Update hmc_learner.py

hmc_learner.py

@@ -50,8 +50,10 @@ def preprocess_data(df: pd.DataFrame) -> pd.DataFrame:
     # risk_score as is (continuous)
     df['risk_score'] = df['risk_score'].astype(float)
 
-    # hour_of_day as cyclic? For simplicity, keep linear for now
-    df['hour'] = df['hour_of_day'].astype(float)
+    # Cyclical encoding of hour
+    hours = df['hour_of_day'].values
+    df['hour_sin'] = np.sin(2 * np.pi * hours / 24)
+    df['hour_cos'] = np.cos(2 * np.pi * hours / 24)
 
     # confidence
     df['confidence'] = df['confidence'].astype(float)
@@ -73,7 +75,8 @@ def build_model(df: pd.DataFrame):
         β_env = pm.Normal("β_env", mu=0, sigma=1)
         β_role = pm.Normal("β_role", mu=0, sigma=1)
         β_risk = pm.Normal("β_risk", mu=0, sigma=1)
-        β_hour = pm.Normal("β_hour", mu=0, sigma=1)
+        β_hour_sin = pm.Normal("β_hour_sin", mu=0, sigma=1)
+        β_hour_cos = pm.Normal("β_hour_cos", mu=0, sigma=1)
         β_conf = pm.Normal("β_conf", mu=0, sigma=1)
 
         # Linear predictor
@@ -81,7 +84,8 @@ def build_model(df: pd.DataFrame):
                    β_env * df['env_prod'].values +
                    β_role * df['role_junior'].values +
                    β_risk * (df['risk_score'].values - 0.5) +  # center risk around 0.5
-                   β_hour * (df['hour'].values - 12) / 12 +    # normalize hour to [-1,1]
+                   β_hour_sin * df['hour_sin'].values +
+                   β_hour_cos * df['hour_cos'].values +
                    β_conf * (df['confidence'].values - 0.5))   # center confidence
 
         # Likelihood