Sync from GitHub (tests passed)

app/features.py

@@ -15,6 +15,12 @@ from typing import Optional
 
 import numpy as np
 import pandas as pd
+
+# Suppress silent downcasting FutureWarnings globally for this module
+try:
+    pd.set_option('future.no_silent_downcasting', True)
+except Exception:
+    pass
 from sqlalchemy import func
 from sqlalchemy.orm import Session
 
@@ -130,7 +136,7 @@ def load_sentiment_data(
 
 def compute_returns(prices: pd.Series, periods: int = 1) -> pd.Series:
     """Compute percentage returns."""
-    return prices.pct_change(periods)
+    return prices.pct_change(periods, fill_method=None)
 
 
 def compute_sma(prices: pd.Series, window: int) -> pd.Series:

deep_learning/data/feature_store.py

@@ -311,7 +311,7 @@ def build_tft_dataframe(
 
     # ---- 6. Target: next-day simple return ----
     close = price_df["close"]
-    target_ret = close.pct_change().shift(-1)
+    target_ret = close.pct_change(fill_method=None).shift(-1)
     target_ret.name = "target"
 
     # ---- Assemble master DataFrame ----

deep_learning/data/lme_warehouse.py

@@ -163,7 +163,7 @@ def compute_lme_features(
     for w in windows:
         change = stock.diff(w)
         features[f"lme_stock_change_{w}d"] = change
-        pct = stock.pct_change(w)
+        pct = stock.pct_change(w, fill_method=None)
         features[f"lme_stock_pct_change_{w}d"] = pct
 
     features["lme_depletion_rate"] = stock.diff(depletion_window) / depletion_window
@@ -207,7 +207,7 @@ def compute_proxy_lme_features(
 
     features["proxy_vol_zscore"] = (vol - vol_ma20) / vol_std20
     features["proxy_vol_spike"] = (features["proxy_vol_zscore"] > 2.0).astype(np.float32)
-    features["proxy_vol_price_interaction"] = features["proxy_vol_zscore"] * close.pct_change()
+    features["proxy_vol_price_interaction"] = features["proxy_vol_zscore"] * close.pct_change(fill_method=None)
 
     spread_5_20 = close.rolling(5).mean() - close.rolling(20).mean()
     features["proxy_momentum_spread"] = spread_5_20 / close.rolling(20).std().replace(0, np.nan)

deep_learning/inference/predictor.py

@@ -31,11 +31,25 @@ warnings.filterwarnings(
     category=UserWarning,
     module="sklearn",
 )
+warnings.filterwarnings(
+    "ignore",
+    message=".*is an instance of `nn.Module` and is already saved during checkpointing.*",
+    category=UserWarning,
+    module="lightning.pytorch.utilities",
+)
+warnings.filterwarnings(
+    "ignore",
+    category=DeprecationWarning,
+    module="lightning.pytorch.utilities",
+)
 
 from deep_learning.config import TFTASROConfig, get_tft_config
 
 logger = logging.getLogger(__name__)
 
+# Suppress PyTorch Lightning promotional tips ("litlogger", "litmodels")
+logging.getLogger("lightning.pytorch.utilities.rank_zero").setLevel(logging.WARNING)
+
 
 class TFTPredictor:
     """

deep_learning/models/tft_copper.py

@@ -150,6 +150,8 @@ def create_tft_model(
         log_val_interval=1,
     )
 
+    model.save_hyperparameters(ignore=['loss', 'logging_metrics'])
+
     n_params = sum(p.numel() for p in model.parameters())
     n_trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
     logger.info("TFT model created: %d total params, %d trainable", n_params, n_trainable)