diff --git a/crypto_spot_bot/strategy.py b/crypto_spot_bot/strategy.py
index 2f807bf..1d6c601 100644
--- a/crypto_spot_bot/strategy.py
+++ b/crypto_spot_bot/strategy.py
@@ -909,6 +909,15 @@ def _torch_forecast_exit_signal(
     if price >= position.take_profit:
         return Signal(position.symbol, "SELL", 0.96, "torch_forecast: take-profit hit", diagnostics)
     if atr_trailing_stop is not None and price <= atr_trailing_stop:
+        if estimated_exit_net_percent < 0:
+            diagnostics["atr_exit_blocked_by_cost"] = True
+            return Signal(
+                position.symbol,
+                "HOLD",
+                0.45,
+                "torch_forecast: ATR trailing touched, but exit is not worth fees",
+                diagnostics,
+            )
         return Signal(position.symbol, "SELL", 0.94, "torch_forecast: ATR trailing stop hit", diagnostics)
     if not _is_torch_forecast(forecast):
         if rebound_fallback_position:
diff --git a/tests/test_strategy.py b/tests/test_strategy.py
index 798fd37..8df6376 100644
--- a/tests/test_strategy.py
+++ b/tests/test_strategy.py
@@ -917,6 +917,43 @@ def test_torch_forecast_holds_fee_churn_exit_after_min_hold(make_settings, tmp_p
     assert signal.diagnostics["forecast_exit_blocked_by_cost"] is True
 
 
+def test_torch_forecast_holds_atr_trailing_exit_that_does_not_cover_fees(make_settings, tmp_path) -> None:
+    settings = make_settings(tmp_path, strategy_mode="torch_forecast", min_hold_seconds=60)
+    strategy = SpotStrategy(settings)
+    candles = _trend_entry_candles(close=100.2)
+    candles[-1].atr_14 = 0.8
+    position = Position(
+        1,
+        "MNTUSDT",
+        1,
+        100,
+        100,
+        0.1,
+        96,
+        120,
+        102,
+        opened_at=utc_now() - timedelta(seconds=600),
+    )
+    ticker = Ticker("MNTUSDT", 100.2, 100.19, 100.21, 10_000_000, 1000, 1.0)
+
+    signal = strategy.exit_signal(
+        position,
+        candles,
+        ticker,
+        forecast={
+            "usable": True,
+            "model": "torch_lstm",
+            "expected_return_percent": 0.4,
+            "probability_up": 0.58,
+            "skill": 0.18,
+            "block_entry": False,
+        },
+    )
+
+    assert signal.action == "HOLD"
+    assert signal.diagnostics["atr_exit_blocked_by_cost"] is True
+
+
 def test_torch_forecast_rebound_fallback_holds_without_model(make_settings, tmp_path) -> None:
     settings = make_settings(tmp_path, strategy_mode="torch_forecast", min_hold_seconds=180)
     strategy = SpotStrategy(settings)