Source 
Walk-forward evaluation was using the final epoch model, which may have
overfit. Now loads the best per-epoch Sharpe checkpoint (saved during
training) back into the agent before running the walk-forward backtest.

Trial 0 showed Sharpe +3.30 at epoch 2 vs +1.84 at epoch 8 (final).
The walk-forward should evaluate the peak model, not the final one.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-03-12 13:21:28 +01:00
parent a074244422
commit cf6e36c8a1
1 changed files with 28 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
crates/ml/src/hyperopt/adapters/dqn.rs
View File

@@ -3005,6 +3005,34 @@ impl HyperparameterOptimizable for DQNTrainer {
tracing::debug!("Reached backtest decision point");
tracing::debug!("enable_backtest = {}", self.enable_backtest);
// Restore best checkpoint before walk-forward evaluation.
// Training may overfit in later epochs; the best per-epoch Sharpe checkpoint
// (saved during training) gives a more accurate walk-forward evaluation.
let best_ckpt_path = self.training_paths.checkpoints_dir()
.join(format!("trial_{}_best.safetensors", current_trial));
if best_ckpt_path.exists() {
let best_epoch = internal_trainer.get_best_epoch();
let agent_arc = internal_trainer.get_agent().clone();
let bt_handle = self.runtime_handle.as_ref().ok_or_else(|| {
MLError::ConfigError("runtime_handle is None".to_owned())
})?;
let mut agent_guard = bt_handle.block_on(agent_arc.write());
let ckpt_str = best_ckpt_path.to_string_lossy().to_string();
match &mut *agent_guard {
crate::trainers::dqn::DQNAgentType::Standard(ref mut dqn) => {
if let Err(e) = dqn.load_from_safetensors(&ckpt_str) {
tracing::warn!("Failed to restore best checkpoint for walk-forward: {}", e);
} else {
info!("Restored best checkpoint (epoch {}) for walk-forward evaluation", best_epoch);
}
}
crate::trainers::dqn::DQNAgentType::RegimeConditional(_) => {
tracing::warn!("Best-checkpoint restore not yet implemented for RegimeConditional");
}
}
drop(agent_guard);
}
// Run backtest if enabled
let backtest_metrics = if self.enable_backtest {
tracing::debug!("Backtest is enabled, starting backtest...");