diff --git a/crates/ml/src/hyperopt/adapters/dqn.rs b/crates/ml/src/hyperopt/adapters/dqn.rs
index c1a2496aa..27f6cf628 100644
--- a/crates/ml/src/hyperopt/adapters/dqn.rs
+++ b/crates/ml/src/hyperopt/adapters/dqn.rs
@@ -3149,6 +3149,7 @@ impl HyperparameterOptimizable for DQNTrainer {
             None
         };
 
+        let actual_best_epoch = internal_trainer.get_best_epoch();
         let metrics = DQNMetrics {
             train_loss: training_metrics.loss,
             val_loss: internal_trainer.get_best_val_loss(), // Use best validation loss for hyperopt
@@ -3302,7 +3303,7 @@ impl HyperparameterOptimizable for DQNTrainer {
         let best_sharpe = metrics.backtest_metrics.as_ref().map(|b| b.sharpe_ratio).unwrap_or(0.0);
         let best_win_rate = metrics.backtest_metrics.as_ref().map(|b| b.win_rate * 100.0).unwrap_or(0.0);
         let best_max_dd = metrics.backtest_metrics.as_ref().map(|b| b.max_drawdown_pct).unwrap_or(0.0);
-        let best_epoch = metrics.epochs_completed;
+        let best_epoch = actual_best_epoch;
 
         info!(
             "TRIAL_SUMMARY: trial={}, objective={:.4}, sharpe={:.4}, win_rate={:.2}%, \