fix(qav3): avoid overlapping label period windows optimization

Signed-off-by: Jérôme Benoit <jerome.benoit@piment-noir.org>

diff --git a/quickadapter/user_data/freqaimodels/LightGBMRegressorQuickAdapterV35.py b/quickadapter/user_data/freqaimodels/LightGBMRegressorQuickAdapterV35.py
index b433644370c7f9e36b13cf4ebda918bf864b0e6e..602b6ab0cf6af5fb032269433a4cdf1673016a9f 100644 (file)
--- a/quickadapter/user_data/freqaimodels/LightGBMRegressorQuickAdapterV35.py
+++ b/quickadapter/user_data/freqaimodels/LightGBMRegressorQuickAdapterV35.py
@@ -130,15 +130,15 @@ class LightGBMRegressorQuickAdapterV35(BaseRegressionModel):
                 train_window = self.__optuna_period_params[dk.pair].get(
                     "train_period_candles"
                 )
-                X = X.tail(train_window)
-                y = y.tail(train_window)
+                X = X.iloc[-train_window:]
+                y = y.iloc[-train_window:]
                 train_weights = train_weights[-train_window:]
 
                 test_window = self.__optuna_period_params[dk.pair].get(
                     "test_period_candles"
                 )
-                X_test = X_test.tail(test_window)
-                y_test = y_test.tail(test_window)
+                X_test = X_test.iloc[-test_window:]
+                y_test = y_test.iloc[-test_window:]
                 test_weights = test_weights[-test_window:]
 
                 # FIXME: find a better way to propagate optuna computed params to strategy
@@ -560,8 +560,8 @@ def period_objective(
     train_window = trial.suggest_int(
         "train_period_candles", min_train_window, max_train_window, step=candles_step
     )
-    X = X.tail(train_window)
-    y = y.tail(train_window)
+    X = X.iloc[-train_window:]
+    y = y.iloc[-train_window:]
     train_weights = train_weights[-train_window:]
 
     min_test_window: int = int(min_train_window * test_size)
@@ -569,8 +569,23 @@ def period_objective(
     test_window = trial.suggest_int(
         "test_period_candles", min_test_window, max_test_window, step=candles_step
     )
-    X_test = X_test.tail(test_window)
-    y_test = y_test.tail(test_window)
+    min_label_period_candles: int = 10
+    max_label_period_candles: int = min(
+        max(fit_live_predictions_candles // 6, min_label_period_candles), test_window
+    )
+    label_period_candles = trial.suggest_int(
+        "label_period_candles",
+        min_label_period_candles,
+        max_label_period_candles,
+        step=candles_step,
+    )
+    test_window = (test_window // label_period_candles) * label_period_candles
+    if test_window < min_label_period_candles:
+        raise optuna.TrialPruned(
+            f"Adjusted test window {test_window} is too small for minimum label period {min_label_period_candles}."
+        )
+    X_test = X_test.iloc[-test_window:]
+    y_test = y_test.iloc[-test_window:]
     test_weights = test_weights[-test_window:]
 
     # Fit the model
@@ -586,28 +601,26 @@ def period_objective(
     )
     y_pred = model.predict(X_test)
 
-    min_label_period_candles: int = 10
-    max_label_period_candles: int = max(
-        fit_live_predictions_candles // 6, min_label_period_candles
-    )
-    label_period_candles = trial.suggest_int(
-        "label_period_candles",
-        min_label_period_candles,
-        max_label_period_candles,
-        step=candles_step,
-    )
+    n_windows = len(y_test) // label_period_candles
     y_test_windows = [
-        y_test[i : i + label_period_candles]
-        for i in range(0, len(y_test) - label_period_candles + 1)
+        y_test.iloc[i : i + label_period_candles].to_numpy()
+        for i in np.arange(0, label_period_candles * n_windows, label_period_candles)
+    ]
+    test_weights_windows = [
+        test_weights[i : i + label_period_candles]
+        for i in np.arange(0, label_period_candles * n_windows, label_period_candles)
     ]
     y_pred_windows = [
         y_pred[i : i + label_period_candles]
-        for i in range(0, len(y_pred) - label_period_candles + 1)
+        for i in np.arange(0, label_period_candles * n_windows, label_period_candles)
     ]
-    y_test = np.array([window.mean() for window in y_test_windows])
-    y_pred = np.array([window.mean() for window in y_pred_windows])
+    y_test = np.array([window for window in y_test_windows])
+    test_weights = np.concatenate(np.array([window for window in test_weights_windows]))
+    y_pred = np.array([window for window in y_pred_windows])
 
-    error = sklearn.metrics.root_mean_squared_error(y_test, y_pred)
+    error = sklearn.metrics.root_mean_squared_error(
+        y_test, y_pred, sample_weight=test_weights
+    )
 
     return error
 
@@ -640,6 +653,8 @@ def hp_objective(
     )
     y_pred = model.predict(X_test)
 
-    error = sklearn.metrics.root_mean_squared_error(y_test, y_pred)
+    error = sklearn.metrics.root_mean_squared_error(
+        y_test, y_pred, sample_weight=test_weights
+    )
 
     return error

diff --git a/quickadapter/user_data/freqaimodels/XGBoostRegressorQuickAdapterV35.py b/quickadapter/user_data/freqaimodels/XGBoostRegressorQuickAdapterV35.py
index 49608d1477ad5e3e177676b980a80132ea3f30b9..a9e03843b534f1ad2466ebe3a2db041bb68462c4 100644 (file)
--- a/quickadapter/user_data/freqaimodels/XGBoostRegressorQuickAdapterV35.py
+++ b/quickadapter/user_data/freqaimodels/XGBoostRegressorQuickAdapterV35.py
@@ -130,15 +130,15 @@ class XGBoostRegressorQuickAdapterV35(BaseRegressionModel):
                 train_window = self.__optuna_period_params[dk.pair].get(
                     "train_period_candles"
                 )
-                X = X.tail(train_window)
-                y = y.tail(train_window)
+                X = X.iloc[-train_window:]
+                y = y.iloc[-train_window:]
                 train_weights = train_weights[-train_window:]
 
                 test_window = self.__optuna_period_params[dk.pair].get(
                     "test_period_candles"
                 )
-                X_test = X_test.tail(test_window)
-                y_test = y_test.tail(test_window)
+                X_test = X_test.iloc[-test_window:]
+                y_test = y_test.iloc[-test_window:]
                 test_weights = test_weights[-test_window:]
 
                 # FIXME: find a better way to propagate optuna computed params to strategy
@@ -563,8 +563,8 @@ def period_objective(
     train_window = trial.suggest_int(
         "train_period_candles", min_train_window, max_train_window, step=candles_step
     )
-    X = X.tail(train_window)
-    y = y.tail(train_window)
+    X = X.iloc[-train_window:]
+    y = y.iloc[-train_window:]
     train_weights = train_weights[-train_window:]
 
     min_test_window: int = int(min_train_window * test_size)
@@ -572,8 +572,23 @@ def period_objective(
     test_window = trial.suggest_int(
         "test_period_candles", min_test_window, max_test_window, step=candles_step
     )
-    X_test = X_test.tail(test_window)
-    y_test = y_test.tail(test_window)
+    min_label_period_candles: int = 10
+    max_label_period_candles: int = min(
+        max(fit_live_predictions_candles // 6, min_label_period_candles), test_window
+    )
+    label_period_candles = trial.suggest_int(
+        "label_period_candles",
+        min_label_period_candles,
+        max_label_period_candles,
+        step=candles_step,
+    )
+    test_window = (test_window // label_period_candles) * label_period_candles
+    if test_window < min_label_period_candles:
+        raise optuna.TrialPruned(
+            f"Adjusted test window {test_window} is too small for minimum label period {min_label_period_candles}."
+        )
+    X_test = X_test.iloc[-test_window:]
+    y_test = y_test.iloc[-test_window:]
     test_weights = test_weights[-test_window:]
 
     # Fit the model
@@ -594,28 +609,26 @@ def period_objective(
     )
     y_pred = model.predict(X_test)
 
-    min_label_period_candles: int = 10
-    max_label_period_candles: int = max(
-        fit_live_predictions_candles // 6, min_label_period_candles
-    )
-    label_period_candles = trial.suggest_int(
-        "label_period_candles",
-        min_label_period_candles,
-        max_label_period_candles,
-        step=candles_step,
-    )
+    n_windows = len(y_test) // label_period_candles
     y_test_windows = [
-        y_test[i : i + label_period_candles]
-        for i in range(0, len(y_test) - label_period_candles + 1)
+        y_test.iloc[i : i + label_period_candles].to_numpy()
+        for i in np.arange(0, label_period_candles * n_windows, label_period_candles)
+    ]
+    test_weights_windows = [
+        test_weights[i : i + label_period_candles]
+        for i in np.arange(0, label_period_candles * n_windows, label_period_candles)
     ]
     y_pred_windows = [
         y_pred[i : i + label_period_candles]
-        for i in range(0, len(y_pred) - label_period_candles + 1)
+        for i in np.arange(0, label_period_candles * n_windows, label_period_candles)
     ]
-    y_test = np.array([window.mean() for window in y_test_windows])
-    y_pred = np.array([window.mean() for window in y_pred_windows])
+    y_test = np.array([window for window in y_test_windows])
+    test_weights = np.concatenate(np.array([window for window in test_weights_windows]))
+    y_pred = np.array([window for window in y_pred_windows])
 
-    error = sklearn.metrics.root_mean_squared_error(y_test, y_pred)
+    error = sklearn.metrics.root_mean_squared_error(
+        y_test, y_pred, sample_weight=test_weights
+    )
 
     return error
 
@@ -652,6 +665,8 @@ def hp_objective(
     )
     y_pred = model.predict(X_test)
 
-    error = sklearn.metrics.root_mean_squared_error(y_test, y_pred)
+    error = sklearn.metrics.root_mean_squared_error(
+        y_test, y_pred, sample_weight=test_weights
+    )
 
     return error