Update app.py

app.py

@@ -169,8 +169,8 @@ with st.expander("Train with Indicators (RSI, EMA, Stochastic)"):
     with tcol1:
         ft_ticker = st.text_input("Ticker", "SPY")
     with tcol3:
-        # Intraday intervals are limited by Yahoo (usually <=60d). Keep it simple here.
         ft_interval = st.selectbox("Interval", ["1d", "60m", "30m", "15m"], index=0)
+
     # Allowed lookbacks depend on interval
     if ft_interval == "1d":
         allowed_periods = ["6mo", "1y", "2y", "5y"]
@@ -210,7 +210,13 @@ with st.expander("Train with Indicators (RSI, EMA, Stochastic)"):
                     )
             if df.empty:
                 st.error("No data returned. Try a shorter lookback for intraday (e.g., 30d/60d) or use Interval=1d.")
-                st.stop()            # Handle MultiIndex columns (yfinance can return 2-level columns)
+                st.stop()
+
+            # Determine frequency alias for AutoGluon and ensure tz-naive index
+            freq_alias = {"1d": "B", "60m": "60min", "30m": "30min", "15m": "15min"}.get(ft_interval, "B")
+            df.index = pd.DatetimeIndex(df.index).tz_localize(None)
+
+            # Handle MultiIndex columns (yfinance can return 2-level columns)
             if isinstance(df.columns, pd.MultiIndex):
                 try:
                     sym = df.columns.get_level_values(1).unique()[0]
@@ -248,6 +254,11 @@ with st.expander("Train with Indicators (RSI, EMA, Stochastic)"):
             tsdf = TimeSeriesDataFrame.from_data_frame(
                 ts, id_column="item_id", timestamp_column="timestamp"
             )
+            # Ensure a regular time grid for AutoGluon
+            try:
+                tsdf = tsdf.convert_frequency(freq=freq_alias)
+            except Exception:
+                pass
 
         with st.spinner("Fine-tuning Chronos-Bolt (small demo)…"):
             # Chronos-Bolt preset via hyperparameters; fine_tune on CPU is OK for small steps
@@ -255,6 +266,7 @@ with st.expander("Train with Indicators (RSI, EMA, Stochastic)"):
                 prediction_length=int(pred_len),              # reuse your UI's pred_len
                 eval_metric="WQL",
                 quantile_levels=[0.1, 0.5, 0.9],
+                freq=freq_alias,
             ).fit(
                 train_data=tsdf,
                 enable_ensemble=False,
@@ -295,4 +307,4 @@ with st.expander("Train with Indicators (RSI, EMA, Stochastic)"):
             out.to_csv(index=False).encode("utf-8"),
             file_name=f"{item}_chronos_finetuned.csv",
             mime="text/csv",
-        )
+        )