feat(qav3): add quantile tunable for predictions smoothing

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 ced1b7ff3da73c1a006a481a7a7f30ab26a794bb..f368eaeb0a25e2dc5ee1c3cb94fdc07b05514b74 100644 (file)
--- a/quickadapter/user_data/freqaimodels/LightGBMRegressorQuickAdapterV35.py
+++ b/quickadapter/user_data/freqaimodels/LightGBMRegressorQuickAdapterV35.py
@@ -271,8 +271,8 @@ class LightGBMRegressorQuickAdapterV35(BaseRegressionModel):
         label_period_candles: int,
     ) -> tuple[float, float]:
         predictions_smoothing = self.freqai_info.get("predictions_smoothing", "mean")
-        if predictions_smoothing == "log-sum-exp":
-            return log_sum_exp_min_max_pred(
+        if predictions_smoothing == "quantile":
+            return quantile_min_max_pred(
                 pred_df, fit_live_predictions_candles, label_period_candles
             )
         elif predictions_smoothing == "mean":
@@ -283,6 +283,10 @@ class LightGBMRegressorQuickAdapterV35(BaseRegressionModel):
             return median_min_max_pred(
                 pred_df, fit_live_predictions_candles, label_period_candles
             )
+        else:
+            raise ValueError(
+                f"Invalid predictions_smoothing value: '{predictions_smoothing}'"
+            )
 
     def optuna_hp_enqueue_previous_best_trial(
         self,
@@ -487,22 +491,24 @@ class LightGBMRegressorQuickAdapterV35(BaseRegressionModel):
             return False
 
 
-def log_sum_exp_min_max_pred(
+def mean_min_max_pred(
     pred_df: pd.DataFrame, fit_live_predictions_candles: int, label_period_candles: int
 ) -> tuple[float, float]:
+    pred_df_sorted = (
+        pred_df.select_dtypes(exclude=["object"])
+        .copy()
+        .apply(lambda col: col.sort_values(ascending=False, ignore_index=True))
+    )
+
     label_period_frequency: int = int(
         fit_live_predictions_candles / (label_period_candles * 2)
     )
-    extrema = pred_df.tail(label_period_candles * label_period_frequency)[
-        EXTREMA_COLUMN
-    ]
-    min_pred = real_soft_min(extrema)
-    max_pred = real_soft_max(extrema)
-
-    return min_pred, max_pred
+    min_pred = pred_df_sorted.iloc[-label_period_frequency:].mean()
+    max_pred = pred_df_sorted.iloc[:label_period_frequency].mean()
+    return min_pred[EXTREMA_COLUMN], max_pred[EXTREMA_COLUMN]
 
 
-def mean_min_max_pred(
+def quantile_min_max_pred(
     pred_df: pd.DataFrame, fit_live_predictions_candles: int, label_period_candles: int
 ) -> tuple[float, float]:
     pred_df_sorted = (
@@ -514,8 +520,8 @@ def mean_min_max_pred(
     label_period_frequency: int = int(
         fit_live_predictions_candles / (label_period_candles * 2)
     )
-    min_pred = pred_df_sorted.iloc[-label_period_frequency:].mean()
-    max_pred = pred_df_sorted.iloc[:label_period_frequency].mean()
+    min_pred = pred_df_sorted.iloc[-label_period_frequency:].quantile(0.25)
+    max_pred = pred_df_sorted.iloc[:label_period_frequency].quantile(0.75)
     return min_pred[EXTREMA_COLUMN], max_pred[EXTREMA_COLUMN]
 
 
@@ -636,13 +642,3 @@ def hp_objective(
     error = sklearn.metrics.root_mean_squared_error(y_test, y_pred)
 
     return error
-
-
-def real_soft_max(series: pd.Series, beta=1.0) -> float:
-    maximum = series.max()
-    return maximum + spy.special.logsumexp(beta * (series - maximum)) / beta
-
-
-def real_soft_min(series: pd.Series, beta=1.0) -> float:
-    minimum = series.min()
-    return minimum - spy.special.logsumexp(-beta * (series - minimum)) / beta

diff --git a/quickadapter/user_data/freqaimodels/XGBoostRegressorQuickAdapterV35.py b/quickadapter/user_data/freqaimodels/XGBoostRegressorQuickAdapterV35.py
index 575ef13e3eefec606783dd6f1db31d468f30815e..0aacf3de6bd15f30825910f6216800a4f01b8a48 100644 (file)
--- a/quickadapter/user_data/freqaimodels/XGBoostRegressorQuickAdapterV35.py
+++ b/quickadapter/user_data/freqaimodels/XGBoostRegressorQuickAdapterV35.py
@@ -272,8 +272,8 @@ class XGBoostRegressorQuickAdapterV35(BaseRegressionModel):
         label_period_candles: int,
     ) -> tuple[float, float]:
         predictions_smoothing = self.freqai_info.get("predictions_smoothing", "mean")
-        if predictions_smoothing == "log-sum-exp":
-            return log_sum_exp_min_max_pred(
+        if predictions_smoothing == "quantile":
+            return quantile_min_max_pred(
                 pred_df, fit_live_predictions_candles, label_period_candles
             )
         elif predictions_smoothing == "mean":
@@ -284,6 +284,10 @@ class XGBoostRegressorQuickAdapterV35(BaseRegressionModel):
             return median_min_max_pred(
                 pred_df, fit_live_predictions_candles, label_period_candles
             )
+        else:
+            raise ValueError(
+                f"Invalid predictions_smoothing value '{predictions_smoothing}'"
+            )
 
     def optuna_hp_enqueue_previous_best_trial(
         self,
@@ -488,22 +492,24 @@ class XGBoostRegressorQuickAdapterV35(BaseRegressionModel):
             return False
 
 
-def log_sum_exp_min_max_pred(
+def mean_min_max_pred(
     pred_df: pd.DataFrame, fit_live_predictions_candles: int, label_period_candles: int
 ) -> tuple[float, float]:
+    pred_df_sorted = (
+        pred_df.select_dtypes(exclude=["object"])
+        .copy()
+        .apply(lambda col: col.sort_values(ascending=False, ignore_index=True))
+    )
+
     label_period_frequency: int = int(
         fit_live_predictions_candles / (label_period_candles * 2)
     )
-    extrema = pred_df.tail(label_period_candles * label_period_frequency)[
-        EXTREMA_COLUMN
-    ]
-    min_pred = real_soft_min(extrema)
-    max_pred = real_soft_max(extrema)
-
-    return min_pred, max_pred
+    min_pred = pred_df_sorted.iloc[-label_period_frequency:].mean()
+    max_pred = pred_df_sorted.iloc[:label_period_frequency].mean()
+    return min_pred[EXTREMA_COLUMN], max_pred[EXTREMA_COLUMN]
 
 
-def mean_min_max_pred(
+def quantile_min_max_pred(
     pred_df: pd.DataFrame, fit_live_predictions_candles: int, label_period_candles: int
 ) -> tuple[float, float]:
     pred_df_sorted = (
@@ -515,8 +521,8 @@ def mean_min_max_pred(
     label_period_frequency: int = int(
         fit_live_predictions_candles / (label_period_candles * 2)
     )
-    min_pred = pred_df_sorted.iloc[-label_period_frequency:].mean()
-    max_pred = pred_df_sorted.iloc[:label_period_frequency].mean()
+    min_pred = pred_df_sorted.iloc[-label_period_frequency:].quantile(0.25)
+    max_pred = pred_df_sorted.iloc[:label_period_frequency].quantile(0.75)
     return min_pred[EXTREMA_COLUMN], max_pred[EXTREMA_COLUMN]
 
 
@@ -642,13 +648,3 @@ def hp_objective(
     error = sklearn.metrics.root_mean_squared_error(y_test, y_pred)
 
     return error
-
-
-def real_soft_max(series: pd.Series, beta=1.0) -> float:
-    maximum = series.max()
-    return maximum + spy.special.logsumexp(beta * (series - maximum)) / beta
-
-
-def real_soft_min(series: pd.Series, beta=1.0) -> float:
-    minimum = series.min()
-    return minimum - spy.special.logsumexp(-beta * (series - minimum)) / beta