Create save/features.py

save/features.py

@@ -0,0 +1,38 @@
+import pandas as pd, numpy as np
+from typing import Tuple
+
+def clip_by_quantile(s: pd.Series, lo=0.01, hi=0.99):
+    ql, qh = s.quantile(lo), s.quantile(hi)
+    return s.clip(lower=ql, upper=qh)
+
+def build_features(df: pd.DataFrame) -> pd.DataFrame:
+    df = df.copy()
+
+    # 基本的な数値→float化（エラーはNaN）
+    for c in [
+        "sales","operating_income","net_income",
+        "current_assets","current_liabilities",
+        "total_assets","total_liabilities","total_equity",
+        "operating_cash_flow",
+        "market_cagr_pct","key_product_count","key_product_growing_count",
+    ]:
+        if c in df.columns:
+            df[c] = pd.to_numeric(df[c], errors="coerce")
+
+    # 比率特徴
+    df["opm"] = np.where(df["sales"]>0, df["operating_income"]/df["sales"], 0.0)
+    df["npm"] = np.where(df["sales"]>0, df["net_income"]/df["sales"], 0.0)
+    df["cur_ratio"] = np.where(df["current_liabilities"]>0, df["current_assets"]/df["current_liabilities"], np.nan)
+    df["eq_ratio"] = np.where(df["total_assets"]>0, df["total_equity"]/df["total_assets"], np.nan)
+    df["de"] = np.where(df["total_equity"]>0, df["total_liabilities"]/df["total_equity"], np.nan)
+    df["ocf_sales"] = np.where(df["sales"]>0, df["operating_cash_flow"]/df["sales"], 0.0)
+
+    # テキスト派生
+    df["prod_grow_ratio"] = np.where(df["key_product_count"]>0,
+                                     df["key_product_growing_count"]/df["key_product_count"], 0.0)
+
+    # 異常値クリップ（winsorize）
+    for c in ["opm","npm","cur_ratio","eq_ratio","de","ocf_sales","market_cagr_pct","prod_grow_ratio"]:
+        df[c] = clip_by_quantile(df[c].astype(float).fillna(0.0), 0.01, 0.99)
+
+    return df